diff options
| author | Felix Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com> | 2023-09-27 17:46:17 +0100 |
|---|---|---|
| committer | felixjohnny.thomasmathibalan <felixjohnny.thomasmathibalan@arm.com> | 2023-09-28 12:08:05 +0000 |
| commit | afd38f0c617d6f89b2b4532c6c44f116617e2b6f (patch) | |
| tree | 03bc7d5a762099989b16a656fa8d397b490ed70e /examples | |
| parent | bdcb4c148ee2fdeaaddf4cf1e57bbb0de02bb894 (diff) | |
| download | ComputeLibrary-afd38f0c617d6f89b2b4532c6c44f116617e2b6f.tar.gz | |
Apply clang-format on repository
Code is formatted as per a revised clang format configuration
file(not part of this delivery). Version 14.0.6 is used.
Exclusion List:
- files with .cl extension
- files that are not strictly C/C++ (e.g. Android.bp, Sconscript ...)
And the following directories
- compute_kernel_writer/validation/
- tests/
- include/
- src/core/NEON/kernels/convolution/
- src/core/NEON/kernels/arm_gemm/
- src/core/NEON/kernels/arm_conv/
- data/
There will be a follow up for formatting of .cl files and the
files under tests/ and compute_kernel_writer/validation/.
Signed-off-by: Felix Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>
Change-Id: Ib7eb1fcf4e7537b9feaefcfc15098a804a3fde0a
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10391
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Diffstat (limited to 'examples')
41 files changed, 5693 insertions, 5422 deletions
diff --git a/examples/cl_cache.cpp b/examples/cl_cache.cpp index 6de62f7c5d..9da5b9176d 100644 --- a/examples/cl_cache.cpp +++ b/examples/cl_cache.cpp @@ -25,8 +25,9 @@ #include "arm_compute/core/Types.h" #include "arm_compute/runtime/CL/CLHelpers.h" #include "arm_compute/runtime/CL/CLScheduler.h" -#include "arm_compute/runtime/CL/Utils.h" #include "arm_compute/runtime/CL/functions/CLPermute.h" +#include "arm_compute/runtime/CL/Utils.h" + #include "utils/Utils.h" using namespace arm_compute; @@ -43,14 +44,15 @@ public: bool do_setup(int argc, char **argv) override { - std::cout << "Once the program has run and created the file cache.bin, rerun with --restore_cache." << std::endl; + std::cout << "Once the program has run and created the file cache.bin, rerun with --restore_cache." + << std::endl; CLScheduler::get().default_init(); - if(argc > 1) + if (argc > 1) { std::string argv1 = argv[1]; std::transform(argv1.begin(), argv1.end(), argv1.begin(), ::tolower); - if(argv1 == "--restore_cache") + if (argv1 == "--restore_cache") { // Load the precompiled kernels from a file into the kernel library, in this way the next time they are needed // compilation won't be required. @@ -110,11 +112,13 @@ private: window.use_tensor_dimensions(reference.info()->tensor_shape()); Iterator it_ref(&reference, window); Iterator it_res(&result, window); - execute_window_loop(window, [&](const Coordinates &) - { - assert(*reinterpret_cast<unsigned char *>(it_ref.ptr()) == *reinterpret_cast<unsigned char *>(it_res.ptr())); - }, - it_ref, it_res); + execute_window_loop( + window, + [&](const Coordinates &) { + assert(*reinterpret_cast<unsigned char *>(it_ref.ptr()) == + *reinterpret_cast<unsigned char *>(it_res.ptr())); + }, + it_ref, it_res); reference.unmap(); result.unmap(); } @@ -126,11 +130,9 @@ private: window.use_tensor_dimensions(tensor.info()->tensor_shape()); Iterator it_tensor(&tensor, window); unsigned char val(0); - execute_window_loop(window, [&](const Coordinates &) - { - *reinterpret_cast<unsigned char *>(it_tensor.ptr()) = val++; - }, - it_tensor); + execute_window_loop( + window, [&](const Coordinates &) { *reinterpret_cast<unsigned char *>(it_tensor.ptr()) = val++; }, + it_tensor); tensor.unmap(); } void init_tensor(const TensorShape shape, CLTensor &tensor, DataType type, DataLayout layout) diff --git a/examples/cl_sgemm.cpp b/examples/cl_sgemm.cpp index 27af228954..68955c52f7 100644 --- a/examples/cl_sgemm.cpp +++ b/examples/cl_sgemm.cpp @@ -29,6 +29,7 @@ #include "arm_compute/runtime/CL/CLScheduler.h" #include "arm_compute/runtime/CL/CLTuner.h" #include "arm_compute/runtime/CL/functions/CLGEMM.h" + #include "utils/Utils.h" #include <cstdlib> @@ -50,15 +51,16 @@ public: CLScheduler::get().default_init(&tuner); std::ifstream stream; - if(argc > 1) + if (argc > 1) { stream.open(argv[1], std::fstream::in); } - if(argc < 3 || (argc < 4 && stream.bad())) + if (argc < 3 || (argc < 4 && stream.bad())) { // Print help - std::cout << "Usage: 1) ./build/cl_sgemm input_matrix_1.npy input_matrix_2.npy [input_matrix_3.npy] [alpha = 1] [beta = 0]\n"; + std::cout << "Usage: 1) ./build/cl_sgemm input_matrix_1.npy input_matrix_2.npy [input_matrix_3.npy] [alpha " + "= 1] [beta = 0]\n"; std::cout << " 2) ./build/cl_sgemm M N K [alpha = 1.0f] [beta = 0.0f]\n\n"; std::cout << "Too few or no input_matrices provided. Using M=7, N=3, K=5, alpha=1.0f and beta=0.0f\n\n"; @@ -68,29 +70,29 @@ public: } else { - if(stream.good()) /* case file1.npy file2.npy [file3.npy] [alpha = 1.0f] [beta = 0.0f] */ + if (stream.good()) /* case file1.npy file2.npy [file3.npy] [alpha = 1.0f] [beta = 0.0f] */ { npy0.open(argv[1]); npy0.init_tensor(src0, DataType::F32); npy1.open(argv[2]); npy1.init_tensor(src1, DataType::F32); - if(argc > 3) + if (argc > 3) { stream.close(); stream.clear(); stream.open(argv[3], std::fstream::in); - if(stream.good()) /* case with third file */ + if (stream.good()) /* case with third file */ { npy2.open(argv[3]); npy2.init_tensor(src2, DataType::F32); - if(argc > 4) + if (argc > 4) { // Convert string to float alpha = strtof(argv[4], nullptr); - if(argc > 5) + if (argc > 5) { // Convert string to float beta = strtof(argv[5], nullptr); @@ -101,7 +103,7 @@ public: { alpha = strtof(argv[3], nullptr); - if(argc > 4) + if (argc > 4) { beta = strtof(argv[4], nullptr); } @@ -118,11 +120,11 @@ public: src1.allocator()->init(TensorInfo(TensorShape(N, K), 1, DataType::F32)); src2.allocator()->init(TensorInfo(TensorShape(N, M), 1, DataType::F32)); - if(argc > 4) + if (argc > 4) { alpha = strtof(argv[4], nullptr); - if(argc > 5) + if (argc > 5) { beta = strtof(argv[5], nullptr); } @@ -141,7 +143,7 @@ public: dst.allocator()->allocate(); // Fill the input images with either the data provided or random data - if(npy0.is_open()) + if (npy0.is_open()) { npy0.fill_tensor(src0); npy1.fill_tensor(src1); @@ -149,7 +151,7 @@ public: output_filename = "sgemm_out.npy"; is_fortran = npy0.is_fortran(); - if(npy2.is_open()) + if (npy2.is_open()) { src2.allocator()->allocate(); npy2.fill_tensor(src2); @@ -179,7 +181,7 @@ public: } void do_teardown() override { - if(!output_filename.empty()) /* Save to .npy file */ + if (!output_filename.empty()) /* Save to .npy file */ { save_to_npy(dst, output_filename, is_fortran); } diff --git a/examples/gemm_tuner/CommonGemmExampleOptions.cpp b/examples/gemm_tuner/CommonGemmExampleOptions.cpp index bee202b99e..c2a465604a 100644 --- a/examples/gemm_tuner/CommonGemmExampleOptions.cpp +++ b/examples/gemm_tuner/CommonGemmExampleOptions.cpp @@ -39,7 +39,8 @@ using namespace utils; return os; } -CommonGemmExampleOptions::CommonGemmExampleOptions(arm_compute::utils::CommandLineParser &parser, arm_compute::DataType default_data_type) +CommonGemmExampleOptions::CommonGemmExampleOptions(arm_compute::utils::CommandLineParser &parser, + arm_compute::DataType default_data_type) : help(parser.add_option<ToggleOption>("help")), M(parser.add_positional_option<SimpleOption<size_t>>("M", 100)), N(parser.add_positional_option<SimpleOption<size_t>>("N", 100)), @@ -48,21 +49,16 @@ CommonGemmExampleOptions::CommonGemmExampleOptions(arm_compute::utils::CommandLi data_type(), tuner_mode() { - const std::set<DataType> supported_data_types - { + const std::set<DataType> supported_data_types{ DataType::F16, DataType::F32, DataType::QASYMM8, }; - const std::set<CLTunerMode> supported_tuner_modes - { - CLTunerMode::EXHAUSTIVE, - CLTunerMode::NORMAL, - CLTunerMode::RAPID - }; + const std::set<CLTunerMode> supported_tuner_modes{CLTunerMode::EXHAUSTIVE, CLTunerMode::NORMAL, CLTunerMode::RAPID}; - ARM_COMPUTE_ERROR_ON_MSG(supported_data_types.find(default_data_type) == supported_data_types.end(), "Default data type unsupported"); + ARM_COMPUTE_ERROR_ON_MSG(supported_data_types.find(default_data_type) == supported_data_types.end(), + "Default data type unsupported"); data_type = parser.add_option<EnumOption<DataType>>("type", supported_data_types, default_data_type); tuner_mode = parser.add_option<EnumOption<CLTunerMode>>("tuner-mode", supported_tuner_modes, CLTunerMode::RAPID); diff --git a/examples/gemm_tuner/CommonGemmExampleOptions.h b/examples/gemm_tuner/CommonGemmExampleOptions.h index f7447e3db3..38178bcef8 100644 --- a/examples/gemm_tuner/CommonGemmExampleOptions.h +++ b/examples/gemm_tuner/CommonGemmExampleOptions.h @@ -27,21 +27,22 @@ #include "arm_compute/core/Types.h" #include "arm_compute/core/Utils.h" #include "arm_compute/runtime/CL/CLTuner.h" -#include "utils/TypePrinter.h" + #include "utils/command_line/CommandLineOptions.h" #include "utils/command_line/CommandLineParser.h" +#include "utils/TypePrinter.h" namespace gemm_tuner { /** Structure holding all the common gemm example parameters */ struct CommonGemmExampleParams { - size_t M{ 100 }; /**< Number of lhs matrix rows */ - size_t N{ 100 }; /**< Number of rhs matrix columns */ - size_t K{ 50 }; /**< Number of lhs matrix columns/rhs matrix rows */ - size_t B{ 1 }; /**< Batch size */ - arm_compute::DataType data_type{ arm_compute::DataType::F32 }; /**< Data type */ - arm_compute::CLTunerMode tuner_mode{ arm_compute::CLTunerMode::RAPID }; /**< OpenCL tuner mode */ + size_t M{100}; /**< Number of lhs matrix rows */ + size_t N{100}; /**< Number of rhs matrix columns */ + size_t K{50}; /**< Number of lhs matrix columns/rhs matrix rows */ + size_t B{1}; /**< Batch size */ + arm_compute::DataType data_type{arm_compute::DataType::F32}; /**< Data type */ + arm_compute::CLTunerMode tuner_mode{arm_compute::CLTunerMode::RAPID}; /**< OpenCL tuner mode */ }; /** Formatted output of the CommonGemmExampleParams type @@ -70,7 +71,8 @@ public: * @param[in,out] parser A parser on which "parse()" hasn't been called yet. * @param[in] default_data_type Default data type if unspecified. */ - CommonGemmExampleOptions(arm_compute::utils::CommandLineParser &parser, arm_compute::DataType default_data_type = arm_compute::DataType::F32); + CommonGemmExampleOptions(arm_compute::utils::CommandLineParser &parser, + arm_compute::DataType default_data_type = arm_compute::DataType::F32); /** Prevent instances of this class from being copied (As this class contains pointers) */ CommonGemmExampleOptions(const CommonGemmExampleOptions &) = delete; /** Prevent instances of this class from being copied (As this class contains pointers) */ @@ -82,11 +84,11 @@ public: /** Default destructor */ ~CommonGemmExampleOptions() = default; - arm_compute::utils::ToggleOption *help; /**< Show help option */ - arm_compute::utils::SimpleOption<size_t> *M; /**< Number of lhs matrix rows option */ - arm_compute::utils::SimpleOption<size_t> *N; /**< Number of rhs matrix columns option */ - arm_compute::utils::SimpleOption<size_t> *K; /**< Number of lhs matrix columns/rhs matrix rows option */ - arm_compute::utils::SimpleOption<size_t> *B; /**< Batch size option */ + arm_compute::utils::ToggleOption *help; /**< Show help option */ + arm_compute::utils::SimpleOption<size_t> *M; /**< Number of lhs matrix rows option */ + arm_compute::utils::SimpleOption<size_t> *N; /**< Number of rhs matrix columns option */ + arm_compute::utils::SimpleOption<size_t> *K; /**< Number of lhs matrix columns/rhs matrix rows option */ + arm_compute::utils::SimpleOption<size_t> *B; /**< Batch size option */ arm_compute::utils::EnumOption<arm_compute::DataType> *data_type; /**< Data type */ arm_compute::utils::EnumOption<arm_compute::CLTunerMode> *tuner_mode; /**< OpenCL tuner mode */ }; diff --git a/examples/gemm_tuner/GemmTunerHelpers.h b/examples/gemm_tuner/GemmTunerHelpers.h index ae5cfbb19e..dbff9e2dff 100644 --- a/examples/gemm_tuner/GemmTunerHelpers.h +++ b/examples/gemm_tuner/GemmTunerHelpers.h @@ -36,9 +36,9 @@ bool update_padding_for_cl_image(arm_compute::ITensorInfo *tensor) constexpr unsigned int num_floats_per_pixel = 4; const unsigned int stride_y_in_elements = tensor->strides_in_bytes()[1] / tensor->element_size(); - const unsigned int pixel_aligment = arm_compute::get_cl_image_pitch_alignment( - arm_compute::CLKernelLibrary::get().get_device()); - if(pixel_aligment == 0) + const unsigned int pixel_aligment = + arm_compute::get_cl_image_pitch_alignment(arm_compute::CLKernelLibrary::get().get_device()); + if (pixel_aligment == 0) { return false; } diff --git a/examples/gemm_tuner/cl_gemm_native.cpp b/examples/gemm_tuner/cl_gemm_native.cpp index dd03873921..7daa0b07d3 100644 --- a/examples/gemm_tuner/cl_gemm_native.cpp +++ b/examples/gemm_tuner/cl_gemm_native.cpp @@ -25,19 +25,20 @@ #error "This example needs to be built with -DARM_COMPUTE_CL" #endif /* ARM_COMPUTE_CL */ -#include "CommonGemmExampleOptions.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/KernelDescriptors.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLScheduler.h" #include "arm_compute/runtime/CL/CLTuner.h" + #include "src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h" #include "tests/CL/Helper.h" -#include "utils/Utils.h" #include "utils/command_line/CommandLineOptions.h" #include "utils/command_line/CommandLineParser.h" +#include "utils/Utils.h" +#include "CommonGemmExampleOptions.h" #include <cstdlib> using namespace arm_compute; @@ -51,9 +52,9 @@ namespace /** Structure holding all tunable gemm configs specific to this example/strategy */ struct GemmConfigs { - size_t m0{ 4 }; /**< Number of rows processed by the matrix multiplication */ - size_t n0{ 4 }; /**< Number of columns processed by the matrix multiplication */ - size_t k0{ 4 }; /**< Number of partial accumulations performed by the matrix multiplication */ + size_t m0{4}; /**< Number of rows processed by the matrix multiplication */ + size_t n0{4}; /**< Number of columns processed by the matrix multiplication */ + size_t k0{4}; /**< Number of partial accumulations performed by the matrix multiplication */ }; /** Formatted output of the GemmConfigs type @@ -145,13 +146,13 @@ public: // Parse command line options parser.parse(argc, argv); - if(param_options.help->is_set() && param_options.help->value()) + if (param_options.help->is_set() && param_options.help->value()) { // Print help message parser.print_help(argv[0]); return false; } - if(!parser.validate()) + if (!parser.validate()) { // Invalid arguments. Use default parameters and configs std::cerr << "Invalid arguments." << std::endl; @@ -198,8 +199,9 @@ public: // Validate argments Status status{}; - status = gemm.validate(lhs.info(), rhs.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); - if(!status) + status = gemm.validate(lhs.info(), rhs.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, + kernel_info); + if (!status) { // Unsupported arguments std::cerr << "Unsupported arguments." << std::endl; @@ -221,11 +223,7 @@ public: void do_run() override { // Execute the function - ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, - { ACL_SRC_1, &rhs }, - { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); + ITensorPack gemm_pack({{ACL_SRC_0, &lhs}, {ACL_SRC_1, &rhs}, {ACL_SRC_2, &bias}, {ACL_DST, &dst}}); gemm.run(gemm_pack); // Make sure all the OpenCL jobs are done executing: diff --git a/examples/gemm_tuner/cl_gemm_reshaped.cpp b/examples/gemm_tuner/cl_gemm_reshaped.cpp index 59044477bf..75f3539cb9 100644 --- a/examples/gemm_tuner/cl_gemm_reshaped.cpp +++ b/examples/gemm_tuner/cl_gemm_reshaped.cpp @@ -31,14 +31,15 @@ #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLScheduler.h" #include "arm_compute/runtime/CL/CLTuner.h" + #include "examples/gemm_tuner/CommonGemmExampleOptions.h" #include "examples/gemm_tuner/GemmTunerHelpers.h" #include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.h" #include "src/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.h" #include "tests/CL/Helper.h" -#include "utils/Utils.h" #include "utils/command_line/CommandLineOptions.h" #include "utils/command_line/CommandLineParser.h" +#include "utils/Utils.h" #include <cstdlib> @@ -53,16 +54,16 @@ namespace /** Structure holding all tunable gemm configs specific to this example/strategy */ struct GemmConfigs { - size_t m0{ 4 }; /**< Number of rows processed by the matrix multiplication */ - size_t n0{ 4 }; /**< Number of columns processed by the matrix multiplication */ - size_t k0{ 4 }; /**< Number of partial accumulations performed by the matrix multiplication */ - size_t v0{ 1 }; /**< Number of vertical blocks of size (m0xk0) stored on the same output row */ - size_t h0{ 1 }; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ - bool interleave_lhs{ true }; /**< Interleave lhs matrix */ - bool transpose_lhs{ true }; /**< Transpose lhs matrix. */ - bool interleave_rhs{ true }; /**< Interleave rhs matrix */ - bool transpose_rhs{ true }; /**< Transpose rhs matrix. */ - bool export_to_cl_image_rhs{ true }; /**< Export rhs matrix to cl_image. */ + size_t m0{4}; /**< Number of rows processed by the matrix multiplication */ + size_t n0{4}; /**< Number of columns processed by the matrix multiplication */ + size_t k0{4}; /**< Number of partial accumulations performed by the matrix multiplication */ + size_t v0{1}; /**< Number of vertical blocks of size (m0xk0) stored on the same output row */ + size_t h0{1}; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ + bool interleave_lhs{true}; /**< Interleave lhs matrix */ + bool transpose_lhs{true}; /**< Transpose lhs matrix. */ + bool interleave_rhs{true}; /**< Interleave rhs matrix */ + bool transpose_rhs{true}; /**< Transpose rhs matrix. */ + bool export_to_cl_image_rhs{true}; /**< Export rhs matrix to cl_image. */ }; /** Formatted output of the GemmConfigs type @@ -119,8 +120,10 @@ public: // FIXME: Currently we only support 2 variants of the gemm reshaped kernels in which transpose_lhs and // transpose_rhs are the opposites of each other. In the future we may extend the kernels to include the other // 2 variants (both transposed and none transposed) - transpose_rhs->set_help("Transpose rhs matrix but not lhs matrix (1) / Do not transpose rhs matrix but do transpose lhs matrix (0)"); - export_to_cl_image_rhs->set_help("Export rhs matrix to cl_image (1) / Do not export rhs matrix to cl_image (0)"); + transpose_rhs->set_help("Transpose rhs matrix but not lhs matrix (1) / Do not transpose rhs matrix but do " + "transpose lhs matrix (0)"); + export_to_cl_image_rhs->set_help( + "Export rhs matrix to cl_image (1) / Do not export rhs matrix to cl_image (0)"); } /** Prevent instances of this class from being copied (As this class contains pointers) */ GemmConfigOptions(const GemmConfigOptions &) = delete; @@ -133,17 +136,18 @@ public: /** Default destructor */ ~GemmConfigOptions() = default; - SimpleOption<size_t> *m0; /**< Number of rows processed by the matrix multiplication option */ - SimpleOption<size_t> *n0; /**< Number of columns processed by the matrix multiplication option */ - SimpleOption<size_t> *k0; /**< Number of partial accumulations performed by the matrix multiplication option */ - SimpleOption<size_t> *v0; /**< Number of vertical blocks of size (m0xk0) stored on the same output row option */ - SimpleOption<size_t> *h0; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row option */ + SimpleOption<size_t> *m0; /**< Number of rows processed by the matrix multiplication option */ + SimpleOption<size_t> *n0; /**< Number of columns processed by the matrix multiplication option */ + SimpleOption<size_t> *k0; /**< Number of partial accumulations performed by the matrix multiplication option */ + SimpleOption<size_t> *v0; /**< Number of vertical blocks of size (m0xk0) stored on the same output row option */ + SimpleOption<size_t> *h0; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row option */ SimpleOption<size_t> *interleave_lhs; /**< Interleave lhs matrix option (1 enable; 0 disable) */ SimpleOption<size_t> *interleave_rhs; /**< Interleave rhs matrix option (1 enable; 0 disable) */ // FIXME: Currently we only support 2 variants of the gemm reshaped kernels in which transpose_lhs and // transpose_rhs are the opposites of each other. In the future we may extend the kernels to include the other // 2 variants (both transposed and none transposed) - SimpleOption<size_t> *transpose_rhs; /**< Transpose rhs matrix option (1 enable; 0 disable). Also set the lhs matrix transpose option to the opposite. */ + SimpleOption<size_t> * + transpose_rhs; /**< Transpose rhs matrix option (1 enable; 0 disable). Also set the lhs matrix transpose option to the opposite. */ SimpleOption<size_t> *export_to_cl_image_rhs; /**< Export rhs matrix to cl_image.*/ }; @@ -198,13 +202,13 @@ public: // Parse command line options parser.parse(argc, argv); - if(param_options.help->is_set() && param_options.help->value()) + if (param_options.help->is_set() && param_options.help->value()) { // Print help message parser.print_help(argv[0]); return false; } - if(!parser.validate()) + if (!parser.validate()) { // Invalid arguments. Use default parameters and configs std::cerr << "Invalid arguments." << std::endl; @@ -256,20 +260,22 @@ public: kernel_info.broadcast_bias = true; kernel_info.activation_info = act_info; - if(rhs_info.h0 == 0) + if (rhs_info.h0 == 0) { rhs_info.h0 = std::max(kernel_info.n / rhs_info.n0, 1U); } // Initialise lhs_reshaped tensor info - lhs_reshaped.allocator()->init(TensorInfo(compute_lhs_reshaped_shape(*lhs.info(), lhs_info), 1, params.data_type)); + lhs_reshaped.allocator()->init( + TensorInfo(compute_lhs_reshaped_shape(*lhs.info(), lhs_info), 1, params.data_type)); // Initialise rhs_reshaped tensor info - rhs_reshaped.allocator()->init(TensorInfo(compute_rhs_reshaped_shape(*rhs.info(), rhs_info), 1, params.data_type)); + rhs_reshaped.allocator()->init( + TensorInfo(compute_rhs_reshaped_shape(*rhs.info(), rhs_info), 1, params.data_type)); - if(rhs_info.export_to_cl_image) + if (rhs_info.export_to_cl_image) { - if(!examples::gemm_tuner_helpers::update_padding_for_cl_image(rhs_reshaped.info())) + if (!examples::gemm_tuner_helpers::update_padding_for_cl_image(rhs_reshaped.info())) { std::cerr << "cl_image is not supported on the device, disable export_to_cl_image" << std::endl; return false; @@ -279,7 +285,7 @@ public: // Validate argments Status status{}; status = reshape_lhs.validate(lhs.info(), lhs_reshaped.info(), lhs_info, kernel_info.reinterpret_input_as_3d); - if(!status) + if (!status) { // Unsupported arguments std::cerr << "Unsupported arguments." << std::endl; @@ -287,8 +293,9 @@ public: return false; } - status = gemm.validate(lhs_reshaped.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); - if(!status) + status = gemm.validate(lhs_reshaped.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, + rhs_info, kernel_info); + if (!status) { // Unsupported arguments std::cerr << "Unsupported arguments." << std::endl; @@ -300,7 +307,8 @@ public: reshape_lhs.configure(lhs.info(), lhs_reshaped.info(), lhs_info); // Configure function - gemm.configure(lhs_reshaped.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); + gemm.configure(lhs_reshaped.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, + rhs_info, kernel_info); // Allocate tensors lhs.allocator()->allocate(); @@ -315,14 +323,11 @@ public: void do_run() override { // Execute the functions - ITensorPack reshape_lsh_pack({ { ACL_SRC, &lhs }, { ACL_DST, &lhs_reshaped } }); + ITensorPack reshape_lsh_pack({{ACL_SRC, &lhs}, {ACL_DST, &lhs_reshaped}}); reshape_lhs.run(reshape_lsh_pack); - ITensorPack gemm_pack({ { ACL_SRC_0, &lhs_reshaped }, - { ACL_SRC_1, &rhs_reshaped }, - { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); + ITensorPack gemm_pack( + {{ACL_SRC_0, &lhs_reshaped}, {ACL_SRC_1, &rhs_reshaped}, {ACL_SRC_2, &bias}, {ACL_DST, &dst}}); gemm.run(gemm_pack); // Make sure all the OpenCL jobs are done executing: diff --git a/examples/gemm_tuner/cl_gemm_reshaped_rhs_only.cpp b/examples/gemm_tuner/cl_gemm_reshaped_rhs_only.cpp index 0ad2a65dc2..cfea2c9bac 100644 --- a/examples/gemm_tuner/cl_gemm_reshaped_rhs_only.cpp +++ b/examples/gemm_tuner/cl_gemm_reshaped_rhs_only.cpp @@ -25,20 +25,21 @@ #error "This example needs to be built with -DARM_COMPUTE_CL" #endif /* ARM_COMPUTE_CL */ -#include "CommonGemmExampleOptions.h" -#include "GemmTunerHelpers.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/KernelDescriptors.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLScheduler.h" #include "arm_compute/runtime/CL/CLTuner.h" + #include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h" #include "tests/CL/Helper.h" -#include "utils/Utils.h" #include "utils/command_line/CommandLineOptions.h" #include "utils/command_line/CommandLineParser.h" +#include "utils/Utils.h" +#include "CommonGemmExampleOptions.h" +#include "GemmTunerHelpers.h" #include <cstdlib> using namespace arm_compute; @@ -52,13 +53,13 @@ namespace /** Structure holding all tunable gemm configs specific to this example/strategy */ struct GemmConfigs { - size_t m0{ 4 }; /**< Number of rows processed by the matrix multiplication */ - size_t n0{ 4 }; /**< Number of columns processed by the matrix multiplication */ - size_t k0{ 4 }; /**< Number of partial accumulations performed by the matrix multiplication */ - size_t h0{ 1 }; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ - bool interleave_rhs{ true }; /**< Interleave rhs matrix */ - bool transpose_rhs{ true }; /**< Transpose rhs matrix */ - bool export_to_cl_image_rhs{ true }; /**< Export rhs matrix to cl_image.*/ + size_t m0{4}; /**< Number of rows processed by the matrix multiplication */ + size_t n0{4}; /**< Number of columns processed by the matrix multiplication */ + size_t k0{4}; /**< Number of partial accumulations performed by the matrix multiplication */ + size_t h0{1}; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ + bool interleave_rhs{true}; /**< Interleave rhs matrix */ + bool transpose_rhs{true}; /**< Transpose rhs matrix */ + bool export_to_cl_image_rhs{true}; /**< Export rhs matrix to cl_image.*/ }; /** Formatted output of the GemmConfigs type @@ -106,7 +107,8 @@ public: h0->set_help("Number of horizontal blocks of size (k0xn0) stored on the same output row"); interleave_rhs->set_help("Interleave rhs matrix (1) / Do not interleave rhs matrix (0)"); transpose_rhs->set_help("Transpose rhs matrix (1) / Do not transpose rhs matrix (0)"); - export_to_cl_image_rhs->set_help("Export rhs matrix to cl_image (1) / Do not export rhs matrix to cl_image (0)"); + export_to_cl_image_rhs->set_help( + "Export rhs matrix to cl_image (1) / Do not export rhs matrix to cl_image (0)"); } /** Prevent instances of this class from being copied (As this class contains pointers) */ GemmConfigOptions(const GemmConfigOptions &) = delete; @@ -119,10 +121,10 @@ public: /** Default destructor */ ~GemmConfigOptions() = default; - SimpleOption<size_t> *m0; /**< Number of rows processed by the matrix multiplication option */ - SimpleOption<size_t> *n0; /**< Number of columns processed by the matrix multiplication option */ - SimpleOption<size_t> *k0; /**< Number of partial accumulations performed by the matrix multiplication option */ - SimpleOption<size_t> *h0; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row option */ + SimpleOption<size_t> *m0; /**< Number of rows processed by the matrix multiplication option */ + SimpleOption<size_t> *n0; /**< Number of columns processed by the matrix multiplication option */ + SimpleOption<size_t> *k0; /**< Number of partial accumulations performed by the matrix multiplication option */ + SimpleOption<size_t> *h0; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row option */ SimpleOption<size_t> *interleave_rhs; /**< Interleave rhs matrix option (1 enable; 0 disable) */ SimpleOption<size_t> *transpose_rhs; /**< Transpose rhs matrix option (1 enable; 0 disable) */ SimpleOption<size_t> *export_to_cl_image_rhs; /**< Export rhs matrix to cl_image.*/ @@ -170,13 +172,13 @@ public: // Parse command line options parser.parse(argc, argv); - if(param_options.help->is_set() && param_options.help->value()) + if (param_options.help->is_set() && param_options.help->value()) { // Print help message parser.print_help(argv[0]); return false; } - if(!parser.validate()) + if (!parser.validate()) { // Invalid arguments. Use default parameters and configs std::cerr << "Invalid arguments." << std::endl; @@ -225,17 +227,18 @@ public: kernel_info.broadcast_bias = true; kernel_info.activation_info = act_info; - if(rhs_info.h0 == 0) + if (rhs_info.h0 == 0) { rhs_info.h0 = std::max(kernel_info.n / rhs_info.n0, 1U); } // Initialise rhs_reshaped tensor info - rhs_reshaped.allocator()->init(TensorInfo(compute_rhs_reshaped_shape(*rhs.info(), rhs_info), 1, params.data_type)); + rhs_reshaped.allocator()->init( + TensorInfo(compute_rhs_reshaped_shape(*rhs.info(), rhs_info), 1, params.data_type)); - if(rhs_info.export_to_cl_image) + if (rhs_info.export_to_cl_image) { - if(!examples::gemm_tuner_helpers::update_padding_for_cl_image(rhs_reshaped.info())) + if (!examples::gemm_tuner_helpers::update_padding_for_cl_image(rhs_reshaped.info())) { std::cerr << "cl_image is not supported on the device, disable export_to_cl_image" << std::endl; return false; @@ -244,8 +247,9 @@ public: // Validate argments Status status{}; - status = gemm.validate(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); - if(!status) + status = gemm.validate(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, + rhs_info, kernel_info); + if (!status) { // Unsupported arguments std::cerr << "Unsupported arguments." << std::endl; @@ -254,7 +258,8 @@ public: } // Configure function - gemm.configure(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); + gemm.configure(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, + kernel_info); // Allocate tensors lhs.allocator()->allocate(); @@ -268,11 +273,7 @@ public: void do_run() override { // Execute the function - ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, - { ACL_SRC_1, &rhs_reshaped }, - { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); + ITensorPack gemm_pack({{ACL_SRC_0, &lhs}, {ACL_SRC_1, &rhs_reshaped}, {ACL_SRC_2, &bias}, {ACL_DST, &dst}}); gemm.run(gemm_pack); // Make sure all the OpenCL jobs are done executing: diff --git a/examples/gemm_tuner/cl_gemmlowp_reshaped.cpp b/examples/gemm_tuner/cl_gemmlowp_reshaped.cpp index 9cf9c9fed0..3808b98b7d 100644 --- a/examples/gemm_tuner/cl_gemmlowp_reshaped.cpp +++ b/examples/gemm_tuner/cl_gemmlowp_reshaped.cpp @@ -31,14 +31,15 @@ #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLScheduler.h" #include "arm_compute/runtime/CL/CLTuner.h" + #include "examples/gemm_tuner/CommonGemmExampleOptions.h" #include "examples/gemm_tuner/GemmTunerHelpers.h" #include "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedKernel.h" #include "src/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.h" #include "tests/CL/Helper.h" -#include "utils/Utils.h" #include "utils/command_line/CommandLineOptions.h" #include "utils/command_line/CommandLineParser.h" +#include "utils/Utils.h" #include <cstdlib> @@ -53,15 +54,15 @@ namespace /** Structure holding all tunable gemm configs specific to this example/strategy */ struct GemmConfigs { - size_t m0{ 4 }; /**< Number of rows processed by the matrix multiplication */ - size_t n0{ 4 }; /**< Number of columns processed by the matrix multiplication */ - size_t k0{ 4 }; /**< Number of partial accumulations performed by the matrix multiplication */ - size_t v0{ 1 }; /**< Number of vertical blocks of size (m0xk0) stored on the same output row */ - size_t h0{ 1 }; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ - bool interleave_lhs{ true }; /**< Interleave lhs matrix */ - bool transpose_lhs{ true }; /**< Transpose lhs matrix. */ - bool interleave_rhs{ true }; /**< Interleave rhs matrix */ - bool transpose_rhs{ true }; /**< Transpose rhs matrix. */ + size_t m0{4}; /**< Number of rows processed by the matrix multiplication */ + size_t n0{4}; /**< Number of columns processed by the matrix multiplication */ + size_t k0{4}; /**< Number of partial accumulations performed by the matrix multiplication */ + size_t v0{1}; /**< Number of vertical blocks of size (m0xk0) stored on the same output row */ + size_t h0{1}; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ + bool interleave_lhs{true}; /**< Interleave lhs matrix */ + bool transpose_lhs{true}; /**< Transpose lhs matrix. */ + bool interleave_rhs{true}; /**< Interleave rhs matrix */ + bool transpose_rhs{true}; /**< Transpose rhs matrix. */ }; /** Formatted output of the GemmConfigs type @@ -116,7 +117,8 @@ public: // FIXME: Currently we only support 2 variants of the gemm reshaped kernels in which transpose_lhs and // transpose_rhs are the opposites of each other. In the future we may extend the kernels to include the other // 2 variants (both transposed and none transposed) - transpose_rhs->set_help("Transpose rhs matrix but not lhs matrix (1) / Do not transpose rhs matrix but do transpose lhs matrix (0)"); + transpose_rhs->set_help("Transpose rhs matrix but not lhs matrix (1) / Do not transpose rhs matrix but do " + "transpose lhs matrix (0)"); } /** Prevent instances of this class from being copied (As this class contains pointers) */ GemmConfigOptions(const GemmConfigOptions &) = delete; @@ -129,17 +131,18 @@ public: /** Default destructor */ ~GemmConfigOptions() = default; - SimpleOption<size_t> *m0; /**< Number of rows processed by the matrix multiplication option */ - SimpleOption<size_t> *n0; /**< Number of columns processed by the matrix multiplication option */ - SimpleOption<size_t> *k0; /**< Number of partial accumulations performed by the matrix multiplication option */ - SimpleOption<size_t> *v0; /**< Number of vertical blocks of size (m0xk0) stored on the same output row option */ - SimpleOption<size_t> *h0; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row option */ + SimpleOption<size_t> *m0; /**< Number of rows processed by the matrix multiplication option */ + SimpleOption<size_t> *n0; /**< Number of columns processed by the matrix multiplication option */ + SimpleOption<size_t> *k0; /**< Number of partial accumulations performed by the matrix multiplication option */ + SimpleOption<size_t> *v0; /**< Number of vertical blocks of size (m0xk0) stored on the same output row option */ + SimpleOption<size_t> *h0; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row option */ SimpleOption<size_t> *interleave_lhs; /**< Interleave lhs matrix option (1 enable; 0 disable) */ SimpleOption<size_t> *interleave_rhs; /**< Interleave rhs matrix option (1 enable; 0 disable) */ // FIXME: Currently we only support 2 variants of the gemm reshaped kernels in which transpose_lhs and // transpose_rhs are the opposites of each other. In the future we may extend the kernels to include the other // 2 variants (both transposed and none transposed) - SimpleOption<size_t> *transpose_rhs; /**< Transpose rhs matrix option (1 enable; 0 disable). Also set the lhs matrix transpose option to the opposite. */ + SimpleOption<size_t> * + transpose_rhs; /**< Transpose rhs matrix option (1 enable; 0 disable). Also set the lhs matrix transpose option to the opposite. */ }; /** Consumes the gemm configuration options and creates a structure containing all information @@ -186,12 +189,12 @@ public: GemmConfigOptions config_options(parser); parser.parse(argc, argv); - if(param_options.help->is_set() && param_options.help->value()) + if (param_options.help->is_set() && param_options.help->value()) { parser.print_help(argv[0]); return false; } - if(!parser.validate()) + if (!parser.validate()) { // Invalid arguments. Use default parameters and configs std::cerr << "Invalid arguments." << std::endl; @@ -217,10 +220,7 @@ public: rhs.allocator()->init(TensorInfo(TensorShape(params.N, params.K, params.B), 1, params.data_type)); // Set arbitrary quantization information - const QuantizationInfo q_info - { - 0.012, 3 - }; + const QuantizationInfo q_info{0.012, 3}; lhs.info()->set_quantization_info(q_info); rhs.info()->set_quantization_info(q_info); dst.info()->set_quantization_info(q_info); @@ -240,45 +240,44 @@ public: rhs_info.transpose = configs.transpose_rhs; rhs_info.export_to_cl_image = false; // CL image not supported for quantized cases yet - if(rhs_info.h0 == 0) + if (rhs_info.h0 == 0) { rhs_info.h0 = std::max(static_cast<unsigned int>(params.N) / rhs_info.n0, 1U); } - lhs_reshaped.allocator()->init(TensorInfo(compute_lhs_reshaped_shape(*lhs.info(), lhs_info), 1, params.data_type)); - rhs_reshaped.allocator()->init(TensorInfo(compute_rhs_reshaped_shape(*rhs.info(), rhs_info), 1, params.data_type)); + lhs_reshaped.allocator()->init( + TensorInfo(compute_lhs_reshaped_shape(*lhs.info(), lhs_info), 1, params.data_type)); + rhs_reshaped.allocator()->init( + TensorInfo(compute_rhs_reshaped_shape(*rhs.info(), rhs_info), 1, params.data_type)); lhs_reshaped.info()->set_quantization_info(q_info); rhs_reshaped.info()->set_quantization_info(q_info); - if(rhs_info.export_to_cl_image) + if (rhs_info.export_to_cl_image) { - if(!examples::gemm_tuner_helpers::update_padding_for_cl_image(rhs_reshaped.info())) + if (!examples::gemm_tuner_helpers::update_padding_for_cl_image(rhs_reshaped.info())) { std::cerr << "cl_image is not supported on the device, disable export_to_cl_image" << std::endl; return false; } } - GEMMReshapeInfo gemm_info - { - static_cast<int>(params.M), - static_cast<int>(params.N), - static_cast<int>(params.K), - static_cast<int>(configs.h0), - static_cast<int>(configs.v0), - 0, - false, - true - }; + GEMMReshapeInfo gemm_info{static_cast<int>(params.M), + static_cast<int>(params.N), + static_cast<int>(params.K), + static_cast<int>(configs.h0), + static_cast<int>(configs.v0), + 0, + false, + true}; // Validate argments - if(!reshape_lhs.validate(lhs.info(), lhs_reshaped.info(), lhs_info, gemm_info.reinterpret_input_as_3d())) + if (!reshape_lhs.validate(lhs.info(), lhs_reshaped.info(), lhs_info, gemm_info.reinterpret_input_as_3d())) { std::cerr << "Invalid arguments for ClGemmReshapeLHSMatrixKernel." << std::endl; return false; } - if(!gemm.validate(lhs_reshaped.info(), rhs_reshaped.info(), dst.info(), lhs_info, rhs_info, gemm_info)) + if (!gemm.validate(lhs_reshaped.info(), rhs_reshaped.info(), dst.info(), lhs_info, rhs_info, gemm_info)) { std::cerr << "Invalid arguments for ClGemmLowpMatrixMultiplyReshapedKernel." << std::endl; return false; @@ -300,10 +299,10 @@ public: } void do_run() override { - ITensorPack reshape_lsh_pack({ { ACL_SRC, &lhs }, { ACL_DST, &lhs_reshaped } }); + ITensorPack reshape_lsh_pack({{ACL_SRC, &lhs}, {ACL_DST, &lhs_reshaped}}); reshape_lhs.run(reshape_lsh_pack); - ITensorPack gemm_pack({ { ACL_SRC_0, &lhs_reshaped }, { ACL_SRC_1, &rhs_reshaped }, { ACL_DST, &dst } }); + ITensorPack gemm_pack({{ACL_SRC_0, &lhs_reshaped}, {ACL_SRC_1, &rhs_reshaped}, {ACL_DST, &dst}}); gemm.run(gemm_pack); // Make sure all the OpenCL jobs are done executing: diff --git a/examples/gemm_tuner/cl_gemmlowp_reshaped_rhs_only_fused_output_stage_fixedpoint.cpp b/examples/gemm_tuner/cl_gemmlowp_reshaped_rhs_only_fused_output_stage_fixedpoint.cpp index 94f3c93166..4acb316a3c 100644 --- a/examples/gemm_tuner/cl_gemmlowp_reshaped_rhs_only_fused_output_stage_fixedpoint.cpp +++ b/examples/gemm_tuner/cl_gemmlowp_reshaped_rhs_only_fused_output_stage_fixedpoint.cpp @@ -25,23 +25,23 @@ #error "This example needs to be built with -DARM_COMPUTE_CL" #endif /* ARM_COMPUTE_CL */ -#include "CommonGemmExampleOptions.h" -#include "GemmTunerHelpers.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/KernelDescriptors.h" -#include "arm_compute/core/KernelDescriptors.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/core/utils/quantization/AsymmHelpers.h" #include "arm_compute/runtime/CL/CLScheduler.h" #include "arm_compute/runtime/CL/CLTuner.h" + #include "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel.h" #include "src/gpu/cl/kernels/ClGemmLowpReductionKernel.h" #include "tests/CL/Helper.h" -#include "utils/Utils.h" #include "utils/command_line/CommandLineOptions.h" #include "utils/command_line/CommandLineParser.h" +#include "utils/Utils.h" +#include "CommonGemmExampleOptions.h" +#include "GemmTunerHelpers.h" #include <cstdlib> #include <memory> @@ -56,12 +56,12 @@ namespace /** Structure holding all tunable gemm configs specific to this example/strategy */ struct GemmConfigs { - size_t m0{ 4 }; /**< Number of rows processed by the matrix multiplication */ - size_t n0{ 4 }; /**< Number of columns processed by the matrix multiplication */ - size_t k0{ 4 }; /**< Number of partial accumulations performed by the matrix multiplication */ - size_t h0{ 1 }; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ - bool interleave_rhs{ true }; /**< Interleave rhs matrix */ - bool transpose_rhs{ true }; /**< Transpose rhs matrix */ + size_t m0{4}; /**< Number of rows processed by the matrix multiplication */ + size_t n0{4}; /**< Number of columns processed by the matrix multiplication */ + size_t k0{4}; /**< Number of partial accumulations performed by the matrix multiplication */ + size_t h0{1}; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ + bool interleave_rhs{true}; /**< Interleave rhs matrix */ + bool transpose_rhs{true}; /**< Transpose rhs matrix */ }; /** Formatted output of the GemmConfigs type @@ -119,10 +119,10 @@ public: /** Default destructor */ ~GemmConfigOptions() = default; - SimpleOption<size_t> *m0; /**< Number of rows processed by the matrix multiplication option */ - SimpleOption<size_t> *n0; /**< Number of columns processed by the matrix multiplication option */ - SimpleOption<size_t> *k0; /**< Number of partial accumulations performed by the matrix multiplication option */ - SimpleOption<size_t> *h0; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row option */ + SimpleOption<size_t> *m0; /**< Number of rows processed by the matrix multiplication option */ + SimpleOption<size_t> *n0; /**< Number of columns processed by the matrix multiplication option */ + SimpleOption<size_t> *k0; /**< Number of partial accumulations performed by the matrix multiplication option */ + SimpleOption<size_t> *h0; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row option */ SimpleOption<size_t> *interleave_rhs; /**< Interleave rhs matrix option (1 enable; 0 disable) */ SimpleOption<size_t> *transpose_rhs; /**< Transpose rhs matrix option (1 enable; 0 disable) */ }; @@ -147,8 +147,9 @@ GemmConfigs consume_gemm_configs(const GemmConfigOptions &options) } // namespace -using ClGemmLowpMatrixMultiplyReshapedOnlyRhs = test::CLSynthetizeOperator<ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel>; -using ClGemmLowpMatrixAReduction = test::CLSynthetizeOperator<ClGemmLowpMatrixAReductionKernel>; +using ClGemmLowpMatrixMultiplyReshapedOnlyRhs = + test::CLSynthetizeOperator<ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel>; +using ClGemmLowpMatrixAReduction = test::CLSynthetizeOperator<ClGemmLowpMatrixAReductionKernel>; class CLGEMMLowpMatrixMultiplyReshapedOnlyRHSFusedOutputStageFixedpointExample : public Example { @@ -165,12 +166,12 @@ public: GemmConfigOptions config_options(parser); parser.parse(argc, argv); - if(param_options.help->is_set() && param_options.help->value()) + if (param_options.help->is_set() && param_options.help->value()) { parser.print_help(argv[0]); return false; } - if(!parser.validate()) + if (!parser.validate()) { // Invalid arguments. Use default parameters and configs std::cerr << "Invalid arguments." << std::endl; @@ -199,10 +200,7 @@ public: // Set arbitrary quantization information (non-zero offset to ensure offset contribution stage is included) // Could be extended in the future to include a user-controlled option for offset == 0 - const QuantizationInfo q_info - { - 0.012, 3 - }; + const QuantizationInfo q_info{0.012, 3}; lhs.info()->set_quantization_info(q_info); rhs.info()->set_quantization_info(q_info); bias.info()->set_quantization_info(q_info); @@ -220,16 +218,17 @@ public: rhs_info.transpose = configs.transpose_rhs; rhs_info.export_to_cl_image = false; // CL image not supported for quantized cases yet - if(rhs_info.h0 == 0) + if (rhs_info.h0 == 0) { rhs_info.h0 = std::max(static_cast<unsigned int>(params.N) / rhs_info.n0, 1U); } - rhs_reshaped.allocator()->init(TensorInfo(compute_rhs_reshaped_shape(*rhs.info(), rhs_info), 1, params.data_type)); + rhs_reshaped.allocator()->init( + TensorInfo(compute_rhs_reshaped_shape(*rhs.info(), rhs_info), 1, params.data_type)); rhs_reshaped.info()->set_quantization_info(q_info); - if(rhs_info.export_to_cl_image) + if (rhs_info.export_to_cl_image) { - if(!examples::gemm_tuner_helpers::update_padding_for_cl_image(rhs_reshaped.info())) + if (!examples::gemm_tuner_helpers::update_padding_for_cl_image(rhs_reshaped.info())) { std::cerr << "cl_image is not supported on the device, disable export_to_cl_image" << std::endl; return false; @@ -251,9 +250,7 @@ public: gemmlowp_output_stage.gemmlowp_multipliers.resize(num_filters); gemmlowp_output_stage.gemmlowp_shifts.resize(num_filters); - quantization::compute_quantized_multipliers_and_shifts(lhs.info(), - rhs.info(), - dst.info(), + quantization::compute_quantized_multipliers_and_shifts(lhs.info(), rhs.info(), dst.info(), gemmlowp_output_stage.gemmlowp_multipliers.data(), gemmlowp_output_stage.gemmlowp_shifts.data()); gemmlowp_output_stage.gemmlowp_multiplier = gemmlowp_output_stage.gemmlowp_multipliers[0]; @@ -290,14 +287,14 @@ public: gemm_info.output_stage = gemmlowp_output_stage; // Initialize Matrix A reduction kernel only if _b_offset is not equal to 0 - if(gemm_info.b_offset != 0) + if (gemm_info.b_offset != 0) { const TensorInfo info_vector_sum_row(compute_reductionB_shape(*lhs.info()), 1, DataType::S32); vector_sum_row.allocator()->init(info_vector_sum_row); mtx_a_reduction = std::make_unique<ClGemmLowpMatrixAReduction>(); - if(!mtx_a_reduction->validate(lhs.info(), vector_sum_row.info(), GEMMLowpReductionKernelInfo{})) + if (!mtx_a_reduction->validate(lhs.info(), vector_sum_row.info(), GEMMLowpReductionKernelInfo{})) { std::cerr << "Invalid arguments for CLGEMMLowpMatrixAReductionKernel." << std::endl; return false; @@ -306,7 +303,7 @@ public: mtx_a_reduction->configure(lhs.info(), vector_sum_row.info(), GEMMLowpReductionKernelInfo{}); } // Initialize matrix B reduction kernel only if _a_offset is not equal to 0 - if(gemm_info.a_offset != 0) + if (gemm_info.a_offset != 0) { const TensorInfo info_vector_sum_col(compute_reductionA_shape(*rhs.info()), 1, DataType::S32); vector_sum_col.allocator()->init(info_vector_sum_col); @@ -314,8 +311,10 @@ public: } // Validate argments - if(!gemm.validate(lhs.info(), rhs_reshaped.info(), dst.info(), gemm_info, gemm_info.a_offset == 0 ? nullptr : vector_sum_col.info(), - gemm_info.b_offset == 0 ? nullptr : vector_sum_row.info(), bias.info(), dst_multipliers.info(), dst_shifts.info())) + if (!gemm.validate(lhs.info(), rhs_reshaped.info(), dst.info(), gemm_info, + gemm_info.a_offset == 0 ? nullptr : vector_sum_col.info(), + gemm_info.b_offset == 0 ? nullptr : vector_sum_row.info(), bias.info(), + dst_multipliers.info(), dst_shifts.info())) { std::cerr << "Invalid arguments for ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel." << std::endl; return false; @@ -323,8 +322,9 @@ public: // Configure function gemm.configure(lhs.info(), rhs_reshaped.info(), dst.info(), gemm_info, - gemm_info.a_offset == 0 ? nullptr : vector_sum_col.info(), gemm_info.b_offset == 0 ? nullptr : vector_sum_row.info(), - bias.info(), dst_multipliers.info(), dst_shifts.info()); + gemm_info.a_offset == 0 ? nullptr : vector_sum_col.info(), + gemm_info.b_offset == 0 ? nullptr : vector_sum_row.info(), bias.info(), dst_multipliers.info(), + dst_shifts.info()); // Allocate tensors lhs.allocator()->allocate(); @@ -341,13 +341,20 @@ public: } void do_run() override { - if(mtx_a_reduction != nullptr) + if (mtx_a_reduction != nullptr) { - ITensorPack red_pack({ { ACL_SRC, &lhs }, { ACL_DST, &dst } }); + ITensorPack red_pack({{ACL_SRC, &lhs}, {ACL_DST, &dst}}); mtx_a_reduction->run(red_pack); } - ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, { ACL_SRC_1, &rhs }, { ACL_BIAS, &bias }, { ACL_VEC_COL_SUM, &vector_sum_col }, { ACL_VEC_ROW_SUM, &vector_sum_row }, { ACL_SHIFTS, &dst_shifts }, { ACL_MULTIPLIERS, &dst_multipliers }, { ACL_DST, &dst } }); + ITensorPack gemm_pack({{ACL_SRC_0, &lhs}, + {ACL_SRC_1, &rhs}, + {ACL_BIAS, &bias}, + {ACL_VEC_COL_SUM, &vector_sum_col}, + {ACL_VEC_ROW_SUM, &vector_sum_row}, + {ACL_SHIFTS, &dst_shifts}, + {ACL_MULTIPLIERS, &dst_multipliers}, + {ACL_DST, &dst}}); gemm.run(gemm_pack); // Make sure all the OpenCL jobs are done executing: @@ -370,7 +377,7 @@ private: CLTensor dst_shifts{}; CLTuner tuner{}; ClGemmLowpMatrixMultiplyReshapedOnlyRhs gemm{}; - std::unique_ptr<ClGemmLowpMatrixAReduction> mtx_a_reduction{ nullptr }; + std::unique_ptr<ClGemmLowpMatrixAReduction> mtx_a_reduction{nullptr}; }; /** Main test program for gemmlowp reshaped rhs only with fused output stage fixedpoint diff --git a/examples/graph_alexnet.cpp b/examples/graph_alexnet.cpp index 53a4547e04..be0b8a7d8a 100644 --- a/examples/graph_alexnet.cpp +++ b/examples/graph_alexnet.cpp @@ -39,8 +39,7 @@ using namespace arm_compute::graph_utils; class GraphAlexnetExample : public Example { public: - GraphAlexnetExample() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "AlexNet") + GraphAlexnetExample() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "AlexNet") { } bool do_setup(int argc, char **argv) override @@ -53,14 +52,15 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; } // Checks - ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph"); + ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), + "QASYMM8 not supported for this graph"); // Print parameter values std::cout << common_params << std::endl; @@ -69,88 +69,80 @@ public: std::string data_path = common_params.data_path; // Create a preprocessor object - const std::array<float, 3> mean_rgb{ { 122.68f, 116.67f, 104.01f } }; + const std::array<float, 3> mean_rgb{{122.68f, 116.67f, 104.01f}}; std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<CaffePreproccessor>(mean_rgb); // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(227U, 227U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(227U, 227U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint - << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor))) - // Layer 1 - << ConvolutionLayer( - 11U, 11U, 96U, - get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv1_b.npy"), - PadStrideInfo(4, 4, 0, 0)) - .set_name("conv1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu1") - << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)).set_name("norm1") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool1") - // Layer 2 - << ConvolutionLayer( - 5U, 5U, 256U, - get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv2_b.npy"), - PadStrideInfo(1, 1, 2, 2), 2) - .set_name("conv2") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu2") - << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)).set_name("norm2") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool2") - // Layer 3 - << ConvolutionLayer( - 3U, 3U, 384U, - get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv3_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv3_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv3") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu3") - // Layer 4 - << ConvolutionLayer( - 3U, 3U, 384U, - get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv4_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv4_b.npy"), - PadStrideInfo(1, 1, 1, 1), 2) - .set_name("conv4") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu4") - // Layer 5 - << ConvolutionLayer( - 3U, 3U, 256U, - get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv5_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv5_b.npy"), - PadStrideInfo(1, 1, 1, 1), 2) - .set_name("conv5") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu5") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool5") - // Layer 6 - << FullyConnectedLayer( - 4096U, - get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc6_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc6_b.npy")) - .set_name("fc6") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu6") - // Layer 7 - << FullyConnectedLayer( - 4096U, - get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc7_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc7_b.npy")) - .set_name("fc7") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu7") - // Layer 8 - << FullyConnectedLayer( - 1000U, - get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc8_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc8_b.npy")) - .set_name("fc8") - // Softmax - << SoftmaxLayer().set_name("prob") - << OutputLayer(get_output_accessor(common_params, 5)); + graph + << common_params.target << common_params.fast_math_hint + << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor))) + // Layer 1 + << ConvolutionLayer(11U, 11U, 96U, + get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv1_b.npy"), + PadStrideInfo(4, 4, 0, 0)) + .set_name("conv1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu1") + << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)).set_name("norm1") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool1") + // Layer 2 + << ConvolutionLayer( + 5U, 5U, 256U, get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv2_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv2_b.npy"), PadStrideInfo(1, 1, 2, 2), 2) + .set_name("conv2") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu2") + << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)).set_name("norm2") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool2") + // Layer 3 + << ConvolutionLayer( + 3U, 3U, 384U, get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv3_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv3_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv3") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu3") + // Layer 4 + << ConvolutionLayer( + 3U, 3U, 384U, get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv4_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv4_b.npy"), PadStrideInfo(1, 1, 1, 1), 2) + .set_name("conv4") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu4") + // Layer 5 + << ConvolutionLayer( + 3U, 3U, 256U, get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv5_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv5_b.npy"), PadStrideInfo(1, 1, 1, 1), 2) + .set_name("conv5") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu5") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool5") + // Layer 6 + << FullyConnectedLayer(4096U, + get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc6_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc6_b.npy")) + .set_name("fc6") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu6") + // Layer 7 + << FullyConnectedLayer(4096U, + get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc7_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc7_b.npy")) + .set_name("fc7") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu7") + // Layer 8 + << FullyConnectedLayer(1000U, + get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc8_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc8_b.npy")) + .set_name("fc8") + // Softmax + << SoftmaxLayer().set_name("prob") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph GraphConfig config; @@ -163,7 +155,7 @@ public: // Load the precompiled kernels from a file into the kernel library, in this way the next time they are needed // compilation won't be required. - if(common_params.enable_cl_cache) + if (common_params.enable_cl_cache) { #ifdef ARM_COMPUTE_CL restore_program_cache_from_file(); @@ -173,7 +165,7 @@ public: graph.finalize(common_params.target, config); // Save the opencl kernels to a file - if(common_opts.enable_cl_cache) + if (common_opts.enable_cl_cache) { #ifdef ARM_COMPUTE_CL save_program_cache_to_file(); diff --git a/examples/graph_deepspeech_v0_4_1.cpp b/examples/graph_deepspeech_v0_4_1.cpp index da163b6493..08cd4a47b1 100644 --- a/examples/graph_deepspeech_v0_4_1.cpp +++ b/examples/graph_deepspeech_v0_4_1.cpp @@ -23,6 +23,7 @@ */ #include "arm_compute/graph.h" #include "arm_compute/graph/Types.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -37,8 +38,7 @@ using namespace arm_compute::graph_utils; class GraphDeepSpeechExample : public Example { public: - GraphDeepSpeechExample() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "DeepSpeech v0.4.1") + GraphDeepSpeechExample() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "DeepSpeech v0.4.1") { } bool do_setup(int argc, char **argv) override @@ -51,7 +51,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -64,7 +64,7 @@ public: std::string data_path = common_params.data_path; const std::string model_path = "/cnn_data/deepspeech_model/"; - if(!data_path.empty()) + if (!data_path.empty()) { data_path += model_path; } @@ -77,131 +77,131 @@ public: const float cell_clip = 20.f; // Create input descriptor - const TensorShape tensor_shape = permute_shape(TensorShape(26U, 19U, n_steps, 1U), DataLayout::NHWC, common_params.data_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(26U, 19U, n_steps, 1U), DataLayout::NHWC, common_params.data_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NHWC; - graph << common_params.target - << common_params.fast_math_hint + graph << common_params.target << common_params.fast_math_hint << InputLayer(input_descriptor, - get_weights_accessor(data_path, "input_values_x" + std::to_string(n_steps) + ".npy", weights_layout)) - .set_name("input_node"); + get_weights_accessor(data_path, "input_values_x" + std::to_string(n_steps) + ".npy", + weights_layout)) + .set_name("input_node"); - if(common_params.data_layout == DataLayout::NCHW) + if (common_params.data_layout == DataLayout::NCHW) { graph << PermuteLayer(PermutationVector(2U, 0U, 1U), common_params.data_layout).set_name("permute_to_nhwc"); } graph << ReshapeLayer(TensorShape(494U, n_steps)).set_name("Reshape_input") // Layer 1 - << FullyConnectedLayer( - 2048U, - get_weights_accessor(data_path, "h1_transpose.npy", weights_layout), - get_weights_accessor(data_path, "MatMul_bias.npy")) - .set_name("fc0") + << FullyConnectedLayer(2048U, get_weights_accessor(data_path, "h1_transpose.npy", weights_layout), + get_weights_accessor(data_path, "MatMul_bias.npy")) + .set_name("fc0") << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, cell_clip)) - .set_name("Relu") + .set_name("Relu") // Layer 2 - << FullyConnectedLayer( - 2048U, - get_weights_accessor(data_path, "h2_transpose.npy", weights_layout), - get_weights_accessor(data_path, "MatMul_1_bias.npy")) - .set_name("fc1") + << FullyConnectedLayer(2048U, get_weights_accessor(data_path, "h2_transpose.npy", weights_layout), + get_weights_accessor(data_path, "MatMul_1_bias.npy")) + .set_name("fc1") << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, cell_clip)) - .set_name("Relu_1") + .set_name("Relu_1") // Layer 3 - << FullyConnectedLayer( - 2048U, - get_weights_accessor(data_path, "h3_transpose.npy", weights_layout), - get_weights_accessor(data_path, "MatMul_2_bias.npy")) - .set_name("fc2") + << FullyConnectedLayer(2048U, get_weights_accessor(data_path, "h3_transpose.npy", weights_layout), + get_weights_accessor(data_path, "MatMul_2_bias.npy")) + .set_name("fc2") << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, cell_clip)) - .set_name("Relu_2") + .set_name("Relu_2") // Layer 4 << ReshapeLayer(TensorShape(2048U, 1U, n_steps)).set_name("Reshape_1"); // Unstack Layer (using SplitLayerNode) - NodeParams unstack_params = { "unstack", graph.hints().target_hint }; - NodeID unstack_nid = GraphBuilder::add_split_node(graph.graph(), unstack_params, { graph.tail_node(), 0 }, n_steps, 2); + NodeParams unstack_params = {"unstack", graph.hints().target_hint}; + NodeID unstack_nid = + GraphBuilder::add_split_node(graph.graph(), unstack_params, {graph.tail_node(), 0}, n_steps, 2); // Create input state descriptor - TensorDescriptor state_descriptor = TensorDescriptor(TensorShape(2048U), common_params.data_type).set_layout(common_params.data_layout); - SubStream previous_state(graph); - SubStream add_y(graph); + TensorDescriptor state_descriptor = + TensorDescriptor(TensorShape(2048U), common_params.data_type).set_layout(common_params.data_layout); + SubStream previous_state(graph); + SubStream add_y(graph); // Initial state for LSTM is all zeroes for both state_h and state_c, therefore only one input is created - previous_state << InputLayer(state_descriptor, - get_weights_accessor(data_path, "zeros.npy")) - .set_name("previous_state_c_h"); - add_y << InputLayer(state_descriptor, - get_weights_accessor(data_path, "ones.npy")) - .set_name("add_y"); + previous_state << InputLayer(state_descriptor, get_weights_accessor(data_path, "zeros.npy")) + .set_name("previous_state_c_h"); + add_y << InputLayer(state_descriptor, get_weights_accessor(data_path, "ones.npy")).set_name("add_y"); // Create LSTM Fully Connected weights and bias descriptors - TensorDescriptor lstm_weights_descriptor = TensorDescriptor(TensorShape(4096U, 8192U), common_params.data_type).set_layout(common_params.data_layout); - TensorDescriptor lstm_bias_descriptor = TensorDescriptor(TensorShape(8192U), common_params.data_type).set_layout(common_params.data_layout); - SubStream lstm_fc_weights(graph); - SubStream lstm_fc_bias(graph); - lstm_fc_weights << ConstantLayer(lstm_weights_descriptor, - get_weights_accessor(data_path, "rnn_lstm_cell_kernel_transpose.npy", weights_layout)) - .set_name("h5/transpose"); + TensorDescriptor lstm_weights_descriptor = + TensorDescriptor(TensorShape(4096U, 8192U), common_params.data_type).set_layout(common_params.data_layout); + TensorDescriptor lstm_bias_descriptor = + TensorDescriptor(TensorShape(8192U), common_params.data_type).set_layout(common_params.data_layout); + SubStream lstm_fc_weights(graph); + SubStream lstm_fc_bias(graph); + lstm_fc_weights << ConstantLayer( + lstm_weights_descriptor, + get_weights_accessor(data_path, "rnn_lstm_cell_kernel_transpose.npy", weights_layout)) + .set_name("h5/transpose"); lstm_fc_bias << ConstantLayer(lstm_bias_descriptor, get_weights_accessor(data_path, "rnn_lstm_cell_MatMul_bias.npy")) - .set_name("MatMul_3_bias"); + .set_name("MatMul_3_bias"); // LSTM Block - std::pair<SubStream, SubStream> new_state_1 = add_lstm_cell(unstack_nid, 0, previous_state, previous_state, add_y, lstm_fc_weights, lstm_fc_bias); - std::pair<SubStream, SubStream> new_state_2 = add_lstm_cell(unstack_nid, 1, new_state_1.first, new_state_1.second, add_y, lstm_fc_weights, lstm_fc_bias); - std::pair<SubStream, SubStream> new_state_3 = add_lstm_cell(unstack_nid, 2, new_state_2.first, new_state_2.second, add_y, lstm_fc_weights, lstm_fc_bias); - std::pair<SubStream, SubStream> new_state_4 = add_lstm_cell(unstack_nid, 3, new_state_3.first, new_state_3.second, add_y, lstm_fc_weights, lstm_fc_bias); - std::pair<SubStream, SubStream> new_state_5 = add_lstm_cell(unstack_nid, 4, new_state_4.first, new_state_4.second, add_y, lstm_fc_weights, lstm_fc_bias); - std::pair<SubStream, SubStream> new_state_6 = add_lstm_cell(unstack_nid, 5, new_state_5.first, new_state_5.second, add_y, lstm_fc_weights, lstm_fc_bias); - std::pair<SubStream, SubStream> new_state_7 = add_lstm_cell(unstack_nid, 6, new_state_6.first, new_state_6.second, add_y, lstm_fc_weights, lstm_fc_bias); - std::pair<SubStream, SubStream> new_state_8 = add_lstm_cell(unstack_nid, 7, new_state_7.first, new_state_7.second, add_y, lstm_fc_weights, lstm_fc_bias); - std::pair<SubStream, SubStream> new_state_9 = add_lstm_cell(unstack_nid, 8, new_state_8.first, new_state_8.second, add_y, lstm_fc_weights, lstm_fc_bias); - std::pair<SubStream, SubStream> new_state_10 = add_lstm_cell(unstack_nid, 9, new_state_9.first, new_state_9.second, add_y, lstm_fc_weights, lstm_fc_bias); - std::pair<SubStream, SubStream> new_state_11 = add_lstm_cell(unstack_nid, 10, new_state_10.first, new_state_10.second, add_y, lstm_fc_weights, lstm_fc_bias); - std::pair<SubStream, SubStream> new_state_12 = add_lstm_cell(unstack_nid, 11, new_state_11.first, new_state_11.second, add_y, lstm_fc_weights, lstm_fc_bias); - std::pair<SubStream, SubStream> new_state_13 = add_lstm_cell(unstack_nid, 12, new_state_12.first, new_state_12.second, add_y, lstm_fc_weights, lstm_fc_bias); - std::pair<SubStream, SubStream> new_state_14 = add_lstm_cell(unstack_nid, 13, new_state_13.first, new_state_13.second, add_y, lstm_fc_weights, lstm_fc_bias); - std::pair<SubStream, SubStream> new_state_15 = add_lstm_cell(unstack_nid, 14, new_state_14.first, new_state_14.second, add_y, lstm_fc_weights, lstm_fc_bias); - std::pair<SubStream, SubStream> new_state_16 = add_lstm_cell(unstack_nid, 15, new_state_15.first, new_state_15.second, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_1 = + add_lstm_cell(unstack_nid, 0, previous_state, previous_state, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_2 = + add_lstm_cell(unstack_nid, 1, new_state_1.first, new_state_1.second, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_3 = + add_lstm_cell(unstack_nid, 2, new_state_2.first, new_state_2.second, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_4 = + add_lstm_cell(unstack_nid, 3, new_state_3.first, new_state_3.second, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_5 = + add_lstm_cell(unstack_nid, 4, new_state_4.first, new_state_4.second, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_6 = + add_lstm_cell(unstack_nid, 5, new_state_5.first, new_state_5.second, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_7 = + add_lstm_cell(unstack_nid, 6, new_state_6.first, new_state_6.second, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_8 = + add_lstm_cell(unstack_nid, 7, new_state_7.first, new_state_7.second, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_9 = + add_lstm_cell(unstack_nid, 8, new_state_8.first, new_state_8.second, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_10 = + add_lstm_cell(unstack_nid, 9, new_state_9.first, new_state_9.second, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_11 = add_lstm_cell( + unstack_nid, 10, new_state_10.first, new_state_10.second, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_12 = add_lstm_cell( + unstack_nid, 11, new_state_11.first, new_state_11.second, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_13 = add_lstm_cell( + unstack_nid, 12, new_state_12.first, new_state_12.second, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_14 = add_lstm_cell( + unstack_nid, 13, new_state_13.first, new_state_13.second, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_15 = add_lstm_cell( + unstack_nid, 14, new_state_14.first, new_state_14.second, add_y, lstm_fc_weights, lstm_fc_bias); + std::pair<SubStream, SubStream> new_state_16 = add_lstm_cell( + unstack_nid, 15, new_state_15.first, new_state_15.second, add_y, lstm_fc_weights, lstm_fc_bias); // Concatenate new states on height const int axis = 1; - graph << StackLayer(axis, - std::move(new_state_1.second), - std::move(new_state_2.second), - std::move(new_state_3.second), - std::move(new_state_4.second), - std::move(new_state_5.second), - std::move(new_state_6.second), - std::move(new_state_7.second), - std::move(new_state_8.second), - std::move(new_state_9.second), - std::move(new_state_10.second), - std::move(new_state_11.second), - std::move(new_state_12.second), - std::move(new_state_13.second), - std::move(new_state_14.second), - std::move(new_state_15.second), - std::move(new_state_16.second)) - .set_name("concat"); - - graph << FullyConnectedLayer( - 2048U, - get_weights_accessor(data_path, "h5_transpose.npy", weights_layout), - get_weights_accessor(data_path, "MatMul_3_bias.npy")) - .set_name("fc3") + graph << StackLayer(axis, std::move(new_state_1.second), std::move(new_state_2.second), + std::move(new_state_3.second), std::move(new_state_4.second), std::move(new_state_5.second), + std::move(new_state_6.second), std::move(new_state_7.second), std::move(new_state_8.second), + std::move(new_state_9.second), std::move(new_state_10.second), + std::move(new_state_11.second), std::move(new_state_12.second), + std::move(new_state_13.second), std::move(new_state_14.second), + std::move(new_state_15.second), std::move(new_state_16.second)) + .set_name("concat"); + + graph << FullyConnectedLayer(2048U, get_weights_accessor(data_path, "h5_transpose.npy", weights_layout), + get_weights_accessor(data_path, "MatMul_3_bias.npy")) + .set_name("fc3") << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, cell_clip)) - .set_name("Relu3") - << FullyConnectedLayer( - 29U, - get_weights_accessor(data_path, "h6_transpose.npy", weights_layout), - get_weights_accessor(data_path, "MatMul_4_bias.npy")) - .set_name("fc3") + .set_name("Relu3") + << FullyConnectedLayer(29U, get_weights_accessor(data_path, "h6_transpose.npy", weights_layout), + get_weights_accessor(data_path, "MatMul_4_bias.npy")) + .set_name("fc3") << SoftmaxLayer().set_name("logits"); graph << OutputLayer(get_output_accessor(common_params, 5)); @@ -241,7 +241,7 @@ private: return Status{}; } - std::pair<SubStream, SubStream> add_lstm_cell(NodeID unstack_nid, + std::pair<SubStream, SubStream> add_lstm_cell(NodeID unstack_nid, unsigned int unstack_idx, SubStream previous_state_c, SubStream previous_state_h, @@ -250,41 +250,41 @@ private: SubStream lstm_fc_bias) { const std::string cell_name("rnn/lstm_cell_" + std::to_string(unstack_idx)); - const DataLayoutDimension concat_dim = (common_params.data_layout == DataLayout::NHWC) ? DataLayoutDimension::CHANNEL : DataLayoutDimension::WIDTH; + const DataLayoutDimension concat_dim = + (common_params.data_layout == DataLayout::NHWC) ? DataLayoutDimension::CHANNEL : DataLayoutDimension::WIDTH; // Concatenate result of Unstack with previous_state_h - NodeParams concat_params = { cell_name + "/concat", graph.hints().target_hint }; + NodeParams concat_params = {cell_name + "/concat", graph.hints().target_hint}; NodeID concat_nid = graph.graph().add_node<ConcatenateLayerNode>(2, concat_dim); graph.graph().add_connection(unstack_nid, unstack_idx, concat_nid, 0); graph.graph().add_connection(previous_state_h.tail_node(), 0, concat_nid, 1); set_node_params(graph.graph(), concat_nid, concat_params); graph.forward_tail(concat_nid); - graph << FullyConnectedLayer( - 8192U, - lstm_fc_weights, - lstm_fc_bias) - .set_name(cell_name + "/BiasAdd"); + graph << FullyConnectedLayer(8192U, lstm_fc_weights, lstm_fc_bias).set_name(cell_name + "/BiasAdd"); // Split Layer const unsigned int num_splits = 4; const unsigned int split_axis = 0; - NodeParams split_params = { cell_name + "/split", graph.hints().target_hint }; - NodeID split_nid = GraphBuilder::add_split_node(graph.graph(), split_params, { graph.tail_node(), 0 }, num_splits, split_axis); + NodeParams split_params = {cell_name + "/split", graph.hints().target_hint}; + NodeID split_nid = + GraphBuilder::add_split_node(graph.graph(), split_params, {graph.tail_node(), 0}, num_splits, split_axis); - NodeParams sigmoid_1_params = { cell_name + "/Sigmoid_1", graph.hints().target_hint }; - NodeParams add_params = { cell_name + "/add", graph.hints().target_hint }; - NodeParams sigmoid_2_params = { cell_name + "/Sigmoid_2", graph.hints().target_hint }; - NodeParams tanh_params = { cell_name + "/Tanh", graph.hints().target_hint }; + NodeParams sigmoid_1_params = {cell_name + "/Sigmoid_1", graph.hints().target_hint}; + NodeParams add_params = {cell_name + "/add", graph.hints().target_hint}; + NodeParams sigmoid_2_params = {cell_name + "/Sigmoid_2", graph.hints().target_hint}; + NodeParams tanh_params = {cell_name + "/Tanh", graph.hints().target_hint}; // Sigmoid 1 (first split) - NodeID sigmoid_1_nid = graph.graph().add_node<ActivationLayerNode>(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); + NodeID sigmoid_1_nid = graph.graph().add_node<ActivationLayerNode>( + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); graph.graph().add_connection(split_nid, 0, sigmoid_1_nid, 0); set_node_params(graph.graph(), sigmoid_1_nid, sigmoid_1_params); // Tanh (second split) - NodeID tanh_nid = graph.graph().add_node<ActivationLayerNode>(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f)); + NodeID tanh_nid = graph.graph().add_node<ActivationLayerNode>( + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f)); graph.graph().add_connection(split_nid, 1, tanh_nid, 0); set_node_params(graph.graph(), tanh_nid, tanh_params); @@ -292,13 +292,15 @@ private: tanh_ss.forward_tail(tanh_nid); // Add (third split) - NodeID add_nid = graph.graph().add_node<EltwiseLayerNode>(descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add }); + NodeID add_nid = + graph.graph().add_node<EltwiseLayerNode>(descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add}); graph.graph().add_connection(split_nid, 2, add_nid, 0); graph.graph().add_connection(add_y.tail_node(), 0, add_nid, 1); set_node_params(graph.graph(), add_nid, add_params); // Sigmoid 2 (fourth split) - NodeID sigmoid_2_nid = graph.graph().add_node<ActivationLayerNode>(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); + NodeID sigmoid_2_nid = graph.graph().add_node<ActivationLayerNode>( + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); graph.graph().add_connection(split_nid, 3, sigmoid_2_nid, 0); set_node_params(graph.graph(), sigmoid_2_nid, sigmoid_2_params); @@ -306,28 +308,28 @@ private: sigmoid_1_ss.forward_tail(sigmoid_1_nid); SubStream mul_1_ss(sigmoid_1_ss); mul_1_ss << EltwiseLayer(std::move(sigmoid_1_ss), std::move(tanh_ss), EltwiseOperation::Mul) - .set_name(cell_name + "/mul_1"); + .set_name(cell_name + "/mul_1"); SubStream tanh_1_ss_tmp(graph); tanh_1_ss_tmp.forward_tail(add_nid); tanh_1_ss_tmp << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)) - .set_name(cell_name + "/Sigmoid"); + .set_name(cell_name + "/Sigmoid"); SubStream tanh_1_ss_tmp2(tanh_1_ss_tmp); tanh_1_ss_tmp2 << EltwiseLayer(std::move(tanh_1_ss_tmp), std::move(previous_state_c), EltwiseOperation::Mul) - .set_name(cell_name + "/mul"); + .set_name(cell_name + "/mul"); SubStream tanh_1_ss(tanh_1_ss_tmp2); tanh_1_ss << EltwiseLayer(std::move(tanh_1_ss_tmp2), std::move(mul_1_ss), EltwiseOperation::Add) - .set_name(cell_name + "/new_state_c"); + .set_name(cell_name + "/new_state_c"); SubStream new_state_c(tanh_1_ss); tanh_1_ss << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f)) - .set_name(cell_name + "/Tanh_1"); + .set_name(cell_name + "/Tanh_1"); SubStream sigmoid_2_ss(graph); sigmoid_2_ss.forward_tail(sigmoid_2_nid); graph << EltwiseLayer(std::move(sigmoid_2_ss), std::move(tanh_1_ss), EltwiseOperation::Mul) - .set_name(cell_name + "/new_state_h"); + .set_name(cell_name + "/new_state_h"); SubStream new_state_h(graph); return std::pair<SubStream, SubStream>(new_state_c, new_state_h); diff --git a/examples/graph_edsr.cpp b/examples/graph_edsr.cpp index 0e41f12155..b4f2fadf4a 100644 --- a/examples/graph_edsr.cpp +++ b/examples/graph_edsr.cpp @@ -22,28 +22,28 @@ * SOFTWARE. */ +#include "graph_edsr.h" + #include "arm_compute/graph/Utils.h" #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/Utils.h" -#include "graph_edsr.h" - using namespace arm_compute::graph; using namespace arm_compute::utils; class GraphEdsrExample : public Example { public: - GraphEdsrExample() - : cmd_parser(), common_opts(cmd_parser), common_params() + GraphEdsrExample() : cmd_parser(), common_opts(cmd_parser), common_params() { expected_output_filename = cmd_parser.add_option<SimpleOption<std::string>>("expected-output-filename", ""); - expected_output_filename->set_help("Name of npy file containing the expected output to validate the graph output."); + expected_output_filename->set_help( + "Name of npy file containing the expected output to validate the graph output."); } - GraphEdsrExample(const GraphEdsrExample &) = delete; + GraphEdsrExample(const GraphEdsrExample &) = delete; GraphEdsrExample &operator=(const GraphEdsrExample &) = delete; ~GraphEdsrExample() override = default; @@ -57,13 +57,14 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; } - ARM_COMPUTE_EXIT_ON_MSG(common_params.data_type != DataType::QASYMM8, "Only QASYMM8 is supported for this graph example"); + ARM_COMPUTE_EXIT_ON_MSG(common_params.data_type != DataType::QASYMM8, + "Only QASYMM8 is supported for this graph example"); // Print parameter values std::cout << common_params << std::endl; @@ -98,7 +99,7 @@ private: GraphContext context{}; GraphManager manager{}; - SimpleOption<std::string> *expected_output_filename{ nullptr }; + SimpleOption<std::string> *expected_output_filename{nullptr}; GraphEdsr model{}; }; diff --git a/examples/graph_edsr.h b/examples/graph_edsr.h index 72012afdcb..1161e4ba38 100644 --- a/examples/graph_edsr.h +++ b/examples/graph_edsr.h @@ -32,12 +32,12 @@ class GraphEdsr { public: - GraphEdsr() - : _graph(0, "EDSR") + GraphEdsr() : _graph(0, "EDSR") { } - bool setup(const arm_compute::utils::CommonGraphParams &common_params, const arm_compute::utils::SimpleOption<std::string> &expected_output_filename) + bool setup(const arm_compute::utils::CommonGraphParams &common_params, + const arm_compute::utils::SimpleOption<std::string> &expected_output_filename) { using namespace arm_compute; using namespace arm_compute::graph; @@ -47,1221 +47,879 @@ public: const auto &data_path = common_params.data_path; const auto &target = common_params.target; - NodeID id_upscale_net_FakeQuantWithMinMaxVars_transposed = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 12, 2, 2, 3 }, - DataType::QASYMM8, - QuantizationInfo(0.00393533194437623, 1), - DataLayout::NHWC }); - INode *node_upscale_net_FakeQuantWithMinMaxVars_transposed = _graph.node(id_upscale_net_FakeQuantWithMinMaxVars_transposed); - node_upscale_net_FakeQuantWithMinMaxVars_transposed->set_common_node_parameters(NodeParams{ "upscale_net_FakeQuantWithMinMaxVars_transposed", target }); - node_upscale_net_FakeQuantWithMinMaxVars_transposed->output(0)->set_accessor(get_weights_accessor(data_path, - "/cnn_data/edsr_model/upscale_net_FakeQuantWithMinMaxVars_transposed.npy", DataLayout::NHWC)); - - NodeID id_pre_upscale_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 12 }, - DataType::S32, - QuantizationInfo(2.9644968435604824e-06), - DataLayout::NHWC }); + NodeID id_upscale_net_FakeQuantWithMinMaxVars_transposed = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{12, 2, 2, 3}, DataType::QASYMM8, QuantizationInfo(0.00393533194437623, 1), DataLayout::NHWC}); + INode *node_upscale_net_FakeQuantWithMinMaxVars_transposed = + _graph.node(id_upscale_net_FakeQuantWithMinMaxVars_transposed); + node_upscale_net_FakeQuantWithMinMaxVars_transposed->set_common_node_parameters( + NodeParams{"upscale_net_FakeQuantWithMinMaxVars_transposed", target}); + node_upscale_net_FakeQuantWithMinMaxVars_transposed->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/upscale_net_FakeQuantWithMinMaxVars_transposed.npy", DataLayout::NHWC)); + + NodeID id_pre_upscale_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{12}, DataType::S32, QuantizationInfo(2.9644968435604824e-06), DataLayout::NHWC}); INode *node_pre_upscale_Conv2D_bias = _graph.node(id_pre_upscale_Conv2D_bias); - node_pre_upscale_Conv2D_bias->set_common_node_parameters(NodeParams{ "pre_upscale_Conv2D_bias", target }); - node_pre_upscale_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/pre_upscale_Conv2D_bias.npy", DataLayout::NHWC)); - - NodeID id_pre_upscale_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 12 }, - DataType::QASYMM8, - QuantizationInfo(0.000455576169770211, 128), - DataLayout::NHWC }); + node_pre_upscale_Conv2D_bias->set_common_node_parameters(NodeParams{"pre_upscale_Conv2D_bias", target}); + node_pre_upscale_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/pre_upscale_Conv2D_bias.npy", DataLayout::NHWC)); + + NodeID id_pre_upscale_FakeQuantWithMinMaxVars = + _graph.add_node<ConstNode>(TensorDescriptor{TensorShape{256, 3, 3, 12}, DataType::QASYMM8, + QuantizationInfo(0.000455576169770211, 128), DataLayout::NHWC}); INode *node_pre_upscale_FakeQuantWithMinMaxVars = _graph.node(id_pre_upscale_FakeQuantWithMinMaxVars); - node_pre_upscale_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "pre_upscale_FakeQuantWithMinMaxVars", target }); - node_pre_upscale_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/pre_upscale_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_post_residual_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.2760000345224398e-06), - DataLayout::NHWC }); + node_pre_upscale_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"pre_upscale_FakeQuantWithMinMaxVars", target}); + node_pre_upscale_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/pre_upscale_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_post_residual_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.2760000345224398e-06), DataLayout::NHWC}); INode *node_post_residual_Conv2D_bias = _graph.node(id_post_residual_Conv2D_bias); - node_post_residual_Conv2D_bias->set_common_node_parameters(NodeParams{ "post_residual_Conv2D_bias", target }); - node_post_residual_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/post_residual_Conv2D_bias.npy", DataLayout::NHWC)); + node_post_residual_Conv2D_bias->set_common_node_parameters(NodeParams{"post_residual_Conv2D_bias", target}); + node_post_residual_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/post_residual_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_post_residual_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.00036424631252884865, 129), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.00036424631252884865, 129), DataLayout::NHWC}); INode *node_post_residual_FakeQuantWithMinMaxVars = _graph.node(id_post_residual_FakeQuantWithMinMaxVars); - node_post_residual_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "post_residual_FakeQuantWithMinMaxVars", target }); - node_post_residual_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/post_residual_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_15_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_post_residual_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"post_residual_FakeQuantWithMinMaxVars", target}); + node_post_residual_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/post_residual_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_15_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_15_y = _graph.node(id_mul_15_y); - node_mul_15_y->set_common_node_parameters(NodeParams{ "mul_15_y", target }); - node_mul_15_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_15_y.npy", DataLayout::NHWC)); - - NodeID id_block_15_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.2441644230420934e-06), - DataLayout::NHWC }); + node_mul_15_y->set_common_node_parameters(NodeParams{"mul_15_y", target}); + node_mul_15_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_15_y.npy", DataLayout::NHWC)); + + NodeID id_block_15_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.2441644230420934e-06), DataLayout::NHWC}); INode *node_block_15_1_Conv2D_bias = _graph.node(id_block_15_1_Conv2D_bias); - node_block_15_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_15_1_Conv2D_bias", target }); - node_block_15_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_15_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_15_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_15_1_Conv2D_bias", target}); + node_block_15_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_15_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_15_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.00037038681330159307, 125), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.00037038681330159307, 125), DataLayout::NHWC}); INode *node_block_15_1_FakeQuantWithMinMaxVars = _graph.node(id_block_15_1_FakeQuantWithMinMaxVars); - node_block_15_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_15_1_FakeQuantWithMinMaxVars", target }); - node_block_15_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_15_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_14_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_block_15_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_15_1_FakeQuantWithMinMaxVars", target}); + node_block_15_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_15_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_14_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_14_y = _graph.node(id_mul_14_y); - node_mul_14_y->set_common_node_parameters(NodeParams{ "mul_14_y", target }); - node_mul_14_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_14_y.npy", DataLayout::NHWC)); - - NodeID id_block_14_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.3417260333881131e-06), - DataLayout::NHWC }); + node_mul_14_y->set_common_node_parameters(NodeParams{"mul_14_y", target}); + node_mul_14_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_14_y.npy", DataLayout::NHWC)); + + NodeID id_block_14_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.3417260333881131e-06), DataLayout::NHWC}); INode *node_block_14_1_Conv2D_bias = _graph.node(id_block_14_1_Conv2D_bias); - node_block_14_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_14_1_Conv2D_bias", target }); - node_block_14_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_14_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_14_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_14_1_Conv2D_bias", target}); + node_block_14_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_14_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_14_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.00040307495510205626, 127), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.00040307495510205626, 127), DataLayout::NHWC}); INode *node_block_14_1_FakeQuantWithMinMaxVars = _graph.node(id_block_14_1_FakeQuantWithMinMaxVars); - node_block_14_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_14_1_FakeQuantWithMinMaxVars", target }); - node_block_14_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_14_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_13_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_block_14_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_14_1_FakeQuantWithMinMaxVars", target}); + node_block_14_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_14_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_13_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_13_y = _graph.node(id_mul_13_y); - node_mul_13_y->set_common_node_parameters(NodeParams{ "mul_13_y", target }); - node_mul_13_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_13_y.npy", DataLayout::NHWC)); - - NodeID id_block_13_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.2636977544389083e-06), - DataLayout::NHWC }); + node_mul_13_y->set_common_node_parameters(NodeParams{"mul_13_y", target}); + node_mul_13_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_13_y.npy", DataLayout::NHWC)); + + NodeID id_block_13_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.2636977544389083e-06), DataLayout::NHWC}); INode *node_block_13_1_Conv2D_bias = _graph.node(id_block_13_1_Conv2D_bias); - node_block_13_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_13_1_Conv2D_bias", target }); - node_block_13_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_13_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_13_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_13_1_Conv2D_bias", target}); + node_block_13_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_13_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_13_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.0003858553245663643, 131), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.0003858553245663643, 131), DataLayout::NHWC}); INode *node_block_13_1_FakeQuantWithMinMaxVars = _graph.node(id_block_13_1_FakeQuantWithMinMaxVars); - node_block_13_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_13_1_FakeQuantWithMinMaxVars", target }); - node_block_13_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_13_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_12_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_block_13_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_13_1_FakeQuantWithMinMaxVars", target}); + node_block_13_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_13_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_12_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_12_y = _graph.node(id_mul_12_y); - node_mul_12_y->set_common_node_parameters(NodeParams{ "mul_12_y", target }); - node_mul_12_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_12_y.npy", DataLayout::NHWC)); - - NodeID id_block_12_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.3479783547154511e-06), - DataLayout::NHWC }); + node_mul_12_y->set_common_node_parameters(NodeParams{"mul_12_y", target}); + node_mul_12_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_12_y.npy", DataLayout::NHWC)); + + NodeID id_block_12_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.3479783547154511e-06), DataLayout::NHWC}); INode *node_block_12_1_Conv2D_bias = _graph.node(id_block_12_1_Conv2D_bias); - node_block_12_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_12_1_Conv2D_bias", target }); - node_block_12_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_12_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_12_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_12_1_Conv2D_bias", target}); + node_block_12_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_12_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_12_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.00041212860378436744, 130), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.00041212860378436744, 130), DataLayout::NHWC}); INode *node_block_12_1_FakeQuantWithMinMaxVars = _graph.node(id_block_12_1_FakeQuantWithMinMaxVars); - node_block_12_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_12_1_FakeQuantWithMinMaxVars", target }); - node_block_12_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_12_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_11_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_block_12_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_12_1_FakeQuantWithMinMaxVars", target}); + node_block_12_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_12_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_11_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_11_y = _graph.node(id_mul_11_y); - node_mul_11_y->set_common_node_parameters(NodeParams{ "mul_11_y", target }); - node_mul_11_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_11_y.npy", DataLayout::NHWC)); - - NodeID id_block_11_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.2847248171965475e-06), - DataLayout::NHWC }); + node_mul_11_y->set_common_node_parameters(NodeParams{"mul_11_y", target}); + node_mul_11_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_11_y.npy", DataLayout::NHWC)); + + NodeID id_block_11_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.2847248171965475e-06), DataLayout::NHWC}); INode *node_block_11_1_Conv2D_bias = _graph.node(id_block_11_1_Conv2D_bias); - node_block_11_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_11_1_Conv2D_bias", target }); - node_block_11_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_11_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_11_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_11_1_Conv2D_bias", target}); + node_block_11_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_11_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_11_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.00040296532097272575, 131), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.00040296532097272575, 131), DataLayout::NHWC}); INode *node_block_11_1_FakeQuantWithMinMaxVars = _graph.node(id_block_11_1_FakeQuantWithMinMaxVars); - node_block_11_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_11_1_FakeQuantWithMinMaxVars", target }); - node_block_11_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_11_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_10_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_block_11_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_11_1_FakeQuantWithMinMaxVars", target}); + node_block_11_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_11_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_10_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_10_y = _graph.node(id_mul_10_y); - node_mul_10_y->set_common_node_parameters(NodeParams{ "mul_10_y", target }); - node_mul_10_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_10_y.npy", DataLayout::NHWC)); - - NodeID id_block_10_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.1997129831797793e-06), - DataLayout::NHWC }); + node_mul_10_y->set_common_node_parameters(NodeParams{"mul_10_y", target}); + node_mul_10_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_10_y.npy", DataLayout::NHWC)); + + NodeID id_block_10_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.1997129831797793e-06), DataLayout::NHWC}); INode *node_block_10_1_Conv2D_bias = _graph.node(id_block_10_1_Conv2D_bias); - node_block_10_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_10_1_Conv2D_bias", target }); - node_block_10_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_10_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_10_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_10_1_Conv2D_bias", target}); + node_block_10_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_10_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_10_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.00036640543839894235, 129), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.00036640543839894235, 129), DataLayout::NHWC}); INode *node_block_10_1_FakeQuantWithMinMaxVars = _graph.node(id_block_10_1_FakeQuantWithMinMaxVars); - node_block_10_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_10_1_FakeQuantWithMinMaxVars", target }); - node_block_10_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_10_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_9_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_block_10_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_10_1_FakeQuantWithMinMaxVars", target}); + node_block_10_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_10_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_9_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_9_y = _graph.node(id_mul_9_y); - node_mul_9_y->set_common_node_parameters(NodeParams{ "mul_9_y", target }); - node_mul_9_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_9_y.npy", DataLayout::NHWC)); - - NodeID id_block_9_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.1920226370421005e-06), - DataLayout::NHWC }); + node_mul_9_y->set_common_node_parameters(NodeParams{"mul_9_y", target}); + node_mul_9_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_9_y.npy", DataLayout::NHWC)); + + NodeID id_block_9_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.1920226370421005e-06), DataLayout::NHWC}); INode *node_block_9_1_Conv2D_bias = _graph.node(id_block_9_1_Conv2D_bias); - node_block_9_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_9_1_Conv2D_bias", target }); - node_block_9_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_9_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_9_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_9_1_Conv2D_bias", target}); + node_block_9_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_9_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_9_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.0003706997958943248, 129), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.0003706997958943248, 129), DataLayout::NHWC}); INode *node_block_9_1_FakeQuantWithMinMaxVars = _graph.node(id_block_9_1_FakeQuantWithMinMaxVars); - node_block_9_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_9_1_FakeQuantWithMinMaxVars", target }); - node_block_9_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_9_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_8_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_block_9_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_9_1_FakeQuantWithMinMaxVars", target}); + node_block_9_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_9_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_8_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_8_y = _graph.node(id_mul_8_y); - node_mul_8_y->set_common_node_parameters(NodeParams{ "mul_8_y", target }); - node_mul_8_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_8_y.npy", DataLayout::NHWC)); - - NodeID id_block_8_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.218903321387188e-06), - DataLayout::NHWC }); + node_mul_8_y->set_common_node_parameters(NodeParams{"mul_8_y", target}); + node_mul_8_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_8_y.npy", DataLayout::NHWC)); + + NodeID id_block_8_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.218903321387188e-06), DataLayout::NHWC}); INode *node_block_8_1_Conv2D_bias = _graph.node(id_block_8_1_Conv2D_bias); - node_block_8_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_8_1_Conv2D_bias", target }); - node_block_8_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_8_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_8_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_8_1_Conv2D_bias", target}); + node_block_8_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_8_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_8_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.00038377835880964994, 127), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.00038377835880964994, 127), DataLayout::NHWC}); INode *node_block_8_1_FakeQuantWithMinMaxVars = _graph.node(id_block_8_1_FakeQuantWithMinMaxVars); - node_block_8_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_8_1_FakeQuantWithMinMaxVars", target }); - node_block_8_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_8_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_7_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_block_8_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_8_1_FakeQuantWithMinMaxVars", target}); + node_block_8_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_8_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_7_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_7_y = _graph.node(id_mul_7_y); - node_mul_7_y->set_common_node_parameters(NodeParams{ "mul_7_y", target }); - node_mul_7_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_7_y.npy", DataLayout::NHWC)); - - NodeID id_block_7_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.257252392861119e-06), - DataLayout::NHWC }); + node_mul_7_y->set_common_node_parameters(NodeParams{"mul_7_y", target}); + node_mul_7_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_7_y.npy", DataLayout::NHWC)); + + NodeID id_block_7_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.257252392861119e-06), DataLayout::NHWC}); INode *node_block_7_1_Conv2D_bias = _graph.node(id_block_7_1_Conv2D_bias); - node_block_7_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_7_1_Conv2D_bias", target }); - node_block_7_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_7_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_7_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_7_1_Conv2D_bias", target}); + node_block_7_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_7_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_7_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.00039844686398282647, 129), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.00039844686398282647, 129), DataLayout::NHWC}); INode *node_block_7_1_FakeQuantWithMinMaxVars = _graph.node(id_block_7_1_FakeQuantWithMinMaxVars); - node_block_7_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_7_1_FakeQuantWithMinMaxVars", target }); - node_block_7_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_7_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_6_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_block_7_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_7_1_FakeQuantWithMinMaxVars", target}); + node_block_7_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_7_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_6_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_6_y = _graph.node(id_mul_6_y); - node_mul_6_y->set_common_node_parameters(NodeParams{ "mul_6_y", target }); - node_mul_6_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_6_y.npy", DataLayout::NHWC)); - - NodeID id_block_6_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.244850636794581e-06), - DataLayout::NHWC }); + node_mul_6_y->set_common_node_parameters(NodeParams{"mul_6_y", target}); + node_mul_6_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_6_y.npy", DataLayout::NHWC)); + + NodeID id_block_6_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.244850636794581e-06), DataLayout::NHWC}); INode *node_block_6_1_Conv2D_bias = _graph.node(id_block_6_1_Conv2D_bias); - node_block_6_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_6_1_Conv2D_bias", target }); - node_block_6_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_6_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_6_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_6_1_Conv2D_bias", target}); + node_block_6_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_6_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_6_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.00040187727427110076, 132), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.00040187727427110076, 132), DataLayout::NHWC}); INode *node_block_6_1_FakeQuantWithMinMaxVars = _graph.node(id_block_6_1_FakeQuantWithMinMaxVars); - node_block_6_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_6_1_FakeQuantWithMinMaxVars", target }); - node_block_6_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_6_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_5_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_block_6_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_6_1_FakeQuantWithMinMaxVars", target}); + node_block_6_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_6_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_5_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_5_y = _graph.node(id_mul_5_y); - node_mul_5_y->set_common_node_parameters(NodeParams{ "mul_5_y", target }); - node_mul_5_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_5_y.npy", DataLayout::NHWC)); - - NodeID id_block_5_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.241092718373693e-06), - DataLayout::NHWC }); + node_mul_5_y->set_common_node_parameters(NodeParams{"mul_5_y", target}); + node_mul_5_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_5_y.npy", DataLayout::NHWC)); + + NodeID id_block_5_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.241092718373693e-06), DataLayout::NHWC}); INode *node_block_5_1_Conv2D_bias = _graph.node(id_block_5_1_Conv2D_bias); - node_block_5_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_5_1_Conv2D_bias", target }); - node_block_5_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_5_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_5_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_5_1_Conv2D_bias", target}); + node_block_5_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_5_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_5_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.0003938926674891263, 129), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.0003938926674891263, 129), DataLayout::NHWC}); INode *node_block_5_1_FakeQuantWithMinMaxVars = _graph.node(id_block_5_1_FakeQuantWithMinMaxVars); - node_block_5_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_5_1_FakeQuantWithMinMaxVars", target }); - node_block_5_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_5_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_4_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_block_5_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_5_1_FakeQuantWithMinMaxVars", target}); + node_block_5_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_5_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_4_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_4_y = _graph.node(id_mul_4_y); - node_mul_4_y->set_common_node_parameters(NodeParams{ "mul_4_y", target }); - node_mul_4_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_4_y.npy", DataLayout::NHWC)); - - NodeID id_block_4_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.1748390988941537e-06), - DataLayout::NHWC }); + node_mul_4_y->set_common_node_parameters(NodeParams{"mul_4_y", target}); + node_mul_4_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_4_y.npy", DataLayout::NHWC)); + + NodeID id_block_4_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.1748390988941537e-06), DataLayout::NHWC}); INode *node_block_4_1_Conv2D_bias = _graph.node(id_block_4_1_Conv2D_bias); - node_block_4_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_4_1_Conv2D_bias", target }); - node_block_4_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_4_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_4_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_4_1_Conv2D_bias", target}); + node_block_4_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_4_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_4_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.0003788181929849088, 129), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.0003788181929849088, 129), DataLayout::NHWC}); INode *node_block_4_1_FakeQuantWithMinMaxVars = _graph.node(id_block_4_1_FakeQuantWithMinMaxVars); - node_block_4_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_4_1_FakeQuantWithMinMaxVars", target }); - node_block_4_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_4_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_3_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_block_4_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_4_1_FakeQuantWithMinMaxVars", target}); + node_block_4_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_4_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_3_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_3_y = _graph.node(id_mul_3_y); - node_mul_3_y->set_common_node_parameters(NodeParams{ "mul_3_y", target }); - node_mul_3_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_3_y.npy", DataLayout::NHWC)); - - NodeID id_block_3_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.1937011095142225e-06), - DataLayout::NHWC }); + node_mul_3_y->set_common_node_parameters(NodeParams{"mul_3_y", target}); + node_mul_3_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_3_y.npy", DataLayout::NHWC)); + + NodeID id_block_3_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.1937011095142225e-06), DataLayout::NHWC}); INode *node_block_3_1_Conv2D_bias = _graph.node(id_block_3_1_Conv2D_bias); - node_block_3_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_3_1_Conv2D_bias", target }); - node_block_3_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_3_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_3_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_3_1_Conv2D_bias", target}); + node_block_3_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_3_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_3_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.0003944312920793891, 129), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.0003944312920793891, 129), DataLayout::NHWC}); INode *node_block_3_1_FakeQuantWithMinMaxVars = _graph.node(id_block_3_1_FakeQuantWithMinMaxVars); - node_block_3_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_3_1_FakeQuantWithMinMaxVars", target }); - node_block_3_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_3_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_2_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_block_3_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_3_1_FakeQuantWithMinMaxVars", target}); + node_block_3_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_3_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_2_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_2_y = _graph.node(id_mul_2_y); - node_mul_2_y->set_common_node_parameters(NodeParams{ "mul_2_y", target }); - node_mul_2_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_2_y.npy", DataLayout::NHWC)); - - NodeID id_block_2_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.1634580232566805e-06), - DataLayout::NHWC }); + node_mul_2_y->set_common_node_parameters(NodeParams{"mul_2_y", target}); + node_mul_2_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_2_y.npy", DataLayout::NHWC)); + + NodeID id_block_2_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.1634580232566805e-06), DataLayout::NHWC}); INode *node_block_2_1_Conv2D_bias = _graph.node(id_block_2_1_Conv2D_bias); - node_block_2_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_2_1_Conv2D_bias", target }); - node_block_2_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_2_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_2_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_2_1_Conv2D_bias", target}); + node_block_2_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_2_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_2_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.0003789655165746808, 132), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.0003789655165746808, 132), DataLayout::NHWC}); INode *node_block_2_1_FakeQuantWithMinMaxVars = _graph.node(id_block_2_1_FakeQuantWithMinMaxVars); - node_block_2_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_2_1_FakeQuantWithMinMaxVars", target }); - node_block_2_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_2_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_1_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_block_2_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_2_1_FakeQuantWithMinMaxVars", target}); + node_block_2_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_2_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_1_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_1_y = _graph.node(id_mul_1_y); - node_mul_1_y->set_common_node_parameters(NodeParams{ "mul_1_y", target }); - node_mul_1_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_1_y.npy", DataLayout::NHWC)); - - NodeID id_block_1_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.197920255435747e-06), - DataLayout::NHWC }); + node_mul_1_y->set_common_node_parameters(NodeParams{"mul_1_y", target}); + node_mul_1_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_1_y.npy", DataLayout::NHWC)); + + NodeID id_block_1_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.197920255435747e-06), DataLayout::NHWC}); INode *node_block_1_1_Conv2D_bias = _graph.node(id_block_1_1_Conv2D_bias); - node_block_1_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_1_1_Conv2D_bias", target }); - node_block_1_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_1_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_1_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_1_1_Conv2D_bias", target}); + node_block_1_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_1_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_1_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.00038527738070115447, 132), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.00038527738070115447, 132), DataLayout::NHWC}); INode *node_block_1_1_FakeQuantWithMinMaxVars = _graph.node(id_block_1_1_FakeQuantWithMinMaxVars); - node_block_1_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_1_1_FakeQuantWithMinMaxVars", target }); - node_block_1_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_1_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_mul_y = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 1 }, - DataType::QASYMM8, - QuantizationInfo(0.0003921568568330258), - DataLayout::NHWC }); + node_block_1_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_1_1_FakeQuantWithMinMaxVars", target}); + node_block_1_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_1_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_mul_y = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{1}, DataType::QASYMM8, QuantizationInfo(0.0003921568568330258), DataLayout::NHWC}); INode *node_mul_y = _graph.node(id_mul_y); - node_mul_y->set_common_node_parameters(NodeParams{ "mul_y", target }); - node_mul_y->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_y.npy", DataLayout::NHWC)); - - NodeID id_block_0_1_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.315485519626236e-06), - DataLayout::NHWC }); + node_mul_y->set_common_node_parameters(NodeParams{"mul_y", target}); + node_mul_y->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/mul_y.npy", DataLayout::NHWC)); + + NodeID id_block_0_1_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.315485519626236e-06), DataLayout::NHWC}); INode *node_block_0_1_Conv2D_bias = _graph.node(id_block_0_1_Conv2D_bias); - node_block_0_1_Conv2D_bias->set_common_node_parameters(NodeParams{ "block_0_1_Conv2D_bias", target }); - node_block_0_1_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_0_1_Conv2D_bias.npy", DataLayout::NHWC)); + node_block_0_1_Conv2D_bias->set_common_node_parameters(NodeParams{"block_0_1_Conv2D_bias", target}); + node_block_0_1_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/block_0_1_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_block_0_1_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.00039420535904355347, 129), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{256, 3, 3, 256}, DataType::QASYMM8, + QuantizationInfo(0.00039420535904355347, 129), DataLayout::NHWC}); INode *node_block_0_1_FakeQuantWithMinMaxVars = _graph.node(id_block_0_1_FakeQuantWithMinMaxVars); - node_block_0_1_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "block_0_1_FakeQuantWithMinMaxVars", target }); - node_block_0_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/block_0_1_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); - - NodeID id_pre_residual_Conv2D_bias = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 256 }, - DataType::S32, - QuantizationInfo(1.7214160834555514e-06), - DataLayout::NHWC }); + node_block_0_1_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"block_0_1_FakeQuantWithMinMaxVars", target}); + node_block_0_1_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/block_0_1_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); + + NodeID id_pre_residual_Conv2D_bias = _graph.add_node<ConstNode>(TensorDescriptor{ + TensorShape{256}, DataType::S32, QuantizationInfo(1.7214160834555514e-06), DataLayout::NHWC}); INode *node_pre_residual_Conv2D_bias = _graph.node(id_pre_residual_Conv2D_bias); - node_pre_residual_Conv2D_bias->set_common_node_parameters(NodeParams{ "pre_residual_Conv2D_bias", target }); - node_pre_residual_Conv2D_bias->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/pre_residual_Conv2D_bias.npy", DataLayout::NHWC)); + node_pre_residual_Conv2D_bias->set_common_node_parameters(NodeParams{"pre_residual_Conv2D_bias", target}); + node_pre_residual_Conv2D_bias->output(0)->set_accessor( + get_weights_accessor(data_path, "/cnn_data/edsr_model/pre_residual_Conv2D_bias.npy", DataLayout::NHWC)); NodeID id_pre_residual_FakeQuantWithMinMaxVars = _graph.add_node<ConstNode>( - TensorDescriptor - { - TensorShape{ 3, 3, 3, 256 }, - DataType::QASYMM8, - QuantizationInfo(0.0004389610840007663, 127), - DataLayout::NHWC }); + TensorDescriptor{TensorShape{3, 3, 3, 256}, DataType::QASYMM8, QuantizationInfo(0.0004389610840007663, 127), + DataLayout::NHWC}); INode *node_pre_residual_FakeQuantWithMinMaxVars = _graph.node(id_pre_residual_FakeQuantWithMinMaxVars); - node_pre_residual_FakeQuantWithMinMaxVars->set_common_node_parameters(NodeParams{ "pre_residual_FakeQuantWithMinMaxVars", target }); - node_pre_residual_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor(data_path, "/cnn_data/edsr_model/pre_residual_FakeQuantWithMinMaxVars.npy", - DataLayout::NHWC)); + node_pre_residual_FakeQuantWithMinMaxVars->set_common_node_parameters( + NodeParams{"pre_residual_FakeQuantWithMinMaxVars", target}); + node_pre_residual_FakeQuantWithMinMaxVars->output(0)->set_accessor(get_weights_accessor( + data_path, "/cnn_data/edsr_model/pre_residual_FakeQuantWithMinMaxVars.npy", DataLayout::NHWC)); TensorShape input_shape{}; input_shape.set(0, 3, false).set(1, 360, false).set(2, 640, false).set(3, 1, false); NodeID id_input = _graph.add_node<InputNode>( - TensorDescriptor - { - input_shape, - DataType::QASYMM8, - QuantizationInfo(0.003921568859368563), - DataLayout::NHWC }); + TensorDescriptor{input_shape, DataType::QASYMM8, QuantizationInfo(0.003921568859368563), DataLayout::NHWC}); INode *node_input = _graph.node(id_input); - node_input->set_common_node_parameters(NodeParams{ "input", target }); + node_input->set_common_node_parameters(NodeParams{"input", target}); node_input->output(0)->set_accessor(get_input_accessor(common_params)); - NodeID id_pre_residual_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.0033370566088706255, 96)); + NodeID id_pre_residual_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.0033370566088706255, 96)); INode *node_pre_residual_BiasAdd = _graph.node(id_pre_residual_BiasAdd); - node_pre_residual_BiasAdd->set_common_node_parameters(NodeParams{ "pre_residual_BiasAdd", target }); + node_pre_residual_BiasAdd->set_common_node_parameters(NodeParams{"pre_residual_BiasAdd", target}); _graph.add_connection(id_input, 0, id_pre_residual_BiasAdd, 0); _graph.add_connection(id_pre_residual_FakeQuantWithMinMaxVars, 0, id_pre_residual_BiasAdd, 1); _graph.add_connection(id_pre_residual_Conv2D_bias, 0, id_pre_residual_BiasAdd, 2); - NodeID id_block_0_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.007344874087721109, 185)); + NodeID id_block_0_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.007344874087721109, 185)); INode *node_block_0_1_BiasAdd = _graph.node(id_block_0_1_BiasAdd); - node_block_0_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_0_1_BiasAdd", target }); + node_block_0_1_BiasAdd->set_common_node_parameters(NodeParams{"block_0_1_BiasAdd", target}); _graph.add_connection(id_pre_residual_BiasAdd, 0, id_block_0_1_BiasAdd, 0); _graph.add_connection(id_block_0_1_FakeQuantWithMinMaxVars, 0, id_block_0_1_BiasAdd, 1); _graph.add_connection(id_block_0_1_Conv2D_bias, 0, id_block_0_1_BiasAdd, 2); NodeID id_mul = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.0006341293919831514, 174 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.0006341293919831514, 174}}); INode *node_mul = _graph.node(id_mul); - node_mul->set_common_node_parameters(NodeParams{ "mul", target }); + node_mul->set_common_node_parameters(NodeParams{"mul", target}); _graph.add_connection(id_block_0_1_BiasAdd, 0, id_mul, 0); _graph.add_connection(id_mul_y, 0, id_mul, 1); NodeID id_add = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.0031092411372810602, 95 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.0031092411372810602, 95}}); INode *node_add = _graph.node(id_add); - node_add->set_common_node_parameters(NodeParams{ "add", target }); + node_add->set_common_node_parameters(NodeParams{"add", target}); _graph.add_connection(id_pre_residual_BiasAdd, 0, id_add, 0); _graph.add_connection(id_mul, 0, id_add, 1); - NodeID id_block_1_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.005333727691322565, 117)); + NodeID id_block_1_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.005333727691322565, 117)); INode *node_block_1_1_BiasAdd = _graph.node(id_block_1_1_BiasAdd); - node_block_1_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_1_1_BiasAdd", target }); + node_block_1_1_BiasAdd->set_common_node_parameters(NodeParams{"block_1_1_BiasAdd", target}); _graph.add_connection(id_add, 0, id_block_1_1_BiasAdd, 0); _graph.add_connection(id_block_1_1_FakeQuantWithMinMaxVars, 0, id_block_1_1_BiasAdd, 1); _graph.add_connection(id_block_1_1_Conv2D_bias, 0, id_block_1_1_BiasAdd, 2); NodeID id_mul_1 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.0004965941770933568, 122 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.0004965941770933568, 122}}); INode *node_mul_1 = _graph.node(id_mul_1); - node_mul_1->set_common_node_parameters(NodeParams{ "mul_1", target }); + node_mul_1->set_common_node_parameters(NodeParams{"mul_1", target}); _graph.add_connection(id_block_1_1_BiasAdd, 0, id_mul_1, 0); _graph.add_connection(id_mul_1_y, 0, id_mul_1, 1); NodeID id_add_1 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.0030700892675668, 96 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.0030700892675668, 96}}); INode *node_add_1 = _graph.node(id_add_1); - node_add_1->set_common_node_parameters(NodeParams{ "add_1", target }); + node_add_1->set_common_node_parameters(NodeParams{"add_1", target}); _graph.add_connection(id_add, 0, id_add_1, 0); _graph.add_connection(id_mul_1, 0, id_add_1, 1); - NodeID id_block_2_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.004199742339551449, 132)); + NodeID id_block_2_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.004199742339551449, 132)); INode *node_block_2_1_BiasAdd = _graph.node(id_block_2_1_BiasAdd); - node_block_2_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_2_1_BiasAdd", target }); + node_block_2_1_BiasAdd->set_common_node_parameters(NodeParams{"block_2_1_BiasAdd", target}); _graph.add_connection(id_add_1, 0, id_block_2_1_BiasAdd, 0); _graph.add_connection(id_block_2_1_FakeQuantWithMinMaxVars, 0, id_block_2_1_BiasAdd, 1); _graph.add_connection(id_block_2_1_Conv2D_bias, 0, id_block_2_1_BiasAdd, 2); NodeID id_mul_2 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.0004133903712499887, 130 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.0004133903712499887, 130}}); INode *node_mul_2 = _graph.node(id_mul_2); - node_mul_2->set_common_node_parameters(NodeParams{ "mul_2", target }); + node_mul_2->set_common_node_parameters(NodeParams{"mul_2", target}); _graph.add_connection(id_block_2_1_BiasAdd, 0, id_mul_2, 0); _graph.add_connection(id_mul_2_y, 0, id_mul_2, 1); NodeID id_add_2 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.003026385325938463, 94 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.003026385325938463, 94}}); INode *node_add_2 = _graph.node(id_add_2); - node_add_2->set_common_node_parameters(NodeParams{ "add_2", target }); + node_add_2->set_common_node_parameters(NodeParams{"add_2", target}); _graph.add_connection(id_add_1, 0, id_add_2, 0); _graph.add_connection(id_mul_2, 0, id_add_2, 1); - NodeID id_block_3_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.003977528307586908, 142)); + NodeID id_block_3_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.003977528307586908, 142)); INode *node_block_3_1_BiasAdd = _graph.node(id_block_3_1_BiasAdd); - node_block_3_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_3_1_BiasAdd", target }); + node_block_3_1_BiasAdd->set_common_node_parameters(NodeParams{"block_3_1_BiasAdd", target}); _graph.add_connection(id_add_2, 0, id_block_3_1_BiasAdd, 0); _graph.add_connection(id_block_3_1_FakeQuantWithMinMaxVars, 0, id_block_3_1_BiasAdd, 1); _graph.add_connection(id_block_3_1_Conv2D_bias, 0, id_block_3_1_BiasAdd, 2); NodeID id_mul_3 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.0003943995980080217, 141 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.0003943995980080217, 141}}); INode *node_mul_3 = _graph.node(id_mul_3); - node_mul_3->set_common_node_parameters(NodeParams{ "mul_3", target }); + node_mul_3->set_common_node_parameters(NodeParams{"mul_3", target}); _graph.add_connection(id_block_3_1_BiasAdd, 0, id_mul_3, 0); _graph.add_connection(id_mul_3_y, 0, id_mul_3, 1); NodeID id_add_3 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.003101327223703265, 98 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.003101327223703265, 98}}); INode *node_add_3 = _graph.node(id_add_3); - node_add_3->set_common_node_parameters(NodeParams{ "add_3", target }); + node_add_3->set_common_node_parameters(NodeParams{"add_3", target}); _graph.add_connection(id_add_2, 0, id_add_3, 0); _graph.add_connection(id_mul_3, 0, id_add_3, 1); - NodeID id_block_4_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.0045388080179691315, 146)); + NodeID id_block_4_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.0045388080179691315, 146)); INode *node_block_4_1_BiasAdd = _graph.node(id_block_4_1_BiasAdd); - node_block_4_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_4_1_BiasAdd", target }); + node_block_4_1_BiasAdd->set_common_node_parameters(NodeParams{"block_4_1_BiasAdd", target}); _graph.add_connection(id_add_3, 0, id_block_4_1_BiasAdd, 0); _graph.add_connection(id_block_4_1_FakeQuantWithMinMaxVars, 0, id_block_4_1_BiasAdd, 1); _graph.add_connection(id_block_4_1_Conv2D_bias, 0, id_block_4_1_BiasAdd, 2); NodeID id_mul_4 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.00044342130422592163, 143 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.00044342130422592163, 143}}); INode *node_mul_4 = _graph.node(id_mul_4); - node_mul_4->set_common_node_parameters(NodeParams{ "mul_4", target }); + node_mul_4->set_common_node_parameters(NodeParams{"mul_4", target}); _graph.add_connection(id_block_4_1_BiasAdd, 0, id_mul_4, 0); _graph.add_connection(id_mul_4_y, 0, id_mul_4, 1); NodeID id_add_4 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.003150839824229479, 98 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.003150839824229479, 98}}); INode *node_add_4 = _graph.node(id_add_4); - node_add_4->set_common_node_parameters(NodeParams{ "add_4", target }); + node_add_4->set_common_node_parameters(NodeParams{"add_4", target}); _graph.add_connection(id_add_3, 0, id_add_4, 0); _graph.add_connection(id_mul_4, 0, id_add_4, 1); - NodeID id_block_5_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.00402890844270587, 132)); + NodeID id_block_5_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.00402890844270587, 132)); INode *node_block_5_1_BiasAdd = _graph.node(id_block_5_1_BiasAdd); - node_block_5_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_5_1_BiasAdd", target }); + node_block_5_1_BiasAdd->set_common_node_parameters(NodeParams{"block_5_1_BiasAdd", target}); _graph.add_connection(id_add_4, 0, id_block_5_1_BiasAdd, 0); _graph.add_connection(id_block_5_1_FakeQuantWithMinMaxVars, 0, id_block_5_1_BiasAdd, 1); _graph.add_connection(id_block_5_1_Conv2D_bias, 0, id_block_5_1_BiasAdd, 2); NodeID id_mul_5 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.0004023382789455354, 132 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.0004023382789455354, 132}}); INode *node_mul_5 = _graph.node(id_mul_5); - node_mul_5->set_common_node_parameters(NodeParams{ "mul_5", target }); + node_mul_5->set_common_node_parameters(NodeParams{"mul_5", target}); _graph.add_connection(id_block_5_1_BiasAdd, 0, id_mul_5, 0); _graph.add_connection(id_mul_5_y, 0, id_mul_5, 1); NodeID id_add_5 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.0030975888948887587, 94 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.0030975888948887587, 94}}); INode *node_add_5 = _graph.node(id_add_5); - node_add_5->set_common_node_parameters(NodeParams{ "add_5", target }); + node_add_5->set_common_node_parameters(NodeParams{"add_5", target}); _graph.add_connection(id_add_4, 0, id_add_5, 0); _graph.add_connection(id_mul_5, 0, id_add_5, 1); - NodeID id_block_6_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.00421866774559021, 125)); + NodeID id_block_6_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.00421866774559021, 125)); INode *node_block_6_1_BiasAdd = _graph.node(id_block_6_1_BiasAdd); - node_block_6_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_6_1_BiasAdd", target }); + node_block_6_1_BiasAdd->set_common_node_parameters(NodeParams{"block_6_1_BiasAdd", target}); _graph.add_connection(id_add_5, 0, id_block_6_1_BiasAdd, 0); _graph.add_connection(id_block_6_1_FakeQuantWithMinMaxVars, 0, id_block_6_1_BiasAdd, 1); _graph.add_connection(id_block_6_1_Conv2D_bias, 0, id_block_6_1_BiasAdd, 2); NodeID id_mul_6 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.00041950203012675047, 125 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.00041950203012675047, 125}}); INode *node_mul_6 = _graph.node(id_mul_6); - node_mul_6->set_common_node_parameters(NodeParams{ "mul_6", target }); + node_mul_6->set_common_node_parameters(NodeParams{"mul_6", target}); _graph.add_connection(id_block_6_1_BiasAdd, 0, id_mul_6, 0); _graph.add_connection(id_mul_6_y, 0, id_mul_6, 1); NodeID id_add_6 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.003155382815748453, 92 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.003155382815748453, 92}}); INode *node_add_6 = _graph.node(id_add_6); - node_add_6->set_common_node_parameters(NodeParams{ "add_6", target }); + node_add_6->set_common_node_parameters(NodeParams{"add_6", target}); _graph.add_connection(id_add_5, 0, id_add_6, 0); _graph.add_connection(id_mul_6, 0, id_add_6, 1); - NodeID id_block_7_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.004250136204063892, 143)); + NodeID id_block_7_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.004250136204063892, 143)); INode *node_block_7_1_BiasAdd = _graph.node(id_block_7_1_BiasAdd); - node_block_7_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_7_1_BiasAdd", target }); + node_block_7_1_BiasAdd->set_common_node_parameters(NodeParams{"block_7_1_BiasAdd", target}); _graph.add_connection(id_add_6, 0, id_block_7_1_BiasAdd, 0); _graph.add_connection(id_block_7_1_FakeQuantWithMinMaxVars, 0, id_block_7_1_BiasAdd, 1); _graph.add_connection(id_block_7_1_Conv2D_bias, 0, id_block_7_1_BiasAdd, 2); NodeID id_mul_7 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.00042401350219734013, 142 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.00042401350219734013, 142}}); INode *node_mul_7 = _graph.node(id_mul_7); - node_mul_7->set_common_node_parameters(NodeParams{ "mul_7", target }); + node_mul_7->set_common_node_parameters(NodeParams{"mul_7", target}); _graph.add_connection(id_block_7_1_BiasAdd, 0, id_mul_7, 0); _graph.add_connection(id_mul_7_y, 0, id_mul_7, 1); NodeID id_add_7 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.0031760605052113533, 86 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.0031760605052113533, 86}}); INode *node_add_7 = _graph.node(id_add_7); - node_add_7->set_common_node_parameters(NodeParams{ "add_7", target }); + node_add_7->set_common_node_parameters(NodeParams{"add_7", target}); _graph.add_connection(id_add_6, 0, id_add_7, 0); _graph.add_connection(id_mul_7, 0, id_add_7, 1); - NodeID id_block_8_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.004277155734598637, 123)); + NodeID id_block_8_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.004277155734598637, 123)); INode *node_block_8_1_BiasAdd = _graph.node(id_block_8_1_BiasAdd); - node_block_8_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_8_1_BiasAdd", target }); + node_block_8_1_BiasAdd->set_common_node_parameters(NodeParams{"block_8_1_BiasAdd", target}); _graph.add_connection(id_add_7, 0, id_block_8_1_BiasAdd, 0); _graph.add_connection(id_block_8_1_FakeQuantWithMinMaxVars, 0, id_block_8_1_BiasAdd, 1); _graph.add_connection(id_block_8_1_Conv2D_bias, 0, id_block_8_1_BiasAdd, 2); NodeID id_mul_8 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.00042673019925132394, 123 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.00042673019925132394, 123}}); INode *node_mul_8 = _graph.node(id_mul_8); - node_mul_8->set_common_node_parameters(NodeParams{ "mul_8", target }); + node_mul_8->set_common_node_parameters(NodeParams{"mul_8", target}); _graph.add_connection(id_block_8_1_BiasAdd, 0, id_mul_8, 0); _graph.add_connection(id_mul_8_y, 0, id_mul_8, 1); NodeID id_add_8 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.0032156009692698717, 86 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.0032156009692698717, 86}}); INode *node_add_8 = _graph.node(id_add_8); - node_add_8->set_common_node_parameters(NodeParams{ "add_8", target }); + node_add_8->set_common_node_parameters(NodeParams{"add_8", target}); _graph.add_connection(id_add_7, 0, id_add_8, 0); _graph.add_connection(id_mul_8, 0, id_add_8, 1); - NodeID id_block_9_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.00445037754252553, 129)); + NodeID id_block_9_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.00445037754252553, 129)); INode *node_block_9_1_BiasAdd = _graph.node(id_block_9_1_BiasAdd); - node_block_9_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_9_1_BiasAdd", target }); + node_block_9_1_BiasAdd->set_common_node_parameters(NodeParams{"block_9_1_BiasAdd", target}); _graph.add_connection(id_add_8, 0, id_block_9_1_BiasAdd, 0); _graph.add_connection(id_block_9_1_FakeQuantWithMinMaxVars, 0, id_block_9_1_BiasAdd, 1); _graph.add_connection(id_block_9_1_Conv2D_bias, 0, id_block_9_1_BiasAdd, 2); NodeID id_mul_9 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.0004448975087143481, 129 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.0004448975087143481, 129}}); INode *node_mul_9 = _graph.node(id_mul_9); - node_mul_9->set_common_node_parameters(NodeParams{ "mul_9", target }); + node_mul_9->set_common_node_parameters(NodeParams{"mul_9", target}); _graph.add_connection(id_block_9_1_BiasAdd, 0, id_mul_9, 0); _graph.add_connection(id_mul_9_y, 0, id_mul_9, 1); NodeID id_add_9 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.0032742770854383707, 80 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.0032742770854383707, 80}}); INode *node_add_9 = _graph.node(id_add_9); - node_add_9->set_common_node_parameters(NodeParams{ "add_9", target }); + node_add_9->set_common_node_parameters(NodeParams{"add_9", target}); _graph.add_connection(id_add_8, 0, id_add_9, 0); _graph.add_connection(id_mul_9, 0, id_add_9, 1); - NodeID id_block_10_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.003614710411056876, 131)); + NodeID id_block_10_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.003614710411056876, 131)); INode *node_block_10_1_BiasAdd = _graph.node(id_block_10_1_BiasAdd); - node_block_10_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_10_1_BiasAdd", target }); + node_block_10_1_BiasAdd->set_common_node_parameters(NodeParams{"block_10_1_BiasAdd", target}); _graph.add_connection(id_add_9, 0, id_block_10_1_BiasAdd, 0); _graph.add_connection(id_block_10_1_FakeQuantWithMinMaxVars, 0, id_block_10_1_BiasAdd, 1); _graph.add_connection(id_block_10_1_Conv2D_bias, 0, id_block_10_1_BiasAdd, 2); NodeID id_mul_10 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.00036083892337046564, 130 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.00036083892337046564, 130}}); INode *node_mul_10 = _graph.node(id_mul_10); - node_mul_10->set_common_node_parameters(NodeParams{ "mul_10", target }); + node_mul_10->set_common_node_parameters(NodeParams{"mul_10", target}); _graph.add_connection(id_block_10_1_BiasAdd, 0, id_mul_10, 0); _graph.add_connection(id_mul_10_y, 0, id_mul_10, 1); NodeID id_add_10 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.0031881770119071007, 81 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.0031881770119071007, 81}}); INode *node_add_10 = _graph.node(id_add_10); - node_add_10->set_common_node_parameters(NodeParams{ "add_10", target }); + node_add_10->set_common_node_parameters(NodeParams{"add_10", target}); _graph.add_connection(id_add_9, 0, id_add_10, 0); _graph.add_connection(id_mul_10, 0, id_add_10, 1); - NodeID id_block_11_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.003969002980738878, 133)); + NodeID id_block_11_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.003969002980738878, 133)); INode *node_block_11_1_BiasAdd = _graph.node(id_block_11_1_BiasAdd); - node_block_11_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_11_1_BiasAdd", target }); + node_block_11_1_BiasAdd->set_common_node_parameters(NodeParams{"block_11_1_BiasAdd", target}); _graph.add_connection(id_add_10, 0, id_block_11_1_BiasAdd, 0); _graph.add_connection(id_block_11_1_FakeQuantWithMinMaxVars, 0, id_block_11_1_BiasAdd, 1); _graph.add_connection(id_block_11_1_Conv2D_bias, 0, id_block_11_1_BiasAdd, 2); NodeID id_mul_11 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.0003968806122429669, 133 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.0003968806122429669, 133}}); INode *node_mul_11 = _graph.node(id_mul_11); - node_mul_11->set_common_node_parameters(NodeParams{ "mul_11", target }); + node_mul_11->set_common_node_parameters(NodeParams{"mul_11", target}); _graph.add_connection(id_block_11_1_BiasAdd, 0, id_mul_11, 0); _graph.add_connection(id_mul_11_y, 0, id_mul_11, 1); NodeID id_add_11 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.0032707711216062307, 80 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.0032707711216062307, 80}}); INode *node_add_11 = _graph.node(id_add_11); - node_add_11->set_common_node_parameters(NodeParams{ "add_11", target }); + node_add_11->set_common_node_parameters(NodeParams{"add_11", target}); _graph.add_connection(id_add_10, 0, id_add_11, 0); _graph.add_connection(id_mul_11, 0, id_add_11, 1); - NodeID id_block_12_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.004366801120340824, 110)); + NodeID id_block_12_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.004366801120340824, 110)); INode *node_block_12_1_BiasAdd = _graph.node(id_block_12_1_BiasAdd); - node_block_12_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_12_1_BiasAdd", target }); + node_block_12_1_BiasAdd->set_common_node_parameters(NodeParams{"block_12_1_BiasAdd", target}); _graph.add_connection(id_add_11, 0, id_block_12_1_BiasAdd, 0); _graph.add_connection(id_block_12_1_FakeQuantWithMinMaxVars, 0, id_block_12_1_BiasAdd, 1); _graph.add_connection(id_block_12_1_Conv2D_bias, 0, id_block_12_1_BiasAdd, 2); NodeID id_mul_12 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.0004365936329122633, 110 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.0004365936329122633, 110}}); INode *node_mul_12 = _graph.node(id_mul_12); - node_mul_12->set_common_node_parameters(NodeParams{ "mul_12", target }); + node_mul_12->set_common_node_parameters(NodeParams{"mul_12", target}); _graph.add_connection(id_block_12_1_BiasAdd, 0, id_mul_12, 0); _graph.add_connection(id_mul_12_y, 0, id_mul_12, 1); NodeID id_add_12 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.003275055903941393, 79 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.003275055903941393, 79}}); INode *node_add_12 = _graph.node(id_add_12); - node_add_12->set_common_node_parameters(NodeParams{ "add_12", target }); + node_add_12->set_common_node_parameters(NodeParams{"add_12", target}); _graph.add_connection(id_add_11, 0, id_add_12, 0); _graph.add_connection(id_mul_12, 0, id_add_12, 1); - NodeID id_block_13_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.004386766813695431, 139)); + NodeID id_block_13_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.004386766813695431, 139)); INode *node_block_13_1_BiasAdd = _graph.node(id_block_13_1_BiasAdd); - node_block_13_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_13_1_BiasAdd", target }); + node_block_13_1_BiasAdd->set_common_node_parameters(NodeParams{"block_13_1_BiasAdd", target}); _graph.add_connection(id_add_12, 0, id_block_13_1_BiasAdd, 0); _graph.add_connection(id_block_13_1_FakeQuantWithMinMaxVars, 0, id_block_13_1_BiasAdd, 1); _graph.add_connection(id_block_13_1_Conv2D_bias, 0, id_block_13_1_BiasAdd, 2); NodeID id_mul_13 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.0004385628562886268, 139 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.0004385628562886268, 139}}); INode *node_mul_13 = _graph.node(id_mul_13); - node_mul_13->set_common_node_parameters(NodeParams{ "mul_13", target }); + node_mul_13->set_common_node_parameters(NodeParams{"mul_13", target}); _graph.add_connection(id_block_13_1_BiasAdd, 0, id_mul_13, 0); _graph.add_connection(id_mul_13_y, 0, id_mul_13, 1); NodeID id_add_13 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.0033287261612713337, 78 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.0033287261612713337, 78}}); INode *node_add_13 = _graph.node(id_add_13); - node_add_13->set_common_node_parameters(NodeParams{ "add_13", target }); + node_add_13->set_common_node_parameters(NodeParams{"add_13", target}); _graph.add_connection(id_add_12, 0, id_add_13, 0); _graph.add_connection(id_mul_13, 0, id_add_13, 1); - NodeID id_block_14_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.0038069337606430054, 130)); + NodeID id_block_14_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.0038069337606430054, 130)); INode *node_block_14_1_BiasAdd = _graph.node(id_block_14_1_BiasAdd); - node_block_14_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_14_1_BiasAdd", target }); + node_block_14_1_BiasAdd->set_common_node_parameters(NodeParams{"block_14_1_BiasAdd", target}); _graph.add_connection(id_add_13, 0, id_block_14_1_BiasAdd, 0); _graph.add_connection(id_block_14_1_FakeQuantWithMinMaxVars, 0, id_block_14_1_BiasAdd, 1); _graph.add_connection(id_block_14_1_Conv2D_bias, 0, id_block_14_1_BiasAdd, 2); NodeID id_mul_14 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.00037829321809113026, 130 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.00037829321809113026, 130}}); INode *node_mul_14 = _graph.node(id_mul_14); - node_mul_14->set_common_node_parameters(NodeParams{ "mul_14", target }); + node_mul_14->set_common_node_parameters(NodeParams{"mul_14", target}); _graph.add_connection(id_block_14_1_BiasAdd, 0, id_mul_14, 0); _graph.add_connection(id_mul_14_y, 0, id_mul_14, 1); NodeID id_add_14 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.0033590947277843952, 77 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.0033590947277843952, 77}}); INode *node_add_14 = _graph.node(id_add_14); - node_add_14->set_common_node_parameters(NodeParams{ "add_14", target }); + node_add_14->set_common_node_parameters(NodeParams{"add_14", target}); _graph.add_connection(id_add_13, 0, id_add_14, 0); _graph.add_connection(id_mul_14, 0, id_add_14, 1); - NodeID id_block_15_1_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.004009159281849861, 130)); + NodeID id_block_15_1_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.004009159281849861, 130)); INode *node_block_15_1_BiasAdd = _graph.node(id_block_15_1_BiasAdd); - node_block_15_1_BiasAdd->set_common_node_parameters(NodeParams{ "block_15_1_BiasAdd", target }); + node_block_15_1_BiasAdd->set_common_node_parameters(NodeParams{"block_15_1_BiasAdd", target}); _graph.add_connection(id_add_14, 0, id_block_15_1_BiasAdd, 0); _graph.add_connection(id_block_15_1_FakeQuantWithMinMaxVars, 0, id_block_15_1_BiasAdd, 1); _graph.add_connection(id_block_15_1_Conv2D_bias, 0, id_block_15_1_BiasAdd, 2); NodeID id_mul_15 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Mul, QuantizationInfo{ 0.0004008286341559142, 130 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Mul, QuantizationInfo{0.0004008286341559142, 130}}); INode *node_mul_15 = _graph.node(id_mul_15); - node_mul_15->set_common_node_parameters(NodeParams{ "mul_15", target }); + node_mul_15->set_common_node_parameters(NodeParams{"mul_15", target}); _graph.add_connection(id_block_15_1_BiasAdd, 0, id_mul_15, 0); _graph.add_connection(id_mul_15_y, 0, id_mul_15, 1); NodeID id_add_15 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.0035031239967793226, 78 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.0035031239967793226, 78}}); INode *node_add_15 = _graph.node(id_add_15); - node_add_15->set_common_node_parameters(NodeParams{ "add_15", target }); + node_add_15->set_common_node_parameters(NodeParams{"add_15", target}); _graph.add_connection(id_add_14, 0, id_add_15, 0); _graph.add_connection(id_mul_15, 0, id_add_15, 1); - NodeID id_post_residual_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.005167999770492315, 112)); + NodeID id_post_residual_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.005167999770492315, 112)); INode *node_post_residual_BiasAdd = _graph.node(id_post_residual_BiasAdd); - node_post_residual_BiasAdd->set_common_node_parameters(NodeParams{ "post_residual_BiasAdd", target }); + node_post_residual_BiasAdd->set_common_node_parameters(NodeParams{"post_residual_BiasAdd", target}); _graph.add_connection(id_add_15, 0, id_post_residual_BiasAdd, 0); _graph.add_connection(id_post_residual_FakeQuantWithMinMaxVars, 0, id_post_residual_BiasAdd, 1); _graph.add_connection(id_post_residual_Conv2D_bias, 0, id_post_residual_BiasAdd, 2); NodeID id_add_16 = _graph.add_node<EltwiseLayerNode>( - descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add, QuantizationInfo{ 0.0065071373246610165, 89 } }); + descriptors::EltwiseLayerDescriptor{EltwiseOperation::Add, QuantizationInfo{0.0065071373246610165, 89}}); INode *node_add_16 = _graph.node(id_add_16); - node_add_16->set_common_node_parameters(NodeParams{ "add_16", target }); + node_add_16->set_common_node_parameters(NodeParams{"add_16", target}); _graph.add_connection(id_post_residual_BiasAdd, 0, id_add_16, 0); _graph.add_connection(id_pre_residual_BiasAdd, 0, id_add_16, 1); - NodeID id_pre_upscale_BiasAdd = _graph.add_node<ConvolutionLayerNode>( - PadStrideInfo - { - 1, 1, - 1, 1, - 1, 1, - DimensionRoundingType::FLOOR }, - 1, - arm_compute::graph::ConvolutionMethod::Default, - FastMathHint::Disabled, - QuantizationInfo(0.005013593938201666, 26)); + NodeID id_pre_upscale_BiasAdd = + _graph.add_node<ConvolutionLayerNode>(PadStrideInfo{1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR}, 1, + arm_compute::graph::ConvolutionMethod::Default, + FastMathHint::Disabled, QuantizationInfo(0.005013593938201666, 26)); INode *node_pre_upscale_BiasAdd = _graph.node(id_pre_upscale_BiasAdd); - node_pre_upscale_BiasAdd->set_common_node_parameters(NodeParams{ "pre_upscale_BiasAdd", target }); + node_pre_upscale_BiasAdd->set_common_node_parameters(NodeParams{"pre_upscale_BiasAdd", target}); _graph.add_connection(id_add_16, 0, id_pre_upscale_BiasAdd, 0); _graph.add_connection(id_pre_upscale_FakeQuantWithMinMaxVars, 0, id_pre_upscale_BiasAdd, 1); _graph.add_connection(id_pre_upscale_Conv2D_bias, 0, id_pre_upscale_BiasAdd, 2); NodeID id_upscale_net_FakeQuantWithMinMaxVars_1 = _graph.add_node<DeconvolutionLayerNode>( - descriptors::DeconvolutionLayerDescriptor - { - PadStrideInfo{ - 2, 2, - 0, 0, - 0, 0, - DimensionRoundingType::FLOOR }, - QuantizationInfo{ 0.004990961868315935, 26 } }); + descriptors::DeconvolutionLayerDescriptor{PadStrideInfo{2, 2, 0, 0, 0, 0, DimensionRoundingType::FLOOR}, + QuantizationInfo{0.004990961868315935, 26}}); INode *node_upscale_net_FakeQuantWithMinMaxVars_1 = _graph.node(id_upscale_net_FakeQuantWithMinMaxVars_1); - node_upscale_net_FakeQuantWithMinMaxVars_1->set_common_node_parameters(NodeParams{ "upscale_net_FakeQuantWithMinMaxVars_1", target }); + node_upscale_net_FakeQuantWithMinMaxVars_1->set_common_node_parameters( + NodeParams{"upscale_net_FakeQuantWithMinMaxVars_1", target}); _graph.add_connection(id_pre_upscale_BiasAdd, 0, id_upscale_net_FakeQuantWithMinMaxVars_1, 0); - _graph.add_connection(id_upscale_net_FakeQuantWithMinMaxVars_transposed, 0, id_upscale_net_FakeQuantWithMinMaxVars_1, 1); + _graph.add_connection(id_upscale_net_FakeQuantWithMinMaxVars_transposed, 0, + id_upscale_net_FakeQuantWithMinMaxVars_1, 1); TensorShape output_shape; output_shape.set(0, 3, false).set(1, 720, false).set(2, 1280, false).set(3, 1, false); NodeID id_output_140211982446376 = _graph.add_node<OutputNode>(); INode *node_output_140211982446376 = _graph.node(id_output_140211982446376); - node_output_140211982446376->set_common_node_parameters(NodeParams{ "output_140211982446376", target }); + node_output_140211982446376->set_common_node_parameters(NodeParams{"output_140211982446376", target}); _graph.add_connection(id_upscale_net_FakeQuantWithMinMaxVars_1, 0, id_output_140211982446376, 0); - node_output_140211982446376->input(0)->set_accessor(get_npy_output_accessor(expected_output_filename.value(), output_shape, common_params.data_type, - common_params.data_layout)); + node_output_140211982446376->input(0)->set_accessor(get_npy_output_accessor( + expected_output_filename.value(), output_shape, common_params.data_type, common_params.data_layout)); return true; } diff --git a/examples/graph_googlenet.cpp b/examples/graph_googlenet.cpp index 683205b3b5..f431fc412b 100644 --- a/examples/graph_googlenet.cpp +++ b/examples/graph_googlenet.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -35,8 +36,7 @@ using namespace arm_compute::graph_utils; class GraphGooglenetExample : public Example { public: - GraphGooglenetExample() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "GoogleNet") + GraphGooglenetExample() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "GoogleNet") { } bool do_setup(int argc, char **argv) override @@ -49,14 +49,15 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; } // Checks - ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph"); + ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), + "QASYMM8 not supported for this graph"); // Print parameter values std::cout << common_params << std::endl; @@ -65,64 +66,99 @@ public: std::string data_path = common_params.data_path; // Create a preprocessor object - const std::array<float, 3> mean_rgb{ { 122.68f, 116.67f, 104.01f } }; + const std::array<float, 3> mean_rgb{{122.68f, 116.67f, 104.01f}}; std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<CaffePreproccessor>(mean_rgb); // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint + graph << common_params.target << common_params.fast_math_hint << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor))) + << ConvolutionLayer(7U, 7U, 64U, + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_b.npy"), + PadStrideInfo(2, 2, 3, 3)) + .set_name("conv1/7x7_s2") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv1/relu_7x7") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool1/3x3_s2") + << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)) + .set_name("pool1/norm1") << ConvolutionLayer( - 7U, 7U, 64U, - get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_b.npy"), - PadStrideInfo(2, 2, 3, 3)) - .set_name("conv1/7x7_s2") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1/relu_7x7") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool1/3x3_s2") - << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)).set_name("pool1/norm1") - << ConvolutionLayer( - 1U, 1U, 64U, - get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2/3x3_reduce") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2/relu_3x3_reduce") + 1U, 1U, 64U, + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2/3x3_reduce") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv2/relu_3x3_reduce") << ConvolutionLayer( - 3U, 3U, 192U, - get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2/3x3") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2/relu_3x3") - << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)).set_name("conv2/norm2") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool2/3x3_s2"); - graph << get_inception_node(data_path, "inception_3a", weights_layout, 64, std::make_tuple(96U, 128U), std::make_tuple(16U, 32U), 32U).set_name("inception_3a/concat"); - graph << get_inception_node(data_path, "inception_3b", weights_layout, 128, std::make_tuple(128U, 192U), std::make_tuple(32U, 96U), 64U).set_name("inception_3b/concat"); - graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool3/3x3_s2"); - graph << get_inception_node(data_path, "inception_4a", weights_layout, 192, std::make_tuple(96U, 208U), std::make_tuple(16U, 48U), 64U).set_name("inception_4a/concat"); - graph << get_inception_node(data_path, "inception_4b", weights_layout, 160, std::make_tuple(112U, 224U), std::make_tuple(24U, 64U), 64U).set_name("inception_4b/concat"); - graph << get_inception_node(data_path, "inception_4c", weights_layout, 128, std::make_tuple(128U, 256U), std::make_tuple(24U, 64U), 64U).set_name("inception_4c/concat"); - graph << get_inception_node(data_path, "inception_4d", weights_layout, 112, std::make_tuple(144U, 288U), std::make_tuple(32U, 64U), 64U).set_name("inception_4d/concat"); - graph << get_inception_node(data_path, "inception_4e", weights_layout, 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U).set_name("inception_4e/concat"); - graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool4/3x3_s2"); - graph << get_inception_node(data_path, "inception_5a", weights_layout, 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U).set_name("inception_5a/concat"); - graph << get_inception_node(data_path, "inception_5b", weights_layout, 384, std::make_tuple(192U, 384U), std::make_tuple(48U, 128U), 128U).set_name("inception_5b/concat"); - graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 7, operation_layout, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL))).set_name("pool5/7x7_s1") + 3U, 3U, 192U, + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_b.npy"), + PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2/3x3") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv2/relu_3x3") + << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)) + .set_name("conv2/norm2") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool2/3x3_s2"); + graph << get_inception_node(data_path, "inception_3a", weights_layout, 64, std::make_tuple(96U, 128U), + std::make_tuple(16U, 32U), 32U) + .set_name("inception_3a/concat"); + graph << get_inception_node(data_path, "inception_3b", weights_layout, 128, std::make_tuple(128U, 192U), + std::make_tuple(32U, 96U), 64U) + .set_name("inception_3b/concat"); + graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool3/3x3_s2"); + graph << get_inception_node(data_path, "inception_4a", weights_layout, 192, std::make_tuple(96U, 208U), + std::make_tuple(16U, 48U), 64U) + .set_name("inception_4a/concat"); + graph << get_inception_node(data_path, "inception_4b", weights_layout, 160, std::make_tuple(112U, 224U), + std::make_tuple(24U, 64U), 64U) + .set_name("inception_4b/concat"); + graph << get_inception_node(data_path, "inception_4c", weights_layout, 128, std::make_tuple(128U, 256U), + std::make_tuple(24U, 64U), 64U) + .set_name("inception_4c/concat"); + graph << get_inception_node(data_path, "inception_4d", weights_layout, 112, std::make_tuple(144U, 288U), + std::make_tuple(32U, 64U), 64U) + .set_name("inception_4d/concat"); + graph << get_inception_node(data_path, "inception_4e", weights_layout, 256, std::make_tuple(160U, 320U), + std::make_tuple(32U, 128U), 128U) + .set_name("inception_4e/concat"); + graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool4/3x3_s2"); + graph << get_inception_node(data_path, "inception_5a", weights_layout, 256, std::make_tuple(160U, 320U), + std::make_tuple(32U, 128U), 128U) + .set_name("inception_5a/concat"); + graph << get_inception_node(data_path, "inception_5b", weights_layout, 384, std::make_tuple(192U, 384U), + std::make_tuple(48U, 128U), 128U) + .set_name("inception_5b/concat"); + graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 7, operation_layout, + PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool5/7x7_s1") << FullyConnectedLayer( - 1000U, - get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_b.npy")) - .set_name("loss3/classifier") - << SoftmaxLayer().set_name("prob") - << OutputLayer(get_output_accessor(common_params, 5)); + 1000U, + get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_b.npy")) + .set_name("loss3/classifier") + << SoftmaxLayer().set_name("prob") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph GraphConfig config; @@ -148,63 +184,63 @@ private: CommonGraphParams common_params; Stream graph; - ConcatLayer get_inception_node(const std::string &data_path, std::string &¶m_path, DataLayout weights_layout, - unsigned int a_filt, + ConcatLayer get_inception_node(const std::string &data_path, + std::string &¶m_path, + DataLayout weights_layout, + unsigned int a_filt, std::tuple<unsigned int, unsigned int> b_filters, std::tuple<unsigned int, unsigned int> c_filters, - unsigned int d_filt) + unsigned int d_filt) { std::string total_path = "/cnn_data/googlenet_model/" + param_path + "/" + param_path + "_"; SubStream i_a(graph); - i_a << ConvolutionLayer( - 1U, 1U, a_filt, - get_weights_accessor(data_path, total_path + "1x1_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_1x1"); + i_a << ConvolutionLayer(1U, 1U, a_filt, + get_weights_accessor(data_path, total_path + "1x1_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "1x1_b.npy"), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_1x1"); SubStream i_b(graph); - i_b << ConvolutionLayer( - 1U, 1U, std::get<0>(b_filters), - get_weights_accessor(data_path, total_path + "3x3_reduce_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "3x3_reduce_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/3x3_reduce") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_3x3_reduce") - << ConvolutionLayer( - 3U, 3U, std::get<1>(b_filters), - get_weights_accessor(data_path, total_path + "3x3_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "3x3_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name(param_path + "/3x3") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_3x3"); + i_b << ConvolutionLayer(1U, 1U, std::get<0>(b_filters), + get_weights_accessor(data_path, total_path + "3x3_reduce_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "3x3_reduce_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/3x3_reduce") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_3x3_reduce") + << ConvolutionLayer(3U, 3U, std::get<1>(b_filters), + get_weights_accessor(data_path, total_path + "3x3_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "3x3_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name(param_path + "/3x3") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_3x3"); SubStream i_c(graph); - i_c << ConvolutionLayer( - 1U, 1U, std::get<0>(c_filters), - get_weights_accessor(data_path, total_path + "5x5_reduce_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "5x5_reduce_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/5x5_reduce") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_5x5_reduce") - << ConvolutionLayer( - 5U, 5U, std::get<1>(c_filters), - get_weights_accessor(data_path, total_path + "5x5_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "5x5_b.npy"), - PadStrideInfo(1, 1, 2, 2)) - .set_name(param_path + "/5x5") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_5x5"); + i_c << ConvolutionLayer(1U, 1U, std::get<0>(c_filters), + get_weights_accessor(data_path, total_path + "5x5_reduce_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "5x5_reduce_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/5x5_reduce") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_5x5_reduce") + << ConvolutionLayer(5U, 5U, std::get<1>(c_filters), + get_weights_accessor(data_path, total_path + "5x5_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "5x5_b.npy"), PadStrideInfo(1, 1, 2, 2)) + .set_name(param_path + "/5x5") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_5x5"); SubStream i_d(graph); - i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL))).set_name(param_path + "/pool") + i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, + PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL))) + .set_name(param_path + "/pool") << ConvolutionLayer( - 1U, 1U, d_filt, - get_weights_accessor(data_path, total_path + "pool_proj_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "pool_proj_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/pool_proj") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_pool_proj"); + 1U, 1U, d_filt, get_weights_accessor(data_path, total_path + "pool_proj_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "pool_proj_b.npy"), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/pool_proj") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_pool_proj"); return ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)); } diff --git a/examples/graph_inception_resnet_v1.cpp b/examples/graph_inception_resnet_v1.cpp index d789d7f6e7..a54a0f7806 100644 --- a/examples/graph_inception_resnet_v1.cpp +++ b/examples/graph_inception_resnet_v1.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -38,7 +39,12 @@ class InceptionResNetV1Example final : public Example { public: InceptionResNetV1Example() - : cmd_parser(), common_opts(cmd_parser), common_params(), model_input_width(nullptr), model_input_height(nullptr), graph(0, "InceptionResNetV1") + : cmd_parser(), + common_opts(cmd_parser), + common_params(), + model_input_width(nullptr), + model_input_height(nullptr), + graph(0, "InceptionResNetV1") { model_input_width = cmd_parser.add_option<SimpleOption<unsigned int>>("image-width", 512); model_input_height = cmd_parser.add_option<SimpleOption<unsigned int>>("image-height", 512); @@ -47,7 +53,7 @@ public: model_input_width->set_help("Input image width."); model_input_height->set_help("Input image height."); } - InceptionResNetV1Example(const InceptionResNetV1Example &) = delete; + InceptionResNetV1Example(const InceptionResNetV1Example &) = delete; InceptionResNetV1Example &operator=(const InceptionResNetV1Example &) = delete; ~InceptionResNetV1Example() override = default; bool do_setup(int argc, char **argv) override @@ -60,7 +66,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -70,13 +76,14 @@ public: const unsigned int image_height = model_input_height->value(); // Set default layout if needed - if(!common_opts.data_layout->is_set() && common_params.target == Target::NEON) + if (!common_opts.data_layout->is_set() && common_params.target == Target::NEON) { common_params.data_layout = DataLayout::NCHW; } // Checks - ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph"); + ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), + "QASYMM8 not supported for this graph"); // Print parameter values std::cout << common_params << std::endl; @@ -86,7 +93,7 @@ public: // Create model path std::string data_path = common_params.data_path; std::string model_path = "/cnn_data/inception_resnet_v1_model/"; - if(!data_path.empty()) + if (!data_path.empty()) { data_path += model_path; } @@ -96,95 +103,98 @@ public: // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(image_width, image_height, 3U, common_params.batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = permute_shape( + TensorShape(image_width, image_height, 3U, common_params.batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint + graph << common_params.target << common_params.fast_math_hint << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false)) // Conv2d_1a_3x3 - << ConvolutionLayer(3U, 3U, 32U, - get_weights_accessor(data_path, "Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name("Conv2d_1a_3x3/convolution") + << ConvolutionLayer( + 3U, 3U, 32U, get_weights_accessor(data_path, "Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Conv2d_1a_3x3/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "Conv2d_1a_3x3_BatchNorm_beta.npy"), batch_norm_epsilon) - .set_name("Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_1a_3x3/Relu") + .set_name("Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_1a_3x3/Relu") // Conv2d_2a_3x3 - << ConvolutionLayer(3U, 3U, 32U, - get_weights_accessor(data_path, "Conv2d_2a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Conv2d_2a_3x3/convolution") + << ConvolutionLayer( + 3U, 3U, 32U, get_weights_accessor(data_path, "Conv2d_2a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Conv2d_2a_3x3/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_2a_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "Conv2d_2a_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "Conv2d_2a_3x3_BatchNorm_beta.npy"), batch_norm_epsilon) - .set_name("Conv2d_2a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2a_3x3/Relu") + .set_name("Conv2d_2a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_2a_3x3/Relu") // Conv2d_2b_3x3 - << ConvolutionLayer(3U, 3U, 64U, - get_weights_accessor(data_path, "Conv2d_2b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("Conv2d_2b_3x3/convolution") + << ConvolutionLayer( + 3U, 3U, 64U, get_weights_accessor(data_path, "Conv2d_2b_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("Conv2d_2b_3x3/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_2b_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "Conv2d_2b_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "Conv2d_2b_3x3_BatchNorm_beta.npy"), batch_norm_epsilon) - .set_name("Conv2d_2b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2b_3x3/Relu") + .set_name("Conv2d_2b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_2b_3x3/Relu") // MaxPool_3a_3x3 - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)).set_name("MaxPool_3a_3x3/MaxPool") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)) + .set_name("MaxPool_3a_3x3/MaxPool") // Conv2d_3b_1x1 - << ConvolutionLayer(1U, 1U, 80U, - get_weights_accessor(data_path, "Conv2d_3b_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Conv2d_3b_1x1/convolution") + << ConvolutionLayer( + 1U, 1U, 80U, get_weights_accessor(data_path, "Conv2d_3b_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Conv2d_3b_1x1/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_3b_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "Conv2d_3b_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "Conv2d_3b_1x1_BatchNorm_beta.npy"), batch_norm_epsilon) - .set_name("Conv2d_3b_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_3b_1x1/Relu") + .set_name("Conv2d_3b_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_3b_1x1/Relu") // Conv2d_4a_3x3 - << ConvolutionLayer(3U, 3U, 192U, - get_weights_accessor(data_path, "Conv2d_4a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Conv2d_4a_3x3/convolution") + << ConvolutionLayer( + 3U, 3U, 192U, get_weights_accessor(data_path, "Conv2d_4a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Conv2d_4a_3x3/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_4a_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "Conv2d_4a_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "Conv2d_4a_3x3_BatchNorm_beta.npy"), batch_norm_epsilon) - .set_name("Conv2d_4a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_4a_3x3/Relu") + .set_name("Conv2d_4a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_4a_3x3/Relu") // Conv2d_4b_3x3 - << ConvolutionLayer(3U, 3U, 256U, - get_weights_accessor(data_path, "Conv2d_4b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name("Conv2d_4a_3x3/convolution") + << ConvolutionLayer( + 3U, 3U, 256U, get_weights_accessor(data_path, "Conv2d_4b_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Conv2d_4a_3x3/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_4b_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "Conv2d_4b_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "Conv2d_4b_3x3_BatchNorm_beta.npy"), batch_norm_epsilon) - .set_name("Conv2d_4b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_4b_3x3/Relu"); + .set_name("Conv2d_4b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_4b_3x3/Relu"); // 5 x Inception-resnet-A block35_repeat(data_path, weights_layout, 5); @@ -202,11 +212,9 @@ public: // Logits tail graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("Logits/AvgPool_1a_8x8") << FlattenLayer().set_name("Logits/Flatten") - << FullyConnectedLayer( - 128U, - get_weights_accessor(data_path, "Logits_Logits_weights.npy", weights_layout), - get_weights_accessor(data_path, "Logits_Logits_biases.npy")) - .set_name("Logits/Logits") + << FullyConnectedLayer(128U, get_weights_accessor(data_path, "Logits_Logits_weights.npy", weights_layout), + get_weights_accessor(data_path, "Logits_Logits_biases.npy")) + .set_name("Logits/Logits") << OutputLayer(std::make_unique<DummyAccessor>(0)); // Finalize graph @@ -231,14 +239,14 @@ private: CommandLineParser cmd_parser; CommonGraphOptions common_opts; CommonGraphParams common_params; - SimpleOption<unsigned int> *model_input_width{ nullptr }; - SimpleOption<unsigned int> *model_input_height{ nullptr }; + SimpleOption<unsigned int> *model_input_width{nullptr}; + SimpleOption<unsigned int> *model_input_height{nullptr}; Stream graph; private: void block35_repeat(const std::string &data_path, DataLayout weights_layout, unsigned int num_blocks) { - for(unsigned int i = 0; i < num_blocks; ++i) + for (unsigned int i = 0; i < num_blocks; ++i) { std::stringstream unit_path_ss; unit_path_ss << "Repeat_block35_" << (i + 1) << "_"; @@ -254,102 +262,128 @@ private: // Branch 0 SubStream i_la(i_l); - i_la << ConvolutionLayer(1U, 1U, 32U, - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_0/Conv2d_1x1/Relu"); + i_la << ConvolutionLayer( + 1U, 1U, 32U, + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_0/Conv2d_1x1/Relu"); // Branch 1 SubStream i_lb(i_l); i_lb << ConvolutionLayer(1U, 1U, 32U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu") + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu") << ConvolutionLayer(3U, 3U, 32U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) - .set_name(unit_name + "Branch_1/Conv2d_0b_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name(unit_name + "Branch_1/Conv2d_0b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0b_3x3/Relu"); + .set_name(unit_name + "Branch_1/Conv2d_0b_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name(unit_name + "Branch_1/Conv2d_0b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0b_3x3/Relu"); // Branch 2 SubStream i_lc(i_l); i_lc << ConvolutionLayer(1U, 1U, 32U, - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Branch_2/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name(unit_name + "Branch_2/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_2/Conv2d_0a_1x1/Relu") + .set_name(unit_name + "Branch_2/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name(unit_name + "Branch_2/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_2/Conv2d_0a_1x1/Relu") << ConvolutionLayer(3U, 3U, 32U, - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) - .set_name(unit_name + "Branch_2/Conv2d_0b_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name(unit_name + "Branch_2/Conv2d_0b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_2/Conv2d_0b_3x3/Relu") + .set_name(unit_name + "Branch_2/Conv2d_0b_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name(unit_name + "Branch_2/Conv2d_0b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_2/Conv2d_0b_3x3/Relu") << ConvolutionLayer(3U, 3U, 32U, - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) - .set_name(unit_name + "Branch_2/Conv2d_0c_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name(unit_name + "Branch_2/Conv2d_0c_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_2/Conv2d_0c_3x3/Relu"); + .set_name(unit_name + "Branch_2/Conv2d_0c_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name(unit_name + "Branch_2/Conv2d_0c_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_2/Conv2d_0c_3x3/Relu"); // Concatenate i_l << ConcatLayer(std::move(i_la), std::move(i_lb), std::move(i_lc)).set_name(unit_name + "concat") - << ConvolutionLayer(1U, 1U, 256U, - get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout), - get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Conv2d_1x1/convolution") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 0.17f, 0.f)).set_name(unit_name + "mul"); + << ConvolutionLayer( + 1U, 1U, 256U, + get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout), + PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "Conv2d_1x1/convolution") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 0.17f, 0.f)) + .set_name(unit_name + "mul"); graph << EltwiseLayer(std::move(i_l), std::move(i_r), EltwiseOperation::Add).set_name(unit_name + "add") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu"); + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Relu"); } } void block17_repeat(const std::string &data_path, DataLayout weights_layout, unsigned int num_blocks) { - for(unsigned int i = 0; i < num_blocks; ++i) + for (unsigned int i = 0; i < num_blocks; ++i) { std::stringstream unit_path_ss; unit_path_ss << "Repeat_1_block17_" << (i + 1) << "_"; @@ -365,79 +399,101 @@ private: // Branch 0 SubStream i_la(i_l); - i_la << ConvolutionLayer(1U, 1U, 128U, - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_0/Conv2d_1x1/Relu"); + i_la << ConvolutionLayer( + 1U, 1U, 128U, + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_0/Conv2d_1x1/Relu"); // Branch 1 SubStream i_lb(i_l); i_lb << ConvolutionLayer(1U, 1U, 128U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu") + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu") << ConvolutionLayer(7U, 1U, 128U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) - .set_name(unit_name + "Branch_1/Conv2d_0b_1x7/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name(unit_name + "Branch_1/Conv2d_0b_1x7/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0b_1x7/Relu") + .set_name(unit_name + "Branch_1/Conv2d_0b_1x7/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name(unit_name + "Branch_1/Conv2d_0b_1x7/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0b_1x7/Relu") << ConvolutionLayer(1U, 7U, 128U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) - .set_name(unit_name + "Branch_1/Conv2d_0c_7x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name(unit_name + "Branch_1/Conv2d_0c_7x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0c_7x1/Relu"); + .set_name(unit_name + "Branch_1/Conv2d_0c_7x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name(unit_name + "Branch_1/Conv2d_0c_7x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0c_7x1/Relu"); // Concatenate i_l << ConcatLayer(std::move(i_la), std::move(i_lb)).set_name(unit_name + "concat") - << ConvolutionLayer(1U, 1U, 896U, - get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout), - get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Conv2d_1x1/convolution") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 0.10f, 0.f)).set_name(unit_name + "mul"); + << ConvolutionLayer( + 1U, 1U, 896U, + get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout), + PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "Conv2d_1x1/convolution") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 0.10f, 0.f)) + .set_name(unit_name + "mul"); graph << EltwiseLayer(std::move(i_l), std::move(i_r), EltwiseOperation::Add).set_name(unit_name + "add") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu"); + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Relu"); } } - void block8_repeat(const std::string &data_path, DataLayout weights_layout, unsigned int num_blocks, float scale, bool has_activation) + void block8_repeat(const std::string &data_path, + DataLayout weights_layout, + unsigned int num_blocks, + float scale, + bool has_activation) { - for(unsigned int i = 0; i < num_blocks; ++i) + for (unsigned int i = 0; i < num_blocks; ++i) { std::stringstream unit_path_ss; std::stringstream unit_name_ss; - if(num_blocks != 1) + if (num_blocks != 1) { unit_path_ss << "Repeat_2_block8_" << (i + 1) << "_"; unit_name_ss << "Repeat_2/block8_" << (i + 1) << "/"; @@ -457,79 +513,97 @@ private: // Branch 0 SubStream i_la(i_l); - i_la << ConvolutionLayer(1U, 1U, 192U, - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_0/Conv2d_1x1/Relu"); + i_la << ConvolutionLayer( + 1U, 1U, 192U, + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_0/Conv2d_1x1/Relu"); // Branch 1 SubStream i_lb(i_l); i_lb << ConvolutionLayer(1U, 1U, 192U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu") + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu") << ConvolutionLayer(3U, 1U, 192U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 0)) - .set_name(unit_name + "Branch_1/Conv2d_0b_1x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name(unit_name + "Branch_1/Conv2d_0b_1x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0b_1x3/Relu") + .set_name(unit_name + "Branch_1/Conv2d_0b_1x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name(unit_name + "Branch_1/Conv2d_0b_1x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0b_1x3/Relu") << ConvolutionLayer(1U, 3U, 192U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 1)) - .set_name(unit_name + "Branch_1/Conv2d_0c_3x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name(unit_name + "Branch_1/Conv2d_0c_3x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0c_3x1/Relu"); + .set_name(unit_name + "Branch_1/Conv2d_0c_3x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name(unit_name + "Branch_1/Conv2d_0c_3x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0c_3x1/Relu"); // Concatenate i_l << ConcatLayer(std::move(i_la), std::move(i_lb)).set_name(unit_name + "concat") - << ConvolutionLayer(1U, 1U, 1792U, - get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout), - get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Conv2d_1x1/convolution"); + << ConvolutionLayer( + 1U, 1U, 1792U, + get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout), + PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "Conv2d_1x1/convolution"); // Scale result - if(scale != 1.f) + if (scale != 1.f) { - i_l << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, scale, 0.f)).set_name(unit_name + "mul"); + i_l << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, scale, 0.f)) + .set_name(unit_name + "mul"); } // Residual add graph << EltwiseLayer(std::move(i_l), std::move(i_r), EltwiseOperation::Add).set_name(unit_name + "add"); // Apply activation if needed - if(has_activation) + if (has_activation) { - graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu"); + graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Relu"); } } } @@ -538,61 +612,71 @@ private: { // Branch 0 SubStream i_a(graph); - i_a << ConvolutionLayer(3U, 3U, 384U, - get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/Relu"); + i_a << ConvolutionLayer( + 3U, 3U, 384U, + get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/Relu"); // Branch 1 SubStream i_b(graph); - i_b << ConvolutionLayer(1U, 1U, 192U, - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(3U, 3U, 192U, - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/Relu") - << ConvolutionLayer(3U, 3U, 256U, - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/Relu"); + i_b << ConvolutionLayer( + 1U, 1U, 192U, + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 3U, 3U, 192U, + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/Relu") + << ConvolutionLayer( + 3U, 3U, 256U, + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/Relu"); // Branch 2 SubStream i_c(graph); - i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(2, 2, 0, 0), true)).set_name("Mixed_6a/Branch_2/MaxPool_1a_3x3"); + i_c << PoolingLayer( + PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(2, 2, 0, 0), true)) + .set_name("Mixed_6a/Branch_2/MaxPool_1a_3x3"); // Concatenate graph << ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c)).set_name("Mixed_6a/concat"); @@ -602,103 +686,120 @@ private: { // Branch 0 SubStream i_a(graph); - i_a << ConvolutionLayer(1U, 1U, 256U, - get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(3U, 3U, 384U, - get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/Relu"); + i_a << ConvolutionLayer( + 1U, 1U, 256U, + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 3U, 3U, 384U, + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/Relu"); // Branch 1 SubStream i_b(graph); - i_b << ConvolutionLayer(1U, 1U, 256U, - get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(3U, 3U, 256U, - get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/Relu"); + i_b << ConvolutionLayer( + 1U, 1U, 256U, + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 3U, 3U, 256U, + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/Relu"); // Branch 2 SubStream i_c(graph); - i_c << ConvolutionLayer(1U, 1U, 256U, - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(3U, 3U, 256U, - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/Relu") - << ConvolutionLayer(3U, 3U, 256U, - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_beta.npy"), - batch_norm_epsilon) - .set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/Relu"); + i_c << ConvolutionLayer( + 1U, 1U, 256U, + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 3U, 3U, 256U, + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/Relu") + << ConvolutionLayer( + 3U, 3U, 256U, + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_beta.npy"), + batch_norm_epsilon) + .set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/Relu"); // Branch 3 SubStream i_d(graph); - i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(2, 2, 0, 0), true)).set_name("Mixed_7a/Branch_3/MaxPool_1a_3x3"); + i_d << PoolingLayer( + PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(2, 2, 0, 0), true)) + .set_name("Mixed_7a/Branch_3/MaxPool_1a_3x3"); // Concatenate - graph << ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)).set_name("Mixed_7a/concat"); + graph + << ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)).set_name("Mixed_7a/concat"); } }; diff --git a/examples/graph_inception_resnet_v2.cpp b/examples/graph_inception_resnet_v2.cpp index 1d0c51e9ad..43e31ee14b 100644 --- a/examples/graph_inception_resnet_v2.cpp +++ b/examples/graph_inception_resnet_v2.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -35,8 +36,7 @@ using namespace arm_compute::graph_utils; class InceptionResNetV2Example final : public Example { public: - InceptionResNetV2Example() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "InceptionResNetV2") + InceptionResNetV2Example() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "InceptionResNetV2") { } bool do_setup(int argc, char **argv) override @@ -49,20 +49,21 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; } // Set default layout if needed - if(!common_opts.data_layout->is_set() && common_params.target == Target::NEON) + if (!common_opts.data_layout->is_set() && common_params.target == Target::NEON) { common_params.data_layout = DataLayout::NCHW; } // Checks - ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph"); + ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), + "QASYMM8 not supported for this graph"); // Print parameter values std::cout << common_params << std::endl; @@ -70,7 +71,7 @@ public: // Create model path std::string data_path = common_params.data_path; std::string model_path = "/cnn_data/inception_resnet_v2_model/"; - if(!data_path.empty()) + if (!data_path.empty()) { data_path += model_path; } @@ -80,84 +81,88 @@ public: // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(299U, 299U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(299U, 299U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint + graph << common_params.target << common_params.fast_math_hint << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false)) // Conv2d_1a_3x3 - << ConvolutionLayer(3U, 3U, 32U, - get_weights_accessor(data_path, "Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name("Conv2d_1a_3x3/convolution") + << ConvolutionLayer( + 3U, 3U, 32U, get_weights_accessor(data_path, "Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Conv2d_1a_3x3/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "Conv2d_1a_3x3_BatchNorm_beta.npy"), 0.0010000000474974513f) - .set_name("Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_1a_3x3/Relu") + .set_name("Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_1a_3x3/Relu") // Conv2d_2a_3x3 - << ConvolutionLayer(3U, 3U, 32U, - get_weights_accessor(data_path, "Conv2d_2a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Conv2d_2a_3x3/convolution") + << ConvolutionLayer( + 3U, 3U, 32U, get_weights_accessor(data_path, "Conv2d_2a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Conv2d_2a_3x3/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_2a_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "Conv2d_2a_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "Conv2d_2a_3x3_BatchNorm_beta.npy"), 0.0010000000474974513f) - .set_name("Conv2d_2a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2a_3x3/Relu") + .set_name("Conv2d_2a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_2a_3x3/Relu") // Conv2d_2b_3x3 - << ConvolutionLayer(3U, 3U, 64U, - get_weights_accessor(data_path, "Conv2d_2b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("Conv2d_2b_3x3/convolution") + << ConvolutionLayer( + 3U, 3U, 64U, get_weights_accessor(data_path, "Conv2d_2b_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("Conv2d_2b_3x3/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_2b_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "Conv2d_2b_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "Conv2d_2b_3x3_BatchNorm_beta.npy"), 0.0010000000474974513f) - .set_name("Conv2d_2b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2b_3x3/Relu") + .set_name("Conv2d_2b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_2b_3x3/Relu") // MaxPool_3a_3x3 - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)).set_name("MaxPool_3a_3x3/MaxPool") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)) + .set_name("MaxPool_3a_3x3/MaxPool") // Conv2d_3b_1x1 - << ConvolutionLayer(1U, 1U, 80U, - get_weights_accessor(data_path, "Conv2d_3b_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Conv2d_3b_1x1/convolution") + << ConvolutionLayer( + 1U, 1U, 80U, get_weights_accessor(data_path, "Conv2d_3b_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Conv2d_3b_1x1/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_3b_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "Conv2d_3b_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "Conv2d_3b_1x1_BatchNorm_beta.npy"), 0.0010000000474974513f) - .set_name("Conv2d_3b_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_3b_1x1/Relu") + .set_name("Conv2d_3b_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_3b_1x1/Relu") // Conv2d_4a_3x3 - << ConvolutionLayer(3U, 3U, 192U, - get_weights_accessor(data_path, "Conv2d_4a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Conv2d_4a_3x3/convolution") + << ConvolutionLayer( + 3U, 3U, 192U, get_weights_accessor(data_path, "Conv2d_4a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Conv2d_4a_3x3/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_4a_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "Conv2d_4a_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "Conv2d_4a_3x3_BatchNorm_beta.npy"), 0.0010000000474974513f) - .set_name("Conv2d_4a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_4a_3x3/Relu") + .set_name("Conv2d_4a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_4a_3x3/Relu") // MaxPool_5a_3x3 - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0), true)).set_name("MaxPool_5a_3x3/MaxPool"); + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0), true)) + .set_name("MaxPool_5a_3x3/MaxPool"); block_mixed_5b(data_path, weights_layout); block35_repeat(data_path, weights_layout, 10); @@ -168,27 +173,25 @@ public: block8_repeat(data_path, weights_layout, 1, 1.f, false); // Conv2d_7b_1x1 - graph << ConvolutionLayer(1U, 1U, 1536U, - get_weights_accessor(data_path, "Conv2d_7b_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Conv2d_7b_1x1/convolution") + graph << ConvolutionLayer( + 1U, 1U, 1536U, get_weights_accessor(data_path, "Conv2d_7b_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Conv2d_7b_1x1/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_7b_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "Conv2d_7b_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "Conv2d_7b_1x1_BatchNorm_beta.npy"), 0.0010000000474974513f) - .set_name("Conv2d_7b_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_7b_1x1/Relu") + .set_name("Conv2d_7b_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_7b_1x1/Relu") << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("Logits/AvgPool_1a_8x8") << FlattenLayer().set_name("Logits/Flatten") - << FullyConnectedLayer( - 1001U, - get_weights_accessor(data_path, "Logits_Logits_weights.npy", weights_layout), - get_weights_accessor(data_path, "Logits_Logits_biases.npy")) - .set_name("Logits/Logits") - << SoftmaxLayer().set_name("Logits/Predictions") - << OutputLayer(get_output_accessor(common_params, 5)); + << FullyConnectedLayer(1001U, + get_weights_accessor(data_path, "Logits_Logits_weights.npy", weights_layout), + get_weights_accessor(data_path, "Logits_Logits_biases.npy")) + .set_name("Logits/Logits") + << SoftmaxLayer().set_name("Logits/Predictions") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph GraphConfig config; @@ -219,164 +222,191 @@ private: { // Branch 0 SubStream i_a(graph); - i_a << ConvolutionLayer(1U, 1U, 96U, - get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_5b/Branch_0/Conv2d_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_5b/Branch_0/Conv2d_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_0/Conv2d_1x1/Relu"); + i_a << ConvolutionLayer( + 1U, 1U, 96U, + get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_5b/Branch_0/Conv2d_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_5b/Branch_0/Conv2d_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_5b/Branch_0/Conv2d_1x1/Relu"); // Branch 1 SubStream i_b(graph); - i_b << ConvolutionLayer(1U, 1U, 48U, - get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_5b/Branch_1/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_5b/Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_1/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(5U, 5U, 64U, - get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 2, 2)) - .set_name("Mixed_5b/Branch_1/Conv2d_0b_5x5/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_5b/Branch_1/Conv2d_0b_5x5/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_1/Conv2d_0b_5x5/Relu"); + i_b << ConvolutionLayer( + 1U, 1U, 48U, + get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_5b/Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_5b/Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_5b/Branch_1/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 5U, 5U, 64U, + get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 2, 2)) + .set_name("Mixed_5b/Branch_1/Conv2d_0b_5x5/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_5b/Branch_1/Conv2d_0b_5x5/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_5b/Branch_1/Conv2d_0b_5x5/Relu"); // Branch 2 SubStream i_c(graph); - i_c << ConvolutionLayer(1U, 1U, 64U, - get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_5b/Branch_2/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_5b/Branch_2/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_2/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(3U, 3U, 96U, - get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("Mixed_5b/Branch_2/Conv2d_0b_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_5b/Branch_2/Conv2d_0b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_2/Conv2d_0b_3x3/Relu") - << ConvolutionLayer(3U, 3U, 96U, - get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("Mixed_5b/Branch_2/Conv2d_0c_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_5b/Branch_2/Conv2d_0c_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_2/Conv2d_0c_3x3/Relu"); + i_c << ConvolutionLayer( + 1U, 1U, 64U, + get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_5b/Branch_2/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_5b/Branch_2/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_5b/Branch_2/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 3U, 3U, 96U, + get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("Mixed_5b/Branch_2/Conv2d_0b_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_5b/Branch_2/Conv2d_0b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_5b/Branch_2/Conv2d_0b_3x3/Relu") + << ConvolutionLayer( + 3U, 3U, 96U, + get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("Mixed_5b/Branch_2/Conv2d_0c_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_5b/Branch_2/Conv2d_0c_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_5b/Branch_2/Conv2d_0c_3x3/Relu"); // Branch 3 SubStream i_d(graph); - i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)).set_name("Mixed_5b/Branch_3/AvgPool_0a_3x3") - << ConvolutionLayer(1U, 1U, 64U, - get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_5b/Branch_3/Conv2d_0b_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_5b/Branch_3/Conv2d_0b_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_3/Conv2d_0b_1x1/Relu"); + i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, + PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) + .set_name("Mixed_5b/Branch_3/AvgPool_0a_3x3") + << ConvolutionLayer( + 1U, 1U, 64U, + get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_5b/Branch_3/Conv2d_0b_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_5b/Branch_3/Conv2d_0b_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_5b/Branch_3/Conv2d_0b_1x1/Relu"); // Concatenate - graph << ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)).set_name("Mixed_5a/concat"); + graph + << ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)).set_name("Mixed_5a/concat"); } void block_mixed_6a(const std::string &data_path, DataLayout weights_layout) { // Branch 0 SubStream i_a(graph); - i_a << ConvolutionLayer(3U, 3U, 384U, - get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/Relu"); + i_a << ConvolutionLayer( + 3U, 3U, 384U, + get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/Relu"); // Branch 1 SubStream i_b(graph); - i_b << ConvolutionLayer(1U, 1U, 256U, - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(3U, 3U, 256U, - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/Relu") - << ConvolutionLayer(3U, 3U, 384U, - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/Relu"); + i_b << ConvolutionLayer( + 1U, 1U, 256U, + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 3U, 3U, 256U, + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/Relu") + << ConvolutionLayer( + 3U, 3U, 384U, + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/Relu"); // Branch 2 SubStream i_c(graph); - i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(2, 2, 0, 0), true)).set_name("Mixed_6a/Branch_2/MaxPool_1a_3x3"); + i_c << PoolingLayer( + PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(2, 2, 0, 0), true)) + .set_name("Mixed_6a/Branch_2/MaxPool_1a_3x3"); // Concatenate graph << ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c)).set_name("Mixed_6a/concat"); @@ -386,108 +416,125 @@ private: { // Branch 0 SubStream i_a(graph); - i_a << ConvolutionLayer(1U, 1U, 256U, - get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(3U, 3U, 384U, - get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/Relu"); + i_a << ConvolutionLayer( + 1U, 1U, 256U, + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 3U, 3U, 384U, + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/Relu"); // Branch 1 SubStream i_b(graph); - i_b << ConvolutionLayer(1U, 1U, 256U, - get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(3U, 3U, 288U, - get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/Relu"); + i_b << ConvolutionLayer( + 1U, 1U, 256U, + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 3U, 3U, 288U, + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/Relu"); // Branch 2 SubStream i_c(graph); - i_c << ConvolutionLayer(1U, 1U, 256U, - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(3U, 3U, 288U, - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/Relu") - << ConvolutionLayer(3U, 3U, 320U, - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/Relu"); + i_c << ConvolutionLayer( + 1U, 1U, 256U, + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 3U, 3U, 288U, + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/Relu") + << ConvolutionLayer( + 3U, 3U, 320U, + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/Relu"); // Branch 3 SubStream i_d(graph); - i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)).set_name("Mixed_7a/Branch_3/MaxPool_1a_3x3"); + i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)) + .set_name("Mixed_7a/Branch_3/MaxPool_1a_3x3"); // Concatenate - graph << ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)).set_name("Mixed_7a/concat"); + graph + << ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)).set_name("Mixed_7a/concat"); } void block35_repeat(const std::string &data_path, DataLayout weights_layout, unsigned int num_blocks) { - for(unsigned int i = 0; i < num_blocks; ++i) + for (unsigned int i = 0; i < num_blocks; ++i) { std::stringstream unit_path_ss; unit_path_ss << "Repeat_block35_" << (i + 1) << "_"; @@ -503,102 +550,128 @@ private: // Branch 0 SubStream i_la(i_l); - i_la << ConvolutionLayer(1U, 1U, 32U, - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_0/Conv2d_1x1/Relu"); + i_la << ConvolutionLayer( + 1U, 1U, 32U, + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_0/Conv2d_1x1/Relu"); // Branch 1 SubStream i_lb(i_l); i_lb << ConvolutionLayer(1U, 1U, 32U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu") + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu") << ConvolutionLayer(3U, 3U, 32U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) - .set_name(unit_name + "Branch_1/Conv2d_0b_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(unit_name + "Branch_1/Conv2d_0b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0b_3x3/Relu"); + .set_name(unit_name + "Branch_1/Conv2d_0b_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(unit_name + "Branch_1/Conv2d_0b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0b_3x3/Relu"); // Branch 2 SubStream i_lc(i_l); i_lc << ConvolutionLayer(1U, 1U, 32U, - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Branch_2/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(unit_name + "Branch_2/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_2/Conv2d_0a_1x1/Relu") + .set_name(unit_name + "Branch_2/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(unit_name + "Branch_2/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_2/Conv2d_0a_1x1/Relu") << ConvolutionLayer(3U, 3U, 48U, - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) - .set_name(unit_name + "Branch_2/Conv2d_0b_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(unit_name + "Branch_2/Conv2d_0b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_2/Conv2d_0b_3x3/Relu") + .set_name(unit_name + "Branch_2/Conv2d_0b_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(unit_name + "Branch_2/Conv2d_0b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_2/Conv2d_0b_3x3/Relu") << ConvolutionLayer(3U, 3U, 64U, - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) - .set_name(unit_name + "Branch_2/Conv2d_0c_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(unit_name + "Branch_2/Conv2d_0c_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_2/Conv2d_0c_3x3/Relu"); + .set_name(unit_name + "Branch_2/Conv2d_0c_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(unit_name + "Branch_2/Conv2d_0c_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_2/Conv2d_0c_3x3/Relu"); // Concatenate i_l << ConcatLayer(std::move(i_la), std::move(i_lb), std::move(i_lc)).set_name(unit_name + "concat") - << ConvolutionLayer(1U, 1U, 320U, - get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout), - get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Conv2d_1x1/convolution") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 0.17f, 0.f)).set_name(unit_name + "mul"); + << ConvolutionLayer( + 1U, 1U, 320U, + get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout), + PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "Conv2d_1x1/convolution") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 0.17f, 0.f)) + .set_name(unit_name + "mul"); graph << EltwiseLayer(std::move(i_l), std::move(i_r), EltwiseOperation::Add).set_name(unit_name + "add") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu"); + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Relu"); } } void block17_repeat(const std::string &data_path, DataLayout weights_layout, unsigned int num_blocks) { - for(unsigned int i = 0; i < num_blocks; ++i) + for (unsigned int i = 0; i < num_blocks; ++i) { std::stringstream unit_path_ss; unit_path_ss << "Repeat_1_block17_" << (i + 1) << "_"; @@ -614,79 +687,101 @@ private: // Branch 0 SubStream i_la(i_l); - i_la << ConvolutionLayer(1U, 1U, 192U, - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_0/Conv2d_1x1/Relu"); + i_la << ConvolutionLayer( + 1U, 1U, 192U, + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_0/Conv2d_1x1/Relu"); // Branch 1 SubStream i_lb(i_l); i_lb << ConvolutionLayer(1U, 1U, 128U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu") + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu") << ConvolutionLayer(7U, 1U, 160U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) - .set_name(unit_name + "Branch_1/Conv2d_0b_1x7/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(unit_name + "Branch_1/Conv2d_0b_1x7/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0b_1x7/Relu") + .set_name(unit_name + "Branch_1/Conv2d_0b_1x7/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(unit_name + "Branch_1/Conv2d_0b_1x7/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0b_1x7/Relu") << ConvolutionLayer(1U, 7U, 192U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) - .set_name(unit_name + "Branch_1/Conv2d_0c_7x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(unit_name + "Branch_1/Conv2d_0c_7x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0c_7x1/Relu"); + .set_name(unit_name + "Branch_1/Conv2d_0c_7x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(unit_name + "Branch_1/Conv2d_0c_7x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0c_7x1/Relu"); // Concatenate i_l << ConcatLayer(std::move(i_la), std::move(i_lb)).set_name(unit_name + "concat") - << ConvolutionLayer(1U, 1U, 1088U, - get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout), - get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Conv2d_1x1/convolution") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 0.10f, 0.f)).set_name(unit_name + "mul"); + << ConvolutionLayer( + 1U, 1U, 1088U, + get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout), + PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "Conv2d_1x1/convolution") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 0.10f, 0.f)) + .set_name(unit_name + "mul"); graph << EltwiseLayer(std::move(i_l), std::move(i_r), EltwiseOperation::Add).set_name(unit_name + "add") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu"); + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Relu"); } } - void block8_repeat(const std::string &data_path, DataLayout weights_layout, unsigned int num_blocks, float scale, bool has_activation) + void block8_repeat(const std::string &data_path, + DataLayout weights_layout, + unsigned int num_blocks, + float scale, + bool has_activation) { - for(unsigned int i = 0; i < num_blocks; ++i) + for (unsigned int i = 0; i < num_blocks; ++i) { std::stringstream unit_path_ss; std::stringstream unit_name_ss; - if(num_blocks != 1) + if (num_blocks != 1) { unit_path_ss << "Repeat_2_block8_" << (i + 1) << "_"; unit_name_ss << "Repeat_2/block8_" << (i + 1) << "/"; @@ -706,79 +801,97 @@ private: // Branch 0 SubStream i_la(i_l); - i_la << ConvolutionLayer(1U, 1U, 192U, - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_0/Conv2d_1x1/Relu"); + i_la << ConvolutionLayer( + 1U, 1U, 192U, + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_0/Conv2d_1x1/Relu"); // Branch 1 SubStream i_lb(i_l); i_lb << ConvolutionLayer(1U, 1U, 192U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu") + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu") << ConvolutionLayer(3U, 1U, 224U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 0)) - .set_name(unit_name + "Branch_1/Conv2d_0b_1x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(unit_name + "Branch_1/Conv2d_0b_1x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0b_1x3/Relu") + .set_name(unit_name + "Branch_1/Conv2d_0b_1x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(unit_name + "Branch_1/Conv2d_0b_1x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0b_1x3/Relu") << ConvolutionLayer(1U, 3U, 256U, - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_weights.npy", + weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 1)) - .set_name(unit_name + "Branch_1/Conv2d_0c_3x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(unit_name + "Branch_1/Conv2d_0c_3x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0c_3x1/Relu"); + .set_name(unit_name + "Branch_1/Conv2d_0c_3x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(unit_name + "Branch_1/Conv2d_0c_3x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Branch_1/Conv2d_0c_3x1/Relu"); // Concatenate i_l << ConcatLayer(std::move(i_la), std::move(i_lb)).set_name(unit_name + "concat") - << ConvolutionLayer(1U, 1U, 2080U, - get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout), - get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "Conv2d_1x1/convolution"); + << ConvolutionLayer( + 1U, 1U, 2080U, + get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout), + PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "Conv2d_1x1/convolution"); // Scale result - if(scale != 1.f) + if (scale != 1.f) { - i_l << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, scale, 0.f)).set_name(unit_name + "mul"); + i_l << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, scale, 0.f)) + .set_name(unit_name + "mul"); } // Residual add graph << EltwiseLayer(std::move(i_l), std::move(i_r), EltwiseOperation::Add).set_name(unit_name + "add"); // Apply activation if needed - if(has_activation) + if (has_activation) { - graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu"); + graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Relu"); } } } diff --git a/examples/graph_inception_v3.cpp b/examples/graph_inception_v3.cpp index 160e7f04f4..75e03fb6b3 100644 --- a/examples/graph_inception_v3.cpp +++ b/examples/graph_inception_v3.cpp @@ -21,9 +21,10 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/graph.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/Utils.h" +#include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -37,8 +38,7 @@ using namespace arm_compute::graph_utils; class InceptionV3Example : public Example { public: - InceptionV3Example() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "InceptionV3") + InceptionV3Example() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "InceptionV3") { } bool do_setup(int argc, char **argv) override @@ -51,7 +51,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -68,134 +68,163 @@ public: // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(299U, 299U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(299U, 299U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint - << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false)) - << ConvolutionLayer(3U, 3U, 32U, - get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) - .set_name("Conv2d_1a_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - nullptr, get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Conv2d_1a_3x3/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_1a_3x3/Relu") - << ConvolutionLayer(3U, 3U, 32U, - get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name("Conv2d_2a_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Conv2d_2a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Conv2d_2a_3x3_BatchNorm_moving_variance.npy"), - nullptr, get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Conv2d_2a_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Conv2d_2a_3x3/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2a_3x3/Relu") - - << ConvolutionLayer(3U, 3U, 64U, - get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) - .set_name("Conv2d_2b_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Conv2d_2b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Conv2d_2b_3x3_BatchNorm_moving_variance.npy"), - nullptr, get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Conv2d_2b_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Conv2d_2b_3x3/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2b_3x3/Relu") - - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("MaxPool_3a_3x3/MaxPool") - - << ConvolutionLayer(1U, 1U, 80U, - get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_3b_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name("Conv2d_3b_1x1/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Conv2d_3b_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Conv2d_3b_1x1_BatchNorm_moving_variance.npy"), - nullptr, get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Conv2d_3b_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name("Conv2d_3b_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_3b_1x1/Relu") - - << ConvolutionLayer(3U, 3U, 192U, - get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_4a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name("Conv2d_4a_3x3/convolution") - << BatchNormalizationLayer(get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Conv2d_4a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Conv2d_4a_3x3_BatchNorm_moving_variance.npy"), - nullptr, get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Conv2d_4a_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Conv2d_4a_3x3/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_4a_3x3/Relu") - - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("MaxPool_5a_3x3/MaxPool"); - - graph << get_inception_node_A(data_path, "Mixed_5b", weights_layout, 64U, std::make_tuple(48U, 64U), std::make_tuple(64U, 96U, 96U), - 32U) - .set_name("Mixed_5b/concat"); - graph << get_inception_node_A(data_path, "Mixed_5c", weights_layout, 64U, std::make_tuple(48U, 64U), std::make_tuple(64U, 96U, 96U), - 64U, true) - .set_name("Mixed_5c/concat"); - graph << get_inception_node_A(data_path, "Mixed_5d", weights_layout, 64U, std::make_tuple(48U, 64U), std::make_tuple(64U, 96U, 96U), - 64U) - .set_name("Mixed_5d/concat"); - - graph << get_inception_node_B(data_path, "Mixed_6a", weights_layout, 384U, std::make_tuple(64U, 96U, 96U)).set_name("Mixed_6a/concat"); + graph + << common_params.target << common_params.fast_math_hint + << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false)) + << ConvolutionLayer(3U, 3U, 32U, + get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_1a_3x3_weights.npy", + weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(2, 2, 0, 0)) + .set_name("Conv2d_1a_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, + "/cnn_data/inceptionv3_model/Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + "/cnn_data/inceptionv3_model/Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + nullptr, + get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_1a_3x3_BatchNorm_beta.npy"), + 0.001f) + .set_name("Conv2d_1a_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_1a_3x3/Relu") + << ConvolutionLayer(3U, 3U, 32U, + get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2a_3x3_weights.npy", + weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(1, 1, 0, 0)) + .set_name("Conv2d_2a_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, + "/cnn_data/inceptionv3_model/Conv2d_2a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + "/cnn_data/inceptionv3_model/Conv2d_2a_3x3_BatchNorm_moving_variance.npy"), + nullptr, + get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2a_3x3_BatchNorm_beta.npy"), + 0.001f) + .set_name("Conv2d_2a_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_2a_3x3/Relu") + + << ConvolutionLayer(3U, 3U, 64U, + get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2b_3x3_weights.npy", + weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(1, 1, 1, 1)) + .set_name("Conv2d_2b_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, + "/cnn_data/inceptionv3_model/Conv2d_2b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + "/cnn_data/inceptionv3_model/Conv2d_2b_3x3_BatchNorm_moving_variance.npy"), + nullptr, + get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2b_3x3_BatchNorm_beta.npy"), + 0.001f) + .set_name("Conv2d_2b_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_2b_3x3/Relu") + + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("MaxPool_3a_3x3/MaxPool") + + << ConvolutionLayer(1U, 1U, 80U, + get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_3b_1x1_weights.npy", + weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(1, 1, 0, 0)) + .set_name("Conv2d_3b_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, + "/cnn_data/inceptionv3_model/Conv2d_3b_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + "/cnn_data/inceptionv3_model/Conv2d_3b_1x1_BatchNorm_moving_variance.npy"), + nullptr, + get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_3b_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name("Conv2d_3b_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_3b_1x1/Relu") + + << ConvolutionLayer(3U, 3U, 192U, + get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_4a_3x3_weights.npy", + weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(1, 1, 0, 0)) + .set_name("Conv2d_4a_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, + "/cnn_data/inceptionv3_model/Conv2d_4a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + "/cnn_data/inceptionv3_model/Conv2d_4a_3x3_BatchNorm_moving_variance.npy"), + nullptr, + get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_4a_3x3_BatchNorm_beta.npy"), + 0.001f) + .set_name("Conv2d_4a_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_4a_3x3/Relu") + + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("MaxPool_5a_3x3/MaxPool"); + + graph << get_inception_node_A(data_path, "Mixed_5b", weights_layout, 64U, std::make_tuple(48U, 64U), + std::make_tuple(64U, 96U, 96U), 32U) + .set_name("Mixed_5b/concat"); + graph << get_inception_node_A(data_path, "Mixed_5c", weights_layout, 64U, std::make_tuple(48U, 64U), + std::make_tuple(64U, 96U, 96U), 64U, true) + .set_name("Mixed_5c/concat"); + graph << get_inception_node_A(data_path, "Mixed_5d", weights_layout, 64U, std::make_tuple(48U, 64U), + std::make_tuple(64U, 96U, 96U), 64U) + .set_name("Mixed_5d/concat"); + + graph << get_inception_node_B(data_path, "Mixed_6a", weights_layout, 384U, std::make_tuple(64U, 96U, 96U)) + .set_name("Mixed_6a/concat"); graph << get_inception_node_C(data_path, "Mixed_6b", weights_layout, 192U, std::make_tuple(128U, 128U, 192U), std::make_tuple(128U, 128U, 128U, 128U, 192U), 192U) - .set_name("Mixed_6b/concat"); + .set_name("Mixed_6b/concat"); graph << get_inception_node_C(data_path, "Mixed_6c", weights_layout, 192U, std::make_tuple(160U, 160U, 192U), std::make_tuple(160U, 160U, 160U, 160U, 192U), 192U) - .set_name("Mixed_6c/concat"); + .set_name("Mixed_6c/concat"); graph << get_inception_node_C(data_path, "Mixed_6d", weights_layout, 192U, std::make_tuple(160U, 160U, 192U), std::make_tuple(160U, 160U, 160U, 160U, 192U), 192U) - .set_name("Mixed_6d/concat"); + .set_name("Mixed_6d/concat"); graph << get_inception_node_C(data_path, "Mixed_6e", weights_layout, 192U, std::make_tuple(192U, 192U, 192U), std::make_tuple(192U, 192U, 192U, 192U, 192U), 192U) - .set_name("Mixed_6e/concat"); + .set_name("Mixed_6e/concat"); graph << get_inception_node_D(data_path, "Mixed_7a", weights_layout, std::make_tuple(192U, 320U), std::make_tuple(192U, 192U, 192U, 192U)) - .set_name("Mixed_7a/concat"); + .set_name("Mixed_7a/concat"); graph << get_inception_node_E(data_path, "Mixed_7b", weights_layout, 320U, std::make_tuple(384U, 384U, 384U), std::make_tuple(448U, 384U, 384U, 384U), 192U) - .set_name("Mixed_7b/concat"); + .set_name("Mixed_7b/concat"); graph << get_inception_node_E(data_path, "Mixed_7c", weights_layout, 320U, std::make_tuple(384U, 384U, 384U), std::make_tuple(448U, 384U, 384U, 384U), 192U, true) - .set_name("Mixed_7c/concat"); - - graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 8, operation_layout, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL))).set_name("Logits/AvgPool_1a_8x8/AvgPool") - << ConvolutionLayer(1U, 1U, 1001U, get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Logits_Conv2d_1c_1x1_weights.npy", weights_layout), - get_weights_accessor(data_path, - "/cnn_data/inceptionv3_model/Logits_Conv2d_1c_1x1_biases.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Logits/Conv2d_1c_1x1/convolution") + .set_name("Mixed_7c/concat"); + + graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 8, operation_layout, + PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL))) + .set_name("Logits/AvgPool_1a_8x8/AvgPool") + << ConvolutionLayer( + 1U, 1U, 1001U, + get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Logits_Conv2d_1c_1x1_weights.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Logits_Conv2d_1c_1x1_biases.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("Logits/Conv2d_1c_1x1/convolution") << ReshapeLayer(TensorShape(1001U)).set_name("Predictions/Reshape") - << SoftmaxLayer().set_name("Predictions/Softmax") - << OutputLayer(get_output_accessor(common_params, 5)); + << SoftmaxLayer().set_name("Predictions/Softmax") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph GraphConfig config; @@ -223,19 +252,21 @@ private: Stream graph; private: - ConcatLayer get_inception_node_A(const std::string &data_path, std::string &¶m_path, DataLayout weights_layout, - unsigned int a_filt, - std::tuple<unsigned int, unsigned int> b_filters, + ConcatLayer get_inception_node_A(const std::string &data_path, + std::string &¶m_path, + DataLayout weights_layout, + unsigned int a_filt, + std::tuple<unsigned int, unsigned int> b_filters, std::tuple<unsigned int, unsigned int, unsigned int> c_filters, - unsigned int d_filt, - bool is_name_different = false) + unsigned int d_filt, + bool is_name_different = false) { std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_"; // This is due to a naming issue in the tf model std::string conv_id0 = "_0a_"; std::string conv_id1 = "2d_0b_"; - if(is_name_different) + if (is_name_different) { conv_id0 = "_0b_"; conv_id1 = "_1_0c_"; @@ -243,457 +274,451 @@ private: SubStream i_a(graph); i_a << ConvolutionLayer( - 1U, 1U, a_filt, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu"); + 1U, 1U, a_filt, + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu"); SubStream i_b(graph); i_b << ConvolutionLayer( - 1U, 1U, std::get<0>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + "1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_1/Conv2d" + conv_id0 + "1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + "1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + "1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + "1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d" + conv_id0 + "1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d" + conv_id0 + "1x1/Relu") + 1U, 1U, std::get<0>(b_filters), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + "1x1_weights.npy", + weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d" + conv_id0 + "1x1/convolution") + << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + + "1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + + "1x1_BatchNorm_moving_variance.npy"), + nullptr, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + + "1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d" + conv_id0 + "1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d" + conv_id0 + "1x1/Relu") << ConvolutionLayer( - 5U, 5U, std::get<1>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 2, 2)) - .set_name(param_path + "/Branch_1/Conv2d" + conv_id1 + "5x5/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d" + conv_id1 + "5x5/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d" + conv_id1 + "5x5/Relu"); + 5U, 5U, std::get<1>(b_filters), + get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_weights.npy", + weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 2, 2)) + .set_name(param_path + "/Branch_1/Conv2d" + conv_id1 + "5x5/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, + total_path + "Branch_1_Conv" + conv_id1 + "5x5_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + total_path + "Branch_1_Conv" + conv_id1 + "5x5_BatchNorm_moving_variance.npy"), + nullptr, + get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d" + conv_id1 + "5x5/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d" + conv_id1 + "5x5/Relu"); SubStream i_c(graph); i_c << ConvolutionLayer( - 1U, 1U, std::get<0>(c_filters), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu") + 1U, 1U, std::get<0>(c_filters), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu") << ConvolutionLayer( - 3U, 3U, std::get<1>(c_filters), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0b_3x3/Relu") + 3U, 3U, std::get<1>(c_filters), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/Relu") << ConvolutionLayer( - 3U, 3U, std::get<2>(c_filters), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name(param_path + "/Branch_2/Conv2d_0c_3x3/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0c_3x3/BatchNorm/batcnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0c_3x3/Relu"); + 3U, 3U, std::get<2>(c_filters), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name(param_path + "/Branch_2/Conv2d_0c_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0c_3x3/BatchNorm/batcnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0c_3x3/Relu"); SubStream i_d(graph); - i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), - true)) - .set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool") + i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, + PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) + .set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool") << ConvolutionLayer( - 1U, 1U, d_filt, - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu"); + 1U, 1U, d_filt, + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu"); return ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)); } - ConcatLayer get_inception_node_B(const std::string &data_path, std::string &¶m_path, DataLayout weights_layout, - unsigned int a_filt, + ConcatLayer get_inception_node_B(const std::string &data_path, + std::string &¶m_path, + DataLayout weights_layout, + unsigned int a_filt, std::tuple<unsigned int, unsigned int, unsigned int> b_filters) { std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_"; SubStream i_a(graph); i_a << ConvolutionLayer( - 3U, 3U, a_filt, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name(param_path + "/Branch_0/Conv2d_1a_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_0/Conv2d_1a_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_1a_1x1/Relu"); + 3U, 3U, a_filt, + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name(param_path + "/Branch_0/Conv2d_1a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_0/Conv2d_1a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_0/Conv2d_1a_1x1/Relu"); SubStream i_b(graph); i_b << ConvolutionLayer( - 1U, 1U, std::get<0>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu") + 1U, 1U, std::get<0>(b_filters), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu") << ConvolutionLayer( - 3U, 3U, std::get<1>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name(param_path + "/Branch_1/Conv2d_0b_3x3/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0b_3x3/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0b_3x3/Relu") + 3U, 3U, std::get<1>(b_filters), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name(param_path + "/Branch_1/Conv2d_0b_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0b_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0b_3x3/Relu") << ConvolutionLayer( - 3U, 3U, std::get<2>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name(param_path + "/Branch_1/Conv2d_1a_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_1a_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_1a_1x1/Relu"); + 3U, 3U, std::get<2>(b_filters), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d_1a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_1a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_1a_1x1/Relu"); SubStream i_c(graph); - i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name(param_path + "/Branch_2/MaxPool_1a_3x3/MaxPool"); + i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name(param_path + "/Branch_2/MaxPool_1a_3x3/MaxPool"); return ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c)); } - ConcatLayer get_inception_node_C(const std::string &data_path, std::string &¶m_path, DataLayout weights_layout, - unsigned int a_filt, - std::tuple<unsigned int, unsigned int, unsigned int> b_filters, - std::tuple<unsigned int, unsigned int, unsigned int, unsigned int, unsigned int> c_filters, - unsigned int d_filt) + ConcatLayer + get_inception_node_C(const std::string &data_path, + std::string &¶m_path, + DataLayout weights_layout, + unsigned int a_filt, + std::tuple<unsigned int, unsigned int, unsigned int> b_filters, + std::tuple<unsigned int, unsigned int, unsigned int, unsigned int, unsigned int> c_filters, + unsigned int d_filt) { std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_"; SubStream i_a(graph); i_a << ConvolutionLayer( - 1U, 1U, a_filt, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu"); + 1U, 1U, a_filt, + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu"); SubStream i_b(graph); i_b << ConvolutionLayer( - 1U, 1U, std::get<0>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu") + 1U, 1U, std::get<0>(b_filters), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu") << ConvolutionLayer( - 7U, 1U, std::get<1>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 3, 0)) - .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0b_1x7/Relu") + 7U, 1U, std::get<1>(b_filters), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/Relu") << ConvolutionLayer( - 1U, 7U, std::get<2>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 3)) - .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0c_7x1/Relu"); + 1U, 7U, std::get<2>(b_filters), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) + .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_0/Conv2d_0c_7x1/Relu"); SubStream i_c(graph); i_c << ConvolutionLayer( - 1U, 1U, std::get<0>(c_filters), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu") + 1U, 1U, std::get<0>(c_filters), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu") << ConvolutionLayer( - 1U, 7U, std::get<1>(c_filters), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 3)) - .set_name(param_path + "/Branch_2/Conv2d_0b_7x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0b_7x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0b_7x1/Relu") + 1U, 7U, std::get<1>(c_filters), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) + .set_name(param_path + "/Branch_2/Conv2d_0b_7x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0b_7x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0b_7x1/Relu") << ConvolutionLayer( - 7U, 1U, std::get<2>(c_filters), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 3, 0)) - .set_name(param_path + "/Branch_2/Conv2d_0c_1x7/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0c_1x7/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0c_1x7/Relu") + 7U, 1U, std::get<2>(c_filters), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0c_1x7/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0c_1x7/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0c_1x7/Relu") << ConvolutionLayer( - 1U, 7U, std::get<3>(c_filters), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 3)) - .set_name(param_path + "/Branch_2/Conv2d_0d_7x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0d_7x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0d_7x1/Relu") + 1U, 7U, std::get<3>(c_filters), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) + .set_name(param_path + "/Branch_2/Conv2d_0d_7x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0d_7x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0d_7x1/Relu") << ConvolutionLayer( - 7U, 1U, std::get<4>(c_filters), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 3, 0)) - .set_name(param_path + "/Branch_2/Conv2d_0e_1x7/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0e_1x7/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0e_1x7/Relu"); + 7U, 1U, std::get<4>(c_filters), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0e_1x7/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0e_1x7/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0e_1x7/Relu"); SubStream i_d(graph); - i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), - true)) - .set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool") + i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, + PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) + .set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool") << ConvolutionLayer( - 1U, 1U, d_filt, - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu"); + 1U, 1U, d_filt, + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu"); return ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)); } - ConcatLayer get_inception_node_D(const std::string &data_path, std::string &¶m_path, DataLayout weights_layout, - std::tuple<unsigned int, unsigned int> a_filters, + ConcatLayer get_inception_node_D(const std::string &data_path, + std::string &¶m_path, + DataLayout weights_layout, + std::tuple<unsigned int, unsigned int> a_filters, std::tuple<unsigned int, unsigned int, unsigned int, unsigned int> b_filters) { std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_"; SubStream i_a(graph); i_a << ConvolutionLayer( - 1U, 1U, std::get<0>(a_filters), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu") + 1U, 1U, std::get<0>(a_filters), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu") << ConvolutionLayer( - 3U, 3U, std::get<1>(a_filters), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name(param_path + "/Branch_0/Conv2d_1a_3x3/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_0/Conv2d_1a_3x3/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_1a_3x3/Relu"); + 3U, 3U, std::get<1>(a_filters), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name(param_path + "/Branch_0/Conv2d_1a_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_0/Conv2d_1a_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_0/Conv2d_1a_3x3/Relu"); SubStream i_b(graph); i_b << ConvolutionLayer( - 1U, 1U, std::get<0>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu") + 1U, 1U, std::get<0>(b_filters), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu") << ConvolutionLayer( - 7U, 1U, std::get<1>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 3, 0)) - .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0b_1x7/Relu") + 7U, 1U, std::get<1>(b_filters), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/Relu") << ConvolutionLayer( - 1U, 7U, std::get<2>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 3)) - .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0c_7x1/Relu") + 1U, 7U, std::get<2>(b_filters), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) + .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/Relu") << ConvolutionLayer( - 3U, 3U, std::get<3>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 0)) - .set_name(param_path + "/Branch_1/Conv2d_1a_3x3/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_1a_3x3/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_1a_3x3/Relu"); + 3U, 3U, std::get<3>(b_filters), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d_1a_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_1a_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_1a_3x3/Relu"); SubStream i_c(graph); - i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name(param_path + "/Branch_2/MaxPool_1a_3x3/MaxPool"); + i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name(param_path + "/Branch_2/MaxPool_1a_3x3/MaxPool"); return ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c)); } - ConcatLayer get_inception_node_E(const std::string &data_path, std::string &¶m_path, DataLayout weights_layout, - unsigned int a_filt, - std::tuple<unsigned int, unsigned int, unsigned int> b_filters, + ConcatLayer get_inception_node_E(const std::string &data_path, + std::string &¶m_path, + DataLayout weights_layout, + unsigned int a_filt, + std::tuple<unsigned int, unsigned int, unsigned int> b_filters, std::tuple<unsigned int, unsigned int, unsigned int, unsigned int> c_filters, - unsigned int d_filt, - bool is_name_different = false) + unsigned int d_filt, + bool is_name_different = false) { // This is due to a naming issue in the tf model std::string conv_id = "_0b_"; - if(is_name_different) + if (is_name_different) { conv_id = "_0c_"; } @@ -701,154 +726,152 @@ private: std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_"; SubStream i_a(graph); i_a << ConvolutionLayer( - 1U, 1U, a_filt, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu"); + 1U, 1U, a_filt, + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu"); SubStream i_b(graph); i_b << ConvolutionLayer( - 1U, 1U, std::get<0>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu"); + 1U, 1U, std::get<0>(b_filters), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu"); SubStream i_b1(i_b); i_b1 << ConvolutionLayer( - 3U, 1U, std::get<1>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 0)) - .set_name(param_path + "/Branch_1/Conv2d_0b_1x3/convolution") + 3U, 1U, std::get<1>(b_filters), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x3/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0b_1x3/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0b_1x3/Relu"); + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x3/Relu"); SubStream i_b2(i_b); i_b2 << ConvolutionLayer( - 1U, 3U, std::get<2>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 1)) - .set_name(param_path + "/Branch_1/Conv2d" + conv_id + "3x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d" + conv_id + "3x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d" + conv_id + "3x1/Relu"); + 1U, 3U, std::get<2>(b_filters), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_weights.npy", + weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 1)) + .set_name(param_path + "/Branch_1/Conv2d" + conv_id + "3x1/convolution") + << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + + "3x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + + "3x1_BatchNorm_moving_variance.npy"), + nullptr, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + + "3x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d" + conv_id + "3x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d" + conv_id + "3x1/Relu"); // Merge b1 and b2 i_b << ConcatLayer(std::move(i_b1), std::move(i_b2)).set_name(param_path + "/Branch_1/concat"); SubStream i_c(graph); i_c << ConvolutionLayer( - 1U, 1U, std::get<0>(c_filters), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu") + 1U, 1U, std::get<0>(c_filters), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu") << ConvolutionLayer( - 3U, 3U, std::get<1>(c_filters), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0b_3x3/Relu"); + 3U, 3U, std::get<1>(c_filters), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/Relu"); SubStream i_c1(i_c); i_c1 << ConvolutionLayer( - 3U, 1U, std::get<2>(c_filters), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 0)) - .set_name(param_path + "/Branch_2/Conv2d_0c_1x3/convolution") + 3U, 1U, std::get<2>(c_filters), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0c_1x3/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0c_1x3/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0c_1x3/Relu"); + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0c_1x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0c_1x3/Relu"); SubStream i_c2(i_c); i_c2 << ConvolutionLayer( - 1U, 3U, std::get<3>(c_filters), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 1)) - .set_name(param_path + "/Branch_2/Conv2d_0d_3x1/convolution") + 1U, 3U, std::get<3>(c_filters), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 1)) + .set_name(param_path + "/Branch_2/Conv2d_0d_3x1/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0d_3x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0d_3x1/Relu"); + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0d_3x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0d_3x1/Relu"); // Merge i_c1 and i_c2 i_c << ConcatLayer(std::move(i_c1), std::move(i_c2)).set_name(param_path + "/Branch_2/concat"); SubStream i_d(graph); - i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), - true)) - .set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool") + i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, + PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) + .set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool") << ConvolutionLayer( - 1U, 1U, d_filt, - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), - nullptr, - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm/batchnorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu"); + 1U, 1U, d_filt, + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), + nullptr, get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu"); return ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)); } diff --git a/examples/graph_inception_v4.cpp b/examples/graph_inception_v4.cpp index 6d8fab4141..052498ad38 100644 --- a/examples/graph_inception_v4.cpp +++ b/examples/graph_inception_v4.cpp @@ -39,8 +39,7 @@ using namespace arm_compute::graph_utils; class InceptionV4Example final : public Example { public: - InceptionV4Example() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "InceptionV4") + InceptionV4Example() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "InceptionV4") { } bool do_setup(int argc, char **argv) override @@ -53,7 +52,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -70,51 +69,70 @@ public: // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(299U, 299U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(299U, 299U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint + graph << common_params.target << common_params.fast_math_hint << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false)) // Conv2d_1a_3x3 - << ConvolutionLayer(3U, 3U, 32U, - get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) - .set_name("Conv2d_1a_3x3/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_1a_3x3/Relu") + << ConvolutionLayer( + 3U, 3U, 32U, + get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_weights.npy", + weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Conv2d_1a_3x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, + "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_beta.npy"), + 0.001f) + .set_name("Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_1a_3x3/Relu") // Conv2d_2a_3x3 - << ConvolutionLayer(3U, 3U, 32U, - get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name("Conv2d_2a_3x3/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Conv2d_2a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2a_3x3/Relu") + << ConvolutionLayer( + 3U, 3U, 32U, + get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_weights.npy", + weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Conv2d_2a_3x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, + "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_beta.npy"), + 0.001f) + .set_name("Conv2d_2a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_2a_3x3/Relu") // Conv2d_2b_3x3 - << ConvolutionLayer(3U, 3U, 64U, - get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) - .set_name("Conv2d_2b_3x3/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Conv2d_2b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2b_3x3/Relu"); + << ConvolutionLayer( + 3U, 3U, 64U, + get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_weights.npy", + weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("Conv2d_2b_3x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, + "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_beta.npy"), + 0.001f) + .set_name("Conv2d_2b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_2b_3x3/Relu"); graph << get_mixed_3a(data_path, weights_layout).set_name("Mixed_3a/concat"); graph << get_mixed_4a(data_path, weights_layout).set_name("Mixed_4a/concat"); @@ -140,15 +158,16 @@ public: graph << get_inceptionC_block(data_path, weights_layout, "Mixed_7b").set_name("Mixed_7b/concat"); graph << get_inceptionC_block(data_path, weights_layout, "Mixed_7c").set_name("Mixed_7c/concat"); graph << get_inceptionC_block(data_path, weights_layout, "Mixed_7d").set_name("Mixed_7d/concat"); - graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("Logits/AvgPool_1a/AvgPool") - << FlattenLayer().set_name("Logits/Flatten") - << FullyConnectedLayer( - 1001U, - get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_weights.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_biases.npy")) - .set_name("Logits/MatMul") - << SoftmaxLayer().set_name("Logits/Predictions") - << OutputLayer(get_output_accessor(common_params, 5)); + graph + << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("Logits/AvgPool_1a/AvgPool") + << FlattenLayer().set_name("Logits/Flatten") + << FullyConnectedLayer( + 1001U, + get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_weights.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_biases.npy")) + .set_name("Logits/MatMul") + << SoftmaxLayer().set_name("Logits/Predictions") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph GraphConfig config; @@ -162,7 +181,7 @@ public: // Load the precompiled kernels from a file into the kernel library, in this way the next time they are needed // compilation won't be required. - if(common_params.enable_cl_cache) + if (common_params.enable_cl_cache) { #ifdef ARM_COMPUTE_CL restore_program_cache_from_file(); @@ -172,7 +191,7 @@ public: graph.finalize(common_params.target, config); // Save the opencl kernels to a file - if(common_opts.enable_cl_cache) + if (common_opts.enable_cl_cache) { #ifdef ARM_COMPUTE_CL save_program_cache_to_file(); @@ -199,22 +218,24 @@ private: std::string total_path = "/cnn_data/inceptionv4_model/Mixed_3a_"; SubStream i_a(graph); - i_a << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), - true)) - .set_name("Mixed_3a/Branch_0/MaxPool_0a_3x3/MaxPool"); + i_a << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)) + .set_name("Mixed_3a/Branch_0/MaxPool_0a_3x3/MaxPool"); SubStream i_b(graph); - i_b << ConvolutionLayer(3U, 3U, 96U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_3a/Branch_1/Conv2d_0a_3x3/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_3a/Branch_1/Conv2d_0a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_3a/Branch_1/Conv2d_0a_3x3/Relu"); + i_b << ConvolutionLayer( + 3U, 3U, 96U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_3a/Branch_1/Conv2d_0a_3x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_3a/Branch_1/Conv2d_0a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_3a/Branch_1/Conv2d_0a_3x3/Relu"); return ConcatLayer(std::move(i_a), std::move(i_b)); } @@ -224,74 +245,86 @@ private: std::string total_path = "/cnn_data/inceptionv4_model/Mixed_4a_"; SubStream i_a(graph); - i_a << ConvolutionLayer(1U, 1U, 64U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_4a/Branch_0/Conv2d_0a_1x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_4a/Branch_0/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_4a/Branch_0/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(3U, 3U, 96U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_4a/Branch_0/Conv2d_1a_3x3/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_4a/Branch_0/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_4a/Branch_0/Conv2d_1a_3x3/Relu"); + i_a << ConvolutionLayer( + 1U, 1U, 64U, + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_4a/Branch_0/Conv2d_0a_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_4a/Branch_0/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_4a/Branch_0/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 3U, 3U, 96U, + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_4a/Branch_0/Conv2d_1a_3x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_4a/Branch_0/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_4a/Branch_0/Conv2d_1a_3x3/Relu"); SubStream i_b(graph); - i_b << ConvolutionLayer(1U, 1U, 64U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_4a/Branch_1/Conv2d_0a_1x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_4a/Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_4a/Branch_1/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(7U, 1U, 64U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) - .set_name("Mixed_4a/Branch_1/Conv2d_0b_1x7/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_4a/Branch_1/Conv2d_0b_1x7/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_4a/Branch_1/Conv2d_0b_1x7/Relu") - << ConvolutionLayer(1U, 7U, 64U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) - .set_name("Mixed_4a/Branch_1/Conv2d_0c_7x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_4a/Branch_1/Conv2d_0c_7x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_4a/Branch_1/Conv2d_0c_7x1/Relu") - << ConvolutionLayer(3U, 3U, 96U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_4a/Branch_1/Conv2d_1a_3x3/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_4a/Branch_1/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_4a/Branch_1/Conv2d_1a_3x3/Relu"); + i_b << ConvolutionLayer( + 1U, 1U, 64U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_4a/Branch_1/Conv2d_0a_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_4a/Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_4a/Branch_1/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 7U, 1U, 64U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) + .set_name("Mixed_4a/Branch_1/Conv2d_0b_1x7/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_4a/Branch_1/Conv2d_0b_1x7/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_4a/Branch_1/Conv2d_0b_1x7/Relu") + << ConvolutionLayer( + 1U, 7U, 64U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) + .set_name("Mixed_4a/Branch_1/Conv2d_0c_7x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_4a/Branch_1/Conv2d_0c_7x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_4a/Branch_1/Conv2d_0c_7x1/Relu") + << ConvolutionLayer( + 3U, 3U, 96U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_4a/Branch_1/Conv2d_1a_3x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_4a/Branch_1/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_4a/Branch_1/Conv2d_1a_3x3/Relu"); return ConcatLayer(std::move(i_a), std::move(i_b)); } @@ -301,22 +334,24 @@ private: std::string total_path = "/cnn_data/inceptionv4_model/Mixed_5a_"; SubStream i_a(graph); - i_a << ConvolutionLayer(3U, 3U, 192U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_5a/Branch_0/Conv2d_1a_3x3/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_5a/Branch_0/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5a/Branch_0/Conv2d_1a_3x3/Relu"); + i_a << ConvolutionLayer( + 3U, 3U, 192U, + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_5a/Branch_0/Conv2d_1a_3x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_5a/Branch_0/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_5a/Branch_0/Conv2d_1a_3x3/Relu"); SubStream i_b(graph); - i_b << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), - true)) - .set_name("Mixed_5a/Branch_1/MaxPool_1a_3x3/MaxPool"); + i_b << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)) + .set_name("Mixed_5a/Branch_1/MaxPool_1a_3x3/MaxPool"); return ConcatLayer(std::move(i_a), std::move(i_b)); } @@ -326,92 +361,106 @@ private: std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_"; SubStream i_a(graph); - i_a << ConvolutionLayer(1U, 1U, 96U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu"); + i_a << ConvolutionLayer( + 1U, 1U, 96U, + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu"); SubStream i_b(graph); - i_b << ConvolutionLayer(1U, 1U, 64U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(3U, 3U, 96U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) - .set_name(param_path + "/Branch_1/Conv2d_0b_3x3/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0b_3x3/Relu"); + i_b << ConvolutionLayer( + 1U, 1U, 64U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 3U, 3U, 96U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name(param_path + "/Branch_1/Conv2d_0b_3x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0b_3x3/Relu"); SubStream i_c(graph); - i_c << ConvolutionLayer(1U, 1U, 64U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(3U, 3U, 96U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) - .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0b_3x3/Relu") - << ConvolutionLayer(3U, 3U, 96U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) - .set_name(param_path + "/Branch_2/Conv2d_0c_3x3/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0c_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0c_3x3/Relu"); + i_c << ConvolutionLayer( + 1U, 1U, 64U, + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 3U, 3U, 96U, + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/Relu") + << ConvolutionLayer( + 3U, 3U, 96U, + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name(param_path + "/Branch_2/Conv2d_0c_3x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0c_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0c_3x3/Relu"); SubStream i_d(graph); - i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), - true)) - .set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool") - << ConvolutionLayer(1U, 1U, 96U, - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu"); + i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, + PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) + .set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool") + << ConvolutionLayer( + 1U, 1U, 96U, + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu"); return ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)); } @@ -421,57 +470,65 @@ private: std::string total_path = "/cnn_data/inceptionv4_model/Mixed_6a_"; SubStream i_a(graph); - i_a << ConvolutionLayer(3U, 3U, 384U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/Relu"); + i_a << ConvolutionLayer( + 3U, 3U, 384U, + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/Relu"); SubStream i_b(graph); - i_b << ConvolutionLayer(1U, 1U, 192U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(3U, 3U, 224U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) - .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/Relu") - << ConvolutionLayer(3U, 3U, 256U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/Relu"); + i_b << ConvolutionLayer( + 1U, 1U, 192U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 3U, 3U, 224U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/Relu") + << ConvolutionLayer( + 3U, 3U, 256U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/Relu"); SubStream i_c(graph); - i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), - true)) - .set_name("Mixed_6a/Branch_2/MaxPool_1a_3x3/MaxPool"); + i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)) + .set_name("Mixed_6a/Branch_2/MaxPool_1a_3x3/MaxPool"); return ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c)); } @@ -481,125 +538,145 @@ private: std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_"; SubStream i_a(graph); - i_a << ConvolutionLayer(1U, 1U, 384U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu"); + i_a << ConvolutionLayer( + 1U, 1U, 384U, + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu"); SubStream i_b(graph); - i_b << ConvolutionLayer(1U, 1U, 192U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(7U, 1U, 224U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) - .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0b_1x7/Relu") - << ConvolutionLayer(1U, 7U, 256U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) - .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0c_7x1/Relu"); + i_b << ConvolutionLayer( + 1U, 1U, 192U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 7U, 1U, 224U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/Relu") + << ConvolutionLayer( + 1U, 7U, 256U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) + .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/Relu"); SubStream i_c(graph); - i_c << ConvolutionLayer(1U, 1U, 192U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(1U, 7U, 192U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) - .set_name(param_path + "/Branch_2/Conv2d_0b_7x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0b_7x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0b_7x1/Relu") - << ConvolutionLayer(7U, 1U, 224U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) - .set_name(param_path + "/Branch_2/Conv2d_0c_1x7/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0c_1x7/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0c_1x7/Relu") - << ConvolutionLayer(1U, 7U, 224U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) - .set_name(param_path + "/Branch_2/Conv2d_0d_7x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0d_7x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0d_7x1/Relu") - << ConvolutionLayer(7U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) - .set_name(param_path + "/Branch_2/Conv2d_0e_1x7/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0e_1x7/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0e_1x7/Relu"); + i_c << ConvolutionLayer( + 1U, 1U, 192U, + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 1U, 7U, 192U, + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) + .set_name(param_path + "/Branch_2/Conv2d_0b_7x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0b_7x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0b_7x1/Relu") + << ConvolutionLayer( + 7U, 1U, 224U, + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0c_1x7/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0c_1x7/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0c_1x7/Relu") + << ConvolutionLayer( + 1U, 7U, 224U, + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) + .set_name(param_path + "/Branch_2/Conv2d_0d_7x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0d_7x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0d_7x1/Relu") + << ConvolutionLayer( + 7U, 1U, 256U, + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0e_1x7/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0e_1x7/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0e_1x7/Relu"); SubStream i_d(graph); - i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), - true)) - .set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool") - << ConvolutionLayer(1U, 1U, 128U, - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu"); + i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, + PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) + .set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool") + << ConvolutionLayer( + 1U, 1U, 128U, + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu"); return ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)); } @@ -609,79 +686,91 @@ private: std::string total_path = "/cnn_data/inceptionv4_model/Mixed_7a_"; SubStream i_a(graph); - i_a << ConvolutionLayer(1U, 1U, 192U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(3U, 3U, 192U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/Relu"); + i_a << ConvolutionLayer( + 1U, 1U, 192U, + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 3U, 3U, 192U, + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/Relu"); SubStream i_b(graph); - i_b << ConvolutionLayer(1U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/Relu") - << ConvolutionLayer(7U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) - .set_name("Mixed_7a/Branch_1/Conv2d_0b_1x7/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_7a/Branch_1/Conv2d_0b_1x7/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_1/Conv2d_0b_1x7/Relu") - << ConvolutionLayer(1U, 7U, 320U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) - .set_name("Mixed_7a/Branch_1/Conv2d_0c_7x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_7a/Branch_1/Conv2d_0c_7x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_1/Conv2d_0c_7x1/Relu") - << ConvolutionLayer(3U, 3U, 320U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) - .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.001f) - .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/Relu"); + i_b << ConvolutionLayer( + 1U, 1U, 256U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 7U, 1U, 256U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) + .set_name("Mixed_7a/Branch_1/Conv2d_0b_1x7/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_7a/Branch_1/Conv2d_0b_1x7/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_1/Conv2d_0b_1x7/Relu") + << ConvolutionLayer( + 1U, 7U, 320U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) + .set_name("Mixed_7a/Branch_1/Conv2d_0c_7x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_7a/Branch_1/Conv2d_0c_7x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_1/Conv2d_0c_7x1/Relu") + << ConvolutionLayer( + 3U, 3U, 320U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), 0.001f) + .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/Relu"); SubStream i_c(graph); - i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), - true)) - .set_name("Mixed_7a/Branch_2/MaxPool_1a_3x3/MaxPool"); + i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)) + .set_name("Mixed_7a/Branch_2/MaxPool_1a_3x3/MaxPool"); return ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c)); } @@ -691,163 +780,163 @@ private: std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_"; SubStream i_a(graph); - i_a << ConvolutionLayer(1U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu"); + i_a << ConvolutionLayer( + 1U, 1U, 256U, + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu"); SubStream i_b(graph); i_b << ConvolutionLayer( - 1U, 1U, 384U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Conv2D") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu"); + 1U, 1U, 384U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu"); SubStream i_b1(i_b); i_b1 << ConvolutionLayer( - 3U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 0)) - .set_name(param_path + "/Branch_1/Conv2d_0b_1x3/Conv2D") + 3U, 1U, 256U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x3/Conv2D") << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0b_1x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0b_1x3/Relu"); + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x3/Relu"); SubStream i_b2(i_b); i_b2 << ConvolutionLayer( - 1U, 3U, 256U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 1)) - .set_name(param_path + "/Branch_1/Conv2d_0c_3x1/Conv2D") + 1U, 3U, 256U, + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 1)) + .set_name(param_path + "/Branch_1/Conv2d_0c_3x1/Conv2D") << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_1/Conv2d_0c_3x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0c_3x1/Relu"); + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0c_3x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0c_3x1/Relu"); // Merge b1 and b2 i_b << ConcatLayer(std::move(i_b1), std::move(i_b2)).set_name(param_path + "/Branch_1/concat"); SubStream i_c(graph); i_c << ConvolutionLayer( - 1U, 1U, 384U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Conv2D") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu") + 1U, 1U, 384U, + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu") << ConvolutionLayer( - 1U, 3U, 448U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 1)) - .set_name(param_path + "/Branch_2/Conv2d_0b_3x1/Conv2D") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0b_3x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0b_3x1/Relu") + 1U, 3U, 448U, + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 1)) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x1/Relu") << ConvolutionLayer( - 3U, 1U, 512U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 0)) - .set_name(param_path + "/Branch_2/Conv2d_0c_1x3/Conv2D") - << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0c_1x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0c_1x3/Relu"); + 3U, 1U, 512U, + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0c_1x3/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0c_1x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0c_1x3/Relu"); SubStream i_c1(i_c); i_c1 << ConvolutionLayer( - 3U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 0)) - .set_name(param_path + "/Branch_2/Conv2d_0d_1x3/Conv2D") + 3U, 1U, 256U, + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0d_1x3/Conv2D") << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0d_1x3/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0d_1x3/Relu"); + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0d_1x3/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0d_1x3/Relu"); SubStream i_c2(i_c); i_c2 << ConvolutionLayer( - 1U, 3U, 256U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 1)) - .set_name(param_path + "/Branch_2/Conv2d_0e_3x1/Conv2D") + 1U, 3U, 256U, + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 1)) + .set_name(param_path + "/Branch_2/Conv2d_0e_3x1/Conv2D") << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_2/Conv2d_0e_3x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0e_3x1/Relu"); + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, + total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0e_3x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0e_3x1/Relu"); // Merge i_c1 and i_c2 i_c << ConcatLayer(std::move(i_c1), std::move(i_c2)).set_name(param_path + "/Branch_2/concat"); SubStream i_d(graph); - i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), - true)) - .set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool") - << ConvolutionLayer(1U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), - get_random_accessor(1.f, 1.f), - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), - 0.001f) - .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu"); + i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, + PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) + .set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool") + << ConvolutionLayer( + 1U, 1U, 256U, + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), 0.001f) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu"); return ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)); } diff --git a/examples/graph_lenet.cpp b/examples/graph_lenet.cpp index 1bcd95fb58..7d6dce7b17 100644 --- a/examples/graph_lenet.cpp +++ b/examples/graph_lenet.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -35,8 +36,7 @@ using namespace arm_compute::graph_utils; class GraphLenetExample : public Example { public: - GraphLenetExample() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "LeNet") + GraphLenetExample() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "LeNet") { } bool do_setup(int argc, char **argv) override @@ -49,14 +49,15 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; } // Checks - ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph"); + ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), + "QASYMM8 not supported for this graph"); // Print parameter values std::cout << common_params << std::endl; @@ -67,43 +68,39 @@ public: // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(28U, 28U, 1U, batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(28U, 28U, 1U, batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; //conv1 << pool1 << conv2 << pool2 << fc1 << act1 << fc2 << smx - graph << common_params.target - << common_params.fast_math_hint + graph << common_params.target << common_params.fast_math_hint << InputLayer(input_descriptor, get_input_accessor(common_params)) << ConvolutionLayer( - 5U, 5U, 20U, - get_weights_accessor(data_path, "/cnn_data/lenet_model/conv1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/lenet_model/conv1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv1") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool1") + 5U, 5U, 20U, get_weights_accessor(data_path, "/cnn_data/lenet_model/conv1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/lenet_model/conv1_b.npy"), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv1") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool1") << ConvolutionLayer( - 5U, 5U, 50U, - get_weights_accessor(data_path, "/cnn_data/lenet_model/conv2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/lenet_model/conv2_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool2") - << FullyConnectedLayer( - 500U, - get_weights_accessor(data_path, "/cnn_data/lenet_model/ip1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/lenet_model/ip1_b.npy")) - .set_name("ip1") + 5U, 5U, 50U, get_weights_accessor(data_path, "/cnn_data/lenet_model/conv2_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/lenet_model/conv2_b.npy"), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool2") + << FullyConnectedLayer(500U, + get_weights_accessor(data_path, "/cnn_data/lenet_model/ip1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/lenet_model/ip1_b.npy")) + .set_name("ip1") << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu") - << FullyConnectedLayer( - 10U, - get_weights_accessor(data_path, "/cnn_data/lenet_model/ip2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/lenet_model/ip2_b.npy")) - .set_name("ip2") - << SoftmaxLayer().set_name("prob") - << OutputLayer(get_output_accessor(common_params)); + << FullyConnectedLayer(10U, + get_weights_accessor(data_path, "/cnn_data/lenet_model/ip2_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/lenet_model/ip2_b.npy")) + .set_name("ip2") + << SoftmaxLayer().set_name("prob") << OutputLayer(get_output_accessor(common_params)); // Finalize graph GraphConfig config; diff --git a/examples/graph_mobilenet.cpp b/examples/graph_mobilenet.cpp index 4630dc958a..e3a6ef116d 100644 --- a/examples/graph_mobilenet.cpp +++ b/examples/graph_mobilenet.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -36,14 +37,13 @@ using namespace arm_compute::graph_utils; class GraphMobilenetExample : public Example { public: - GraphMobilenetExample() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "MobileNetV1") + GraphMobilenetExample() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "MobileNetV1") { // Add model id option model_id_opt = cmd_parser.add_option<SimpleOption<int>>("model-id", 0); model_id_opt->set_help("Mobilenet model id (0: 1.0_224, else: 0.75_160"); } - GraphMobilenetExample(const GraphMobilenetExample &) = delete; + GraphMobilenetExample(const GraphMobilenetExample &) = delete; GraphMobilenetExample &operator=(const GraphMobilenetExample &) = delete; ~GraphMobilenetExample() override = default; bool do_setup(int argc, char **argv) override @@ -56,7 +56,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -72,15 +72,17 @@ public: unsigned int spatial_size = (model_id == 0 || common_params.data_type == DataType::QASYMM8) ? 224 : 160; // Create input descriptor - const TensorShape tensor_shape = permute_shape(TensorShape(spatial_size, spatial_size, 3U, common_params.batches), DataLayout::NCHW, common_params.data_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(spatial_size, spatial_size, 3U, common_params.batches), DataLayout::NCHW, + common_params.data_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); // Set graph hints - graph << common_params.target - << common_params.fast_math_hint; + graph << common_params.target << common_params.fast_math_hint; // Create core graph - if(arm_compute::is_data_type_float(common_params.data_type)) + if (arm_compute::is_data_type_float(common_params.data_type)) { create_graph_float(input_descriptor, model_id); } @@ -90,8 +92,7 @@ public: } // Create common tail - graph << ReshapeLayer(TensorShape(1001U)).set_name("Reshape") - << SoftmaxLayer().set_name("Softmax") + graph << ReshapeLayer(TensorShape(1001U)).set_name("Reshape") << SoftmaxLayer().set_name("Softmax") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph @@ -115,14 +116,15 @@ public: private: CommandLineParser cmd_parser; CommonGraphOptions common_opts; - SimpleOption<int> *model_id_opt{ nullptr }; + SimpleOption<int> *model_id_opt{nullptr}; CommonGraphParams common_params; Stream graph; void create_graph_float(TensorDescriptor &input_descriptor, int model_id) { float depth_scale = (model_id == 0) ? 1.f : 0.75; - std::string model_path = (model_id == 0) ? "/cnn_data/mobilenet_v1_1_224_model/" : "/cnn_data/mobilenet_v1_075_160_model/"; + std::string model_path = + (model_id == 0) ? "/cnn_data/mobilenet_v1_1_224_model/" : "/cnn_data/mobilenet_v1_075_160_model/"; // Create a preprocessor object std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<TFPreproccessor>(); @@ -131,47 +133,68 @@ private: std::string data_path = common_params.data_path; // Add model path to data path - if(!data_path.empty()) + if (!data_path.empty()) { data_path += model_path; } - graph << InputLayer(input_descriptor, - get_input_accessor(common_params, std::move(preprocessor), false)) - << ConvolutionLayer( - 3U, 3U, 32U * depth_scale, - get_weights_accessor(data_path, "Conv2d_0_weights.npy", DataLayout::NCHW), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR)) - .set_name("Conv2d_0") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "Conv2d_0_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "Conv2d_0_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, "Conv2d_0_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, "Conv2d_0_BatchNorm_beta.npy"), - 0.001f) - .set_name("Conv2d_0/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name("Conv2d_0/Relu6"); - graph << get_dwsc_node_float(data_path, "Conv2d_1", 64 * depth_scale, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0)); - graph << get_dwsc_node_float(data_path, "Conv2d_2", 128 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); - graph << get_dwsc_node_float(data_path, "Conv2d_3", 128 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); - graph << get_dwsc_node_float(data_path, "Conv2d_4", 256 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); - graph << get_dwsc_node_float(data_path, "Conv2d_5", 256 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); - graph << get_dwsc_node_float(data_path, "Conv2d_6", 512 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); - graph << get_dwsc_node_float(data_path, "Conv2d_7", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); - graph << get_dwsc_node_float(data_path, "Conv2d_8", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); - graph << get_dwsc_node_float(data_path, "Conv2d_9", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); - graph << get_dwsc_node_float(data_path, "Conv2d_10", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); - graph << get_dwsc_node_float(data_path, "Conv2d_11", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); - graph << get_dwsc_node_float(data_path, "Conv2d_12", 1024 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); - graph << get_dwsc_node_float(data_path, "Conv2d_13", 1024 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); - graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, common_params.data_layout)).set_name("Logits/AvgPool_1a") - << ConvolutionLayer( - 1U, 1U, 1001U, - get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy", DataLayout::NCHW), - get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_biases.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Logits/Conv2d_1c_1x1"); + graph << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false)) + << ConvolutionLayer(3U, 3U, 32U * depth_scale, + get_weights_accessor(data_path, "Conv2d_0_weights.npy", DataLayout::NCHW), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR)) + .set_name("Conv2d_0") + << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_0_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Conv2d_0_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, "Conv2d_0_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, "Conv2d_0_BatchNorm_beta.npy"), 0.001f) + .set_name("Conv2d_0/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) + .set_name("Conv2d_0/Relu6"); + graph << get_dwsc_node_float(data_path, "Conv2d_1", 64 * depth_scale, PadStrideInfo(1, 1, 1, 1), + PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node_float(data_path, "Conv2d_2", 128 * depth_scale, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), + PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node_float(data_path, "Conv2d_3", 128 * depth_scale, + PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), + PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node_float(data_path, "Conv2d_4", 256 * depth_scale, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), + PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node_float(data_path, "Conv2d_5", 256 * depth_scale, + PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), + PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node_float(data_path, "Conv2d_6", 512 * depth_scale, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), + PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node_float(data_path, "Conv2d_7", 512 * depth_scale, + PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), + PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node_float(data_path, "Conv2d_8", 512 * depth_scale, + PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), + PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node_float(data_path, "Conv2d_9", 512 * depth_scale, + PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), + PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node_float(data_path, "Conv2d_10", 512 * depth_scale, + PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), + PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node_float(data_path, "Conv2d_11", 512 * depth_scale, + PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), + PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node_float(data_path, "Conv2d_12", 1024 * depth_scale, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), + PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node_float(data_path, "Conv2d_13", 1024 * depth_scale, + PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), + PadStrideInfo(1, 1, 0, 0)); + graph + << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, common_params.data_layout)).set_name("Logits/AvgPool_1a") + << ConvolutionLayer( + 1U, 1U, 1001U, get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy", DataLayout::NCHW), + get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_biases.npy"), PadStrideInfo(1, 1, 0, 0)) + .set_name("Logits/Conv2d_1c_1x1"); } void create_graph_qasymm(TensorDescriptor &input_descriptor) @@ -180,7 +203,7 @@ private: std::string data_path = common_params.data_path; // Add model path to data path - if(!data_path.empty()) + if (!data_path.empty()) { data_path += "/cnn_data/mobilenet_qasymm8_model/"; } @@ -188,19 +211,16 @@ private: // Quantization info taken from the AndroidNN QASYMM8 MobileNet example const QuantizationInfo in_quant_info = QuantizationInfo(0.0078125f, 128); - const std::vector<QuantizationInfo> conv_weights_quant_info = - { + const std::vector<QuantizationInfo> conv_weights_quant_info = { QuantizationInfo(0.02182667888700962f, 151), // conv0 QuantizationInfo(0.004986600950360298f, 74) // conv14 }; - const std::vector<QuantizationInfo> conv_out_quant_info = - { + const std::vector<QuantizationInfo> conv_out_quant_info = { QuantizationInfo(0.023528477177023888f, 0), // conv0 QuantizationInfo(0.16609922051429749f, 66) // conv14 }; - const std::vector<QuantizationInfo> depth_weights_quant_info = - { + const std::vector<QuantizationInfo> depth_weights_quant_info = { QuantizationInfo(0.29219913482666016f, 110), // dwsc1 QuantizationInfo(0.40277284383773804f, 130), // dwsc2 QuantizationInfo(0.06053730100393295f, 160), // dwsc3 @@ -216,8 +236,7 @@ private: QuantizationInfo(0.12616927921772003f, 211) // dwsc13 }; - const std::vector<QuantizationInfo> point_weights_quant_info = - { + const std::vector<QuantizationInfo> point_weights_quant_info = { QuantizationInfo(0.030420949682593346f, 121), // dwsc1 QuantizationInfo(0.015148180536925793f, 104), // dwsc2 QuantizationInfo(0.013755458407104015f, 94), // dwsc3 @@ -235,108 +254,121 @@ private: graph << InputLayer(input_descriptor.set_quantization_info(in_quant_info), get_input_accessor(common_params, nullptr, false)) - << ConvolutionLayer( - 3U, 3U, 32U, - get_weights_accessor(data_path, "Conv2d_0_weights.npy"), - get_weights_accessor(data_path, "Conv2d_0_bias.npy"), - PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), - 1, conv_weights_quant_info.at(0), conv_out_quant_info.at(0)) - .set_name("Conv2d_0") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)).set_name("Conv2d_0/Relu6"); - graph << get_dwsc_node_qasymm(data_path, "Conv2d_1", 64U, PadStrideInfo(1U, 1U, 1U, 1U), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(0), point_weights_quant_info.at(0)); - graph << get_dwsc_node_qasymm(data_path, "Conv2d_2", 128U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(1), - point_weights_quant_info.at(1)); - graph << get_dwsc_node_qasymm(data_path, "Conv2d_3", 128U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(2), - point_weights_quant_info.at(2)); - graph << get_dwsc_node_qasymm(data_path, "Conv2d_4", 256U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(3), - point_weights_quant_info.at(3)); - graph << get_dwsc_node_qasymm(data_path, "Conv2d_5", 256U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(4), - point_weights_quant_info.at(4)); - graph << get_dwsc_node_qasymm(data_path, "Conv2d_6", 512U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(5), - point_weights_quant_info.at(5)); - graph << get_dwsc_node_qasymm(data_path, "Conv2d_7", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(6), - point_weights_quant_info.at(6)); - graph << get_dwsc_node_qasymm(data_path, "Conv2d_8", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(7), - point_weights_quant_info.at(7)); - graph << get_dwsc_node_qasymm(data_path, "Conv2d_9", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(8), - point_weights_quant_info.at(8)); - graph << get_dwsc_node_qasymm(data_path, "Conv2d_10", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(9), - point_weights_quant_info.at(9)); - graph << get_dwsc_node_qasymm(data_path, "Conv2d_11", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(10), - point_weights_quant_info.at(10)); - graph << get_dwsc_node_qasymm(data_path, "Conv2d_12", 1024U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(11), - point_weights_quant_info.at(11)); - graph << get_dwsc_node_qasymm(data_path, "Conv2d_13", 1024U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(12), - point_weights_quant_info.at(12)) - << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, common_params.data_layout)).set_name("Logits/AvgPool_1a") - << ConvolutionLayer( - 1U, 1U, 1001U, - get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy"), - get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_bias.npy"), - PadStrideInfo(1U, 1U, 0U, 0U), 1, conv_weights_quant_info.at(1), conv_out_quant_info.at(1)) - .set_name("Logits/Conv2d_1c_1x1"); + << ConvolutionLayer(3U, 3U, 32U, get_weights_accessor(data_path, "Conv2d_0_weights.npy"), + get_weights_accessor(data_path, "Conv2d_0_bias.npy"), + PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), 1, + conv_weights_quant_info.at(0), conv_out_quant_info.at(0)) + .set_name("Conv2d_0") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)) + .set_name("Conv2d_0/Relu6"); + graph << get_dwsc_node_qasymm(data_path, "Conv2d_1", 64U, PadStrideInfo(1U, 1U, 1U, 1U), + PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(0), + point_weights_quant_info.at(0)); + graph << get_dwsc_node_qasymm( + data_path, "Conv2d_2", 128U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), + PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(1), point_weights_quant_info.at(1)); + graph << get_dwsc_node_qasymm( + data_path, "Conv2d_3", 128U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), + PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(2), point_weights_quant_info.at(2)); + graph << get_dwsc_node_qasymm( + data_path, "Conv2d_4", 256U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), + PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(3), point_weights_quant_info.at(3)); + graph << get_dwsc_node_qasymm( + data_path, "Conv2d_5", 256U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), + PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(4), point_weights_quant_info.at(4)); + graph << get_dwsc_node_qasymm( + data_path, "Conv2d_6", 512U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), + PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(5), point_weights_quant_info.at(5)); + graph << get_dwsc_node_qasymm( + data_path, "Conv2d_7", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), + PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(6), point_weights_quant_info.at(6)); + graph << get_dwsc_node_qasymm( + data_path, "Conv2d_8", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), + PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(7), point_weights_quant_info.at(7)); + graph << get_dwsc_node_qasymm( + data_path, "Conv2d_9", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), + PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(8), point_weights_quant_info.at(8)); + graph << get_dwsc_node_qasymm( + data_path, "Conv2d_10", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), + PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(9), point_weights_quant_info.at(9)); + graph << get_dwsc_node_qasymm( + data_path, "Conv2d_11", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), + PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(10), point_weights_quant_info.at(10)); + graph << get_dwsc_node_qasymm( + data_path, "Conv2d_12", 1024U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), + PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(11), point_weights_quant_info.at(11)); + graph + << get_dwsc_node_qasymm( + data_path, "Conv2d_13", 1024U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), + PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(12), point_weights_quant_info.at(12)) + << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, common_params.data_layout)).set_name("Logits/AvgPool_1a") + << ConvolutionLayer(1U, 1U, 1001U, get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy"), + get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_bias.npy"), + PadStrideInfo(1U, 1U, 0U, 0U), 1, conv_weights_quant_info.at(1), + conv_out_quant_info.at(1)) + .set_name("Logits/Conv2d_1c_1x1"); } - ConcatLayer get_dwsc_node_float(const std::string &data_path, std::string &¶m_path, - unsigned int conv_filt, - PadStrideInfo dwc_pad_stride_info, PadStrideInfo conv_pad_stride_info) + ConcatLayer get_dwsc_node_float(const std::string &data_path, + std::string &¶m_path, + unsigned int conv_filt, + PadStrideInfo dwc_pad_stride_info, + PadStrideInfo conv_pad_stride_info) { std::string total_path = param_path + "_"; SubStream sg(graph); sg << DepthwiseConvolutionLayer( - 3U, 3U, - get_weights_accessor(data_path, total_path + "depthwise_depthwise_weights.npy", DataLayout::NCHW), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - dwc_pad_stride_info) - .set_name(total_path + "depthwise/depthwise") + 3U, 3U, + get_weights_accessor(data_path, total_path + "depthwise_depthwise_weights.npy", DataLayout::NCHW), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), dwc_pad_stride_info) + .set_name(total_path + "depthwise/depthwise") << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_beta.npy"), - 0.001f) - .set_name(total_path + "depthwise/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name(total_path + "depthwise/Relu6") - << ConvolutionLayer( - 1U, 1U, conv_filt, - get_weights_accessor(data_path, total_path + "pointwise_weights.npy", DataLayout::NCHW), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - conv_pad_stride_info) - .set_name(total_path + "pointwise/Conv2D") + get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_beta.npy"), 0.001f) + .set_name(total_path + "depthwise/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) + .set_name(total_path + "depthwise/Relu6") + << ConvolutionLayer(1U, 1U, conv_filt, + get_weights_accessor(data_path, total_path + "pointwise_weights.npy", DataLayout::NCHW), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), conv_pad_stride_info) + .set_name(total_path + "pointwise/Conv2D") << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_beta.npy"), - 0.001f) - .set_name(total_path + "pointwise/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name(total_path + "pointwise/Relu6"); + get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_beta.npy"), 0.001f) + .set_name(total_path + "pointwise/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) + .set_name(total_path + "pointwise/Relu6"); return ConcatLayer(std::move(sg)); } - ConcatLayer get_dwsc_node_qasymm(const std::string &data_path, std::string &¶m_path, + ConcatLayer get_dwsc_node_qasymm(const std::string &data_path, + std::string &¶m_path, const unsigned int conv_filt, - PadStrideInfo dwc_pad_stride_info, PadStrideInfo conv_pad_stride_info, - QuantizationInfo depth_weights_quant_info, QuantizationInfo point_weights_quant_info) + PadStrideInfo dwc_pad_stride_info, + PadStrideInfo conv_pad_stride_info, + QuantizationInfo depth_weights_quant_info, + QuantizationInfo point_weights_quant_info) { std::string total_path = param_path + "_"; SubStream sg(graph); - sg << DepthwiseConvolutionLayer( - 3U, 3U, - get_weights_accessor(data_path, total_path + "depthwise_weights.npy"), - get_weights_accessor(data_path, total_path + "depthwise_bias.npy"), - dwc_pad_stride_info, 1, std::move(depth_weights_quant_info)) - .set_name(total_path + "depthwise/depthwise") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)).set_name(total_path + "depthwise/Relu6") - << ConvolutionLayer( - 1U, 1U, conv_filt, - get_weights_accessor(data_path, total_path + "pointwise_weights.npy"), - get_weights_accessor(data_path, total_path + "pointwise_bias.npy"), - conv_pad_stride_info, 1, std::move(point_weights_quant_info)) - .set_name(total_path + "pointwise/Conv2D") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)).set_name(total_path + "pointwise/Relu6"); + sg << DepthwiseConvolutionLayer(3U, 3U, get_weights_accessor(data_path, total_path + "depthwise_weights.npy"), + get_weights_accessor(data_path, total_path + "depthwise_bias.npy"), + dwc_pad_stride_info, 1, std::move(depth_weights_quant_info)) + .set_name(total_path + "depthwise/depthwise") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)) + .set_name(total_path + "depthwise/Relu6") + << ConvolutionLayer(1U, 1U, conv_filt, get_weights_accessor(data_path, total_path + "pointwise_weights.npy"), + get_weights_accessor(data_path, total_path + "pointwise_bias.npy"), conv_pad_stride_info, + 1, std::move(point_weights_quant_info)) + .set_name(total_path + "pointwise/Conv2D") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)) + .set_name(total_path + "pointwise/Relu6"); return ConcatLayer(std::move(sg)); } diff --git a/examples/graph_mobilenet_v2.cpp b/examples/graph_mobilenet_v2.cpp index c027e6f13e..9bc21c42c5 100644 --- a/examples/graph_mobilenet_v2.cpp +++ b/examples/graph_mobilenet_v2.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -36,11 +37,10 @@ using namespace arm_compute::graph_utils; class GraphMobilenetV2Example : public Example { public: - GraphMobilenetV2Example() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "MobileNetV2") + GraphMobilenetV2Example() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "MobileNetV2") { } - GraphMobilenetV2Example(const GraphMobilenetV2Example &) = delete; + GraphMobilenetV2Example(const GraphMobilenetV2Example &) = delete; GraphMobilenetV2Example &operator=(const GraphMobilenetV2Example &) = delete; ~GraphMobilenetV2Example() override = default; @@ -54,7 +54,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -64,15 +64,16 @@ public: std::cout << common_params << std::endl; // Create input descriptor - const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, common_params.data_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); + const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), + DataLayout::NCHW, common_params.data_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); // Set graph hints - graph << common_params.target - << common_params.fast_math_hint; + graph << common_params.target << common_params.fast_math_hint; // Create core graph - if(arm_compute::is_data_type_float(common_params.data_type)) + if (arm_compute::is_data_type_float(common_params.data_type)) { create_graph_float(input_descriptor); } @@ -82,8 +83,7 @@ public: } // Create common tail graph << ReshapeLayer(TensorShape(1001U)).set_name("Predictions/Reshape") - << SoftmaxLayer().set_name("Predictions/Softmax") - << OutputLayer(get_output_accessor(common_params, 5)); + << SoftmaxLayer().set_name("Predictions/Softmax") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph GraphConfig config; @@ -136,123 +136,143 @@ private: std::string data_path = common_params.data_path; // Add model path to data path - if(!data_path.empty()) + if (!data_path.empty()) { data_path += model_path; } graph << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false)) - << ConvolutionLayer(3U, 3U, 32U, - get_weights_accessor(data_path, "Conv_weights.npy", DataLayout::NCHW), + << ConvolutionLayer(3U, 3U, 32U, get_weights_accessor(data_path, "Conv_weights.npy", DataLayout::NCHW), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL)) - .set_name("Conv") + .set_name("Conv") << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "Conv_BatchNorm_moving_variance.npy"), get_weights_accessor(data_path, "Conv_BatchNorm_gamma.npy"), get_weights_accessor(data_path, "Conv_BatchNorm_beta.npy"), 0.0010000000474974513f) - .set_name("Conv/BatchNorm") + .set_name("Conv/BatchNorm") << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) - .set_name("Conv/Relu6"); + .set_name("Conv/Relu6"); get_expanded_conv_float(data_path, "expanded_conv", 32U, 16U, PadStrideInfo(1, 1, 1, 1)); - get_expanded_conv_float(data_path, "expanded_conv_1", 16U, 24U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), HasExpand::Yes); - get_expanded_conv_float(data_path, "expanded_conv_2", 24U, 24U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes); - get_expanded_conv_float(data_path, "expanded_conv_3", 24U, 32U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), HasExpand::Yes); - get_expanded_conv_float(data_path, "expanded_conv_4", 32U, 32U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes); - get_expanded_conv_float(data_path, "expanded_conv_5", 32U, 32U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes); - get_expanded_conv_float(data_path, "expanded_conv_6", 32U, 64U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), HasExpand::Yes); - get_expanded_conv_float(data_path, "expanded_conv_7", 64U, 64U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes); - get_expanded_conv_float(data_path, "expanded_conv_8", 64U, 64U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes); - get_expanded_conv_float(data_path, "expanded_conv_9", 64U, 64U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes); + get_expanded_conv_float(data_path, "expanded_conv_1", 16U, 24U, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), HasExpand::Yes); + get_expanded_conv_float(data_path, "expanded_conv_2", 24U, 24U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, + IsResidual::Yes); + get_expanded_conv_float(data_path, "expanded_conv_3", 24U, 32U, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), HasExpand::Yes); + get_expanded_conv_float(data_path, "expanded_conv_4", 32U, 32U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, + IsResidual::Yes); + get_expanded_conv_float(data_path, "expanded_conv_5", 32U, 32U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, + IsResidual::Yes); + get_expanded_conv_float(data_path, "expanded_conv_6", 32U, 64U, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), HasExpand::Yes); + get_expanded_conv_float(data_path, "expanded_conv_7", 64U, 64U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, + IsResidual::Yes); + get_expanded_conv_float(data_path, "expanded_conv_8", 64U, 64U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, + IsResidual::Yes); + get_expanded_conv_float(data_path, "expanded_conv_9", 64U, 64U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, + IsResidual::Yes); get_expanded_conv_float(data_path, "expanded_conv_10", 64U, 96U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes); - get_expanded_conv_float(data_path, "expanded_conv_11", 96U, 96U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes); - get_expanded_conv_float(data_path, "expanded_conv_12", 96U, 96U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes); - get_expanded_conv_float(data_path, "expanded_conv_13", 96U, 160U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), HasExpand::Yes); - get_expanded_conv_float(data_path, "expanded_conv_14", 160U, 160U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes); - get_expanded_conv_float(data_path, "expanded_conv_15", 160U, 160U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, IsResidual::Yes); + get_expanded_conv_float(data_path, "expanded_conv_11", 96U, 96U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, + IsResidual::Yes); + get_expanded_conv_float(data_path, "expanded_conv_12", 96U, 96U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, + IsResidual::Yes); + get_expanded_conv_float(data_path, "expanded_conv_13", 96U, 160U, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), HasExpand::Yes); + get_expanded_conv_float(data_path, "expanded_conv_14", 160U, 160U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, + IsResidual::Yes); + get_expanded_conv_float(data_path, "expanded_conv_15", 160U, 160U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes, + IsResidual::Yes); get_expanded_conv_float(data_path, "expanded_conv_16", 160U, 320U, PadStrideInfo(1, 1, 1, 1), HasExpand::Yes); - graph << ConvolutionLayer(1U, 1U, 1280U, - get_weights_accessor(data_path, "Conv_1_weights.npy", DataLayout::NCHW), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("Conv_1") + graph << ConvolutionLayer( + 1U, 1U, 1280U, get_weights_accessor(data_path, "Conv_1_weights.npy", DataLayout::NCHW), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Conv_1") << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv_1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "Conv_1_BatchNorm_moving_variance.npy"), get_weights_accessor(data_path, "Conv_1_BatchNorm_gamma.npy"), get_weights_accessor(data_path, "Conv_1_BatchNorm_beta.npy"), 0.0010000000474974513f) - .set_name("Conv_1/BatchNorm") + .set_name("Conv_1/BatchNorm") << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) - .set_name("Conv_1/Relu6") + .set_name("Conv_1/Relu6") << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, common_params.data_layout)).set_name("Logits/AvgPool") << ConvolutionLayer(1U, 1U, 1001U, get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy", DataLayout::NCHW), get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_biases.npy"), PadStrideInfo(1, 1, 0, 0)) - .set_name("Logits/Conv2d_1c_1x1"); + .set_name("Logits/Conv2d_1c_1x1"); } - void get_expanded_conv_float(const std::string &data_path, std::string &¶m_path, - unsigned int input_channels, unsigned int output_channels, - PadStrideInfo dwc_pad_stride_info, - HasExpand has_expand = HasExpand::No, IsResidual is_residual = IsResidual::No, - unsigned int expansion_size = 6) + void get_expanded_conv_float(const std::string &data_path, + std::string &¶m_path, + unsigned int input_channels, + unsigned int output_channels, + PadStrideInfo dwc_pad_stride_info, + HasExpand has_expand = HasExpand::No, + IsResidual is_residual = IsResidual::No, + unsigned int expansion_size = 6) { std::string total_path = param_path + "_"; SubStream left(graph); // Add expand node - if(has_expand == HasExpand::Yes) + if (has_expand == HasExpand::Yes) { - left << ConvolutionLayer(1U, 1U, input_channels * expansion_size, - get_weights_accessor(data_path, total_path + "expand_weights.npy", DataLayout::NCHW), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/expand/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "expand_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "expand_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, total_path + "expand_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, total_path + "expand_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(param_path + "/expand/BatchNorm") + left << ConvolutionLayer( + 1U, 1U, input_channels * expansion_size, + get_weights_accessor(data_path, total_path + "expand_weights.npy", DataLayout::NCHW), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/expand/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "expand_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "expand_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, total_path + "expand_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, total_path + "expand_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(param_path + "/expand/BatchNorm") << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) - .set_name(param_path + "/expand/Relu6"); + .set_name(param_path + "/expand/Relu6"); } // Add depthwise node - left << DepthwiseConvolutionLayer(3U, 3U, - get_weights_accessor(data_path, total_path + "depthwise_depthwise_weights.npy", DataLayout::NCHW), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - dwc_pad_stride_info) - .set_name(param_path + "/depthwise/depthwise") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_beta.npy"), - 0.0010000000474974513f) - .set_name(param_path + "/depthwise/BatchNorm") + left << DepthwiseConvolutionLayer( + 3U, 3U, + get_weights_accessor(data_path, total_path + "depthwise_depthwise_weights.npy", DataLayout::NCHW), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), dwc_pad_stride_info) + .set_name(param_path + "/depthwise/depthwise") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_beta.npy"), + 0.0010000000474974513f) + .set_name(param_path + "/depthwise/BatchNorm") << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) - .set_name(param_path + "/depthwise/Relu6"); + .set_name(param_path + "/depthwise/Relu6"); // Add project node left << ConvolutionLayer(1U, 1U, output_channels, get_weights_accessor(data_path, total_path + "project_weights.npy", DataLayout::NCHW), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/project/Conv2D") - << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "project_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, total_path + "project_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, total_path + "project_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, total_path + "project_BatchNorm_beta.npy"), - 0.0010000000474974513) - .set_name(param_path + "/project/BatchNorm"); - - if(is_residual == IsResidual::Yes) + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/project/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "project_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "project_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, total_path + "project_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, total_path + "project_BatchNorm_beta.npy"), 0.0010000000474974513) + .set_name(param_path + "/project/BatchNorm"); + + if (is_residual == IsResidual::Yes) { // Add residual node SubStream right(graph); - graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(param_path + "/add"); + graph + << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(param_path + "/add"); } else { @@ -269,7 +289,7 @@ private: std::string data_path = common_params.data_path; // Add model path to data path - if(!data_path.empty()) + if (!data_path.empty()) { data_path += model_path; } @@ -277,16 +297,14 @@ private: const QuantizationInfo in_quant_info = QuantizationInfo(0.0078125f, 128); const QuantizationInfo mid_quant_info = QuantizationInfo(0.023528477177023888f, 128); - const std::vector<QuantizationInfo> conv_weights_quant_info = - { + const std::vector<QuantizationInfo> conv_weights_quant_info = { QuantizationInfo(0.03396892547607422f, 122), // Conv QuantizationInfo(0.005167067516595125f, 125), // Conv1 QuantizationInfo(0.0016910821432247758f, 113) // Conv2d_1c_1x1 }; // Pointwise expand convolution quantization info - const std::vector<QuantizationInfo> pwc_q = - { + const std::vector<QuantizationInfo> pwc_q = { QuantizationInfo(0.254282623529f, 129), // expand_0 (Dummy) QuantizationInfo(0.009758507832884789f, 127), // expand_1 QuantizationInfo(0.0036556976847350597f, 144), // expand_2 @@ -306,8 +324,7 @@ private: QuantizationInfo(0.002046825597062707f, 135) // expand_16 }; // Depthwise expand convolution quantization info - const std::vector<QuantizationInfo> dwc_q = - { + const std::vector<QuantizationInfo> dwc_q = { QuantizationInfo(0.3436955213546753f, 165), // expand_0 QuantizationInfo(0.020969120785593987f, 109), // expand_1 QuantizationInfo(0.16981913149356842f, 52), // expand_2 @@ -327,8 +344,7 @@ private: QuantizationInfo(0.16456253826618195, 201) // expand_16 }; // Project convolution quantization info - const std::vector<QuantizationInfo> prwc_q = - { + const std::vector<QuantizationInfo> prwc_q = { QuantizationInfo(0.03737175464630127f, 140), // expand_0 QuantizationInfo(0.0225360207259655f, 156), // expand_1 QuantizationInfo(0.02740888111293316f, 122), // expand_2 @@ -350,65 +366,84 @@ private: graph << InputLayer(input_descriptor.set_quantization_info(in_quant_info), get_weights_accessor(data_path, common_params.image)) - << ConvolutionLayer( - 3U, 3U, 32U, - get_weights_accessor(data_path, "Conv_weights.npy"), - get_weights_accessor(data_path, "Conv_bias.npy"), - PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), - 1, conv_weights_quant_info.at(0), mid_quant_info) - .set_name("Conv") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)).set_name("Conv/Relu6") - << DepthwiseConvolutionLayer(3U, 3U, - get_weights_accessor(data_path, "expanded_conv_depthwise_depthwise_weights.npy"), - get_weights_accessor(data_path, "expanded_conv_depthwise_depthwise_biases.npy"), - PadStrideInfo(1, 1, 1, 1), 1, dwc_q.at(0)) - .set_name("expanded_conv/depthwise/depthwise") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)).set_name("expanded_conv/depthwise/Relu6") - << ConvolutionLayer(1U, 1U, 16U, - get_weights_accessor(data_path, "expanded_conv_project_weights.npy"), + << ConvolutionLayer(3U, 3U, 32U, get_weights_accessor(data_path, "Conv_weights.npy"), + get_weights_accessor(data_path, "Conv_bias.npy"), + PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), 1, + conv_weights_quant_info.at(0), mid_quant_info) + .set_name("Conv") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)) + .set_name("Conv/Relu6") + << DepthwiseConvolutionLayer( + 3U, 3U, get_weights_accessor(data_path, "expanded_conv_depthwise_depthwise_weights.npy"), + get_weights_accessor(data_path, "expanded_conv_depthwise_depthwise_biases.npy"), + PadStrideInfo(1, 1, 1, 1), 1, dwc_q.at(0)) + .set_name("expanded_conv/depthwise/depthwise") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)) + .set_name("expanded_conv/depthwise/Relu6") + << ConvolutionLayer(1U, 1U, 16U, get_weights_accessor(data_path, "expanded_conv_project_weights.npy"), get_weights_accessor(data_path, "expanded_conv_project_biases.npy"), PadStrideInfo(1, 1, 0, 0), 1, prwc_q.at(0)) - .set_name("expanded_conv/project/Conv2D"); - - get_expanded_conv_qasymm8(data_path, "expanded_conv_1", IsResidual::No, 96U, 24U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), - pwc_q.at(1), dwc_q.at(1), prwc_q.at(1)); - get_expanded_conv_qasymm8(data_path, "expanded_conv_2", IsResidual::Yes, 144U, 24U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(2), dwc_q.at(2), prwc_q.at(2)); - get_expanded_conv_qasymm8(data_path, "expanded_conv_3", IsResidual::No, 144U, 32U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), - pwc_q.at(3), dwc_q.at(3), prwc_q.at(3)); - get_expanded_conv_qasymm8(data_path, "expanded_conv_4", IsResidual::Yes, 192U, 32U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(4), dwc_q.at(4), prwc_q.at(4)); - get_expanded_conv_qasymm8(data_path, "expanded_conv_5", IsResidual::Yes, 192U, 32U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(5), dwc_q.at(5), prwc_q.at(5)); - get_expanded_conv_qasymm8(data_path, "expanded_conv_6", IsResidual::No, 192U, 64U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), - pwc_q.at(6), dwc_q.at(6), prwc_q.at(6)); - get_expanded_conv_qasymm8(data_path, "expanded_conv_7", IsResidual::Yes, 384U, 64U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(7), dwc_q.at(7), prwc_q.at(7)); - get_expanded_conv_qasymm8(data_path, "expanded_conv_8", IsResidual::Yes, 384U, 64U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(8), dwc_q.at(8), prwc_q.at(8)); - get_expanded_conv_qasymm8(data_path, "expanded_conv_9", IsResidual::Yes, 384U, 64U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(9), dwc_q.at(9), prwc_q.at(9)); - get_expanded_conv_qasymm8(data_path, "expanded_conv_10", IsResidual::No, 384U, 96U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(10), dwc_q.at(10), prwc_q.at(10)); - get_expanded_conv_qasymm8(data_path, "expanded_conv_11", IsResidual::Yes, 576U, 96U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(11), dwc_q.at(11), prwc_q.at(11)); - get_expanded_conv_qasymm8(data_path, "expanded_conv_12", IsResidual::Yes, 576U, 96U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(12), dwc_q.at(12), prwc_q.at(12)); - get_expanded_conv_qasymm8(data_path, "expanded_conv_13", IsResidual::No, 576U, 160U, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), - pwc_q.at(13), dwc_q.at(13), prwc_q.at(13)); - get_expanded_conv_qasymm8(data_path, "expanded_conv_14", IsResidual::Yes, 960U, 160U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(14), dwc_q.at(14), prwc_q.at(14)); - get_expanded_conv_qasymm8(data_path, "expanded_conv_15", IsResidual::Yes, 960U, 160U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(15), dwc_q.at(15), prwc_q.at(15)); - get_expanded_conv_qasymm8(data_path, "expanded_conv_16", IsResidual::No, 960U, 320U, PadStrideInfo(1, 1, 1, 1), pwc_q.at(16), dwc_q.at(16), prwc_q.at(16)); - - graph << ConvolutionLayer(1U, 1U, 1280U, - get_weights_accessor(data_path, "Conv_1_weights.npy"), - get_weights_accessor(data_path, "Conv_1_biases.npy"), - PadStrideInfo(1, 1, 0, 0), 1, conv_weights_quant_info.at(1)) - .set_name("Conv_1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)).set_name("Conv_1/Relu6") + .set_name("expanded_conv/project/Conv2D"); + + get_expanded_conv_qasymm8(data_path, "expanded_conv_1", IsResidual::No, 96U, 24U, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), pwc_q.at(1), + dwc_q.at(1), prwc_q.at(1)); + get_expanded_conv_qasymm8(data_path, "expanded_conv_2", IsResidual::Yes, 144U, 24U, PadStrideInfo(1, 1, 1, 1), + pwc_q.at(2), dwc_q.at(2), prwc_q.at(2)); + get_expanded_conv_qasymm8(data_path, "expanded_conv_3", IsResidual::No, 144U, 32U, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), pwc_q.at(3), + dwc_q.at(3), prwc_q.at(3)); + get_expanded_conv_qasymm8(data_path, "expanded_conv_4", IsResidual::Yes, 192U, 32U, PadStrideInfo(1, 1, 1, 1), + pwc_q.at(4), dwc_q.at(4), prwc_q.at(4)); + get_expanded_conv_qasymm8(data_path, "expanded_conv_5", IsResidual::Yes, 192U, 32U, PadStrideInfo(1, 1, 1, 1), + pwc_q.at(5), dwc_q.at(5), prwc_q.at(5)); + get_expanded_conv_qasymm8(data_path, "expanded_conv_6", IsResidual::No, 192U, 64U, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), pwc_q.at(6), + dwc_q.at(6), prwc_q.at(6)); + get_expanded_conv_qasymm8(data_path, "expanded_conv_7", IsResidual::Yes, 384U, 64U, PadStrideInfo(1, 1, 1, 1), + pwc_q.at(7), dwc_q.at(7), prwc_q.at(7)); + get_expanded_conv_qasymm8(data_path, "expanded_conv_8", IsResidual::Yes, 384U, 64U, PadStrideInfo(1, 1, 1, 1), + pwc_q.at(8), dwc_q.at(8), prwc_q.at(8)); + get_expanded_conv_qasymm8(data_path, "expanded_conv_9", IsResidual::Yes, 384U, 64U, PadStrideInfo(1, 1, 1, 1), + pwc_q.at(9), dwc_q.at(9), prwc_q.at(9)); + get_expanded_conv_qasymm8(data_path, "expanded_conv_10", IsResidual::No, 384U, 96U, PadStrideInfo(1, 1, 1, 1), + pwc_q.at(10), dwc_q.at(10), prwc_q.at(10)); + get_expanded_conv_qasymm8(data_path, "expanded_conv_11", IsResidual::Yes, 576U, 96U, PadStrideInfo(1, 1, 1, 1), + pwc_q.at(11), dwc_q.at(11), prwc_q.at(11)); + get_expanded_conv_qasymm8(data_path, "expanded_conv_12", IsResidual::Yes, 576U, 96U, PadStrideInfo(1, 1, 1, 1), + pwc_q.at(12), dwc_q.at(12), prwc_q.at(12)); + get_expanded_conv_qasymm8(data_path, "expanded_conv_13", IsResidual::No, 576U, 160U, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), pwc_q.at(13), + dwc_q.at(13), prwc_q.at(13)); + get_expanded_conv_qasymm8(data_path, "expanded_conv_14", IsResidual::Yes, 960U, 160U, PadStrideInfo(1, 1, 1, 1), + pwc_q.at(14), dwc_q.at(14), prwc_q.at(14)); + get_expanded_conv_qasymm8(data_path, "expanded_conv_15", IsResidual::Yes, 960U, 160U, PadStrideInfo(1, 1, 1, 1), + pwc_q.at(15), dwc_q.at(15), prwc_q.at(15)); + get_expanded_conv_qasymm8(data_path, "expanded_conv_16", IsResidual::No, 960U, 320U, PadStrideInfo(1, 1, 1, 1), + pwc_q.at(16), dwc_q.at(16), prwc_q.at(16)); + + graph << ConvolutionLayer(1U, 1U, 1280U, get_weights_accessor(data_path, "Conv_1_weights.npy"), + get_weights_accessor(data_path, "Conv_1_biases.npy"), PadStrideInfo(1, 1, 0, 0), 1, + conv_weights_quant_info.at(1)) + .set_name("Conv_1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)) + .set_name("Conv_1/Relu6") << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, common_params.data_layout)).set_name("Logits/AvgPool") - << ConvolutionLayer(1U, 1U, 1001U, - get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy"), + << ConvolutionLayer(1U, 1U, 1001U, get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy"), get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_biases.npy"), PadStrideInfo(1, 1, 0, 0), 1, conv_weights_quant_info.at(2)) - .set_name("Logits/Conv2d_1c_1x1"); + .set_name("Logits/Conv2d_1c_1x1"); } - void get_expanded_conv_qasymm8(const std::string &data_path, std::string &¶m_path, IsResidual is_residual, - unsigned int input_channels, unsigned int output_channels, + void get_expanded_conv_qasymm8(const std::string &data_path, + std::string &¶m_path, + IsResidual is_residual, + unsigned int input_channels, + unsigned int output_channels, PadStrideInfo dwc_pad_stride_info, - const QuantizationInfo &pwi, const QuantizationInfo &dwi, const QuantizationInfo &pji) + const QuantizationInfo &pwi, + const QuantizationInfo &dwi, + const QuantizationInfo &pji) { std::string total_path = param_path + "_"; @@ -417,25 +452,28 @@ private: get_weights_accessor(data_path, total_path + "project_weights.npy"), get_weights_accessor(data_path, total_path + "project_biases.npy"), PadStrideInfo(1, 1, 0, 0), 1, pwi) - .set_name(param_path + "/Conv2D") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)).set_name(param_path + "/Conv2D/Relu6") - << DepthwiseConvolutionLayer(3U, 3U, - get_weights_accessor(data_path, total_path + "depthwise_depthwise_weights.npy"), - get_weights_accessor(data_path, total_path + "depthwise_depthwise_biases.npy"), - dwc_pad_stride_info, 1, dwi) - .set_name(param_path + "/depthwise/depthwise") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)).set_name(param_path + "/depthwise/Relu6") + .set_name(param_path + "/Conv2D") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)) + .set_name(param_path + "/Conv2D/Relu6") + << DepthwiseConvolutionLayer( + 3U, 3U, get_weights_accessor(data_path, total_path + "depthwise_depthwise_weights.npy"), + get_weights_accessor(data_path, total_path + "depthwise_depthwise_biases.npy"), dwc_pad_stride_info, + 1, dwi) + .set_name(param_path + "/depthwise/depthwise") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f)) + .set_name(param_path + "/depthwise/Relu6") << ConvolutionLayer(1U, 1U, output_channels, get_weights_accessor(data_path, total_path + "project_weights.npy"), get_weights_accessor(data_path, total_path + "project_biases.npy"), PadStrideInfo(1, 1, 0, 0), 1, pji) - .set_name(param_path + "/project/Conv2D"); + .set_name(param_path + "/project/Conv2D"); - if(is_residual == IsResidual::Yes) + if (is_residual == IsResidual::Yes) { // Add residual node SubStream right(graph); - graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(param_path + "/add"); + graph + << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(param_path + "/add"); } else { diff --git a/examples/graph_resnet12.cpp b/examples/graph_resnet12.cpp index 48708ce29a..80db826be5 100644 --- a/examples/graph_resnet12.cpp +++ b/examples/graph_resnet12.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -36,7 +37,12 @@ class GraphResNet12Example : public Example { public: GraphResNet12Example() - : cmd_parser(), common_opts(cmd_parser), model_input_width(nullptr), model_input_height(nullptr), common_params(), graph(0, "ResNet12") + : cmd_parser(), + common_opts(cmd_parser), + model_input_width(nullptr), + model_input_height(nullptr), + common_params(), + graph(0, "ResNet12") { model_input_width = cmd_parser.add_option<SimpleOption<unsigned int>>("image-width", 192); model_input_height = cmd_parser.add_option<SimpleOption<unsigned int>>("image-height", 128); @@ -45,7 +51,7 @@ public: model_input_width->set_help("Input image width."); model_input_height->set_help("Input image height."); } - GraphResNet12Example(const GraphResNet12Example &) = delete; + GraphResNet12Example(const GraphResNet12Example &) = delete; GraphResNet12Example &operator=(const GraphResNet12Example &) = delete; ~GraphResNet12Example() override = default; bool do_setup(int argc, char **argv) override @@ -58,7 +64,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -69,7 +75,8 @@ public: const unsigned int image_height = model_input_height->value(); // Checks - ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph"); + ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), + "QASYMM8 not supported for this graph"); // Print parameter values std::cout << common_params << std::endl; @@ -84,50 +91,47 @@ public: std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<TFPreproccessor>(); // Create input descriptor - const TensorShape tensor_shape = permute_shape(TensorShape(image_width, image_height, 3U, common_params.batches), DataLayout::NCHW, common_params.data_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(image_width, image_height, 3U, common_params.batches), DataLayout::NCHW, + common_params.data_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint - << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false /* Do not convert to BGR */)) - << ConvolutionLayer( - 9U, 9U, 64U, - get_weights_accessor(data_path, "conv1_weights.npy", weights_layout), - get_weights_accessor(data_path, "conv1_biases.npy", weights_layout), - PadStrideInfo(1, 1, 4, 4)) - .set_name("conv1/convolution") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1/Relu"); + graph << common_params.target << common_params.fast_math_hint + << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), + false /* Do not convert to BGR */)) + << ConvolutionLayer(9U, 9U, 64U, get_weights_accessor(data_path, "conv1_weights.npy", weights_layout), + get_weights_accessor(data_path, "conv1_biases.npy", weights_layout), + PadStrideInfo(1, 1, 4, 4)) + .set_name("conv1/convolution") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv1/Relu"); add_residual_block(data_path, "block1", weights_layout); add_residual_block(data_path, "block2", weights_layout); add_residual_block(data_path, "block3", weights_layout); add_residual_block(data_path, "block4", weights_layout); - graph << ConvolutionLayer( - 3U, 3U, 64U, - get_weights_accessor(data_path, "conv10_weights.npy", weights_layout), - get_weights_accessor(data_path, "conv10_biases.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv10/convolution") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv10/Relu") - << ConvolutionLayer( - 3U, 3U, 64U, - get_weights_accessor(data_path, "conv11_weights.npy", weights_layout), - get_weights_accessor(data_path, "conv11_biases.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv11/convolution") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv11/Relu") - << ConvolutionLayer( - 9U, 9U, 3U, - get_weights_accessor(data_path, "conv12_weights.npy", weights_layout), - get_weights_accessor(data_path, "conv12_biases.npy"), - PadStrideInfo(1, 1, 4, 4)) - .set_name("conv12/convolution") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH)).set_name("conv12/Tanh") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 0.58f, 0.5f)).set_name("conv12/Linear") + graph << ConvolutionLayer(3U, 3U, 64U, get_weights_accessor(data_path, "conv10_weights.npy", weights_layout), + get_weights_accessor(data_path, "conv10_biases.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv10/convolution") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv10/Relu") + << ConvolutionLayer(3U, 3U, 64U, get_weights_accessor(data_path, "conv11_weights.npy", weights_layout), + get_weights_accessor(data_path, "conv11_biases.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv11/convolution") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv11/Relu") + << ConvolutionLayer(9U, 9U, 3U, get_weights_accessor(data_path, "conv12_weights.npy", weights_layout), + get_weights_accessor(data_path, "conv12_biases.npy"), PadStrideInfo(1, 1, 4, 4)) + .set_name("conv12/convolution") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH)) + .set_name("conv12/Tanh") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 0.58f, 0.5f)) + .set_name("conv12/Linear") << OutputLayer(std::make_unique<DummyAccessor>(0)); // Finalize graph @@ -152,8 +156,8 @@ public: private: CommandLineParser cmd_parser; CommonGraphOptions common_opts; - SimpleOption<unsigned int> *model_input_width{ nullptr }; - SimpleOption<unsigned int> *model_input_height{ nullptr }; + SimpleOption<unsigned int> *model_input_width{nullptr}; + SimpleOption<unsigned int> *model_input_height{nullptr}; CommonGraphParams common_params; Stream graph; @@ -170,35 +174,33 @@ private: SubStream left(graph); SubStream right(graph); - right << ConvolutionLayer( - 3U, 3U, 64U, - get_weights_accessor(data_path, unit_path + "conv1_weights.npy", weights_layout), - get_weights_accessor(data_path, unit_path + "conv1_biases.npy", weights_layout), - PadStrideInfo(1, 1, 1, 1)) - .set_name(unit_name + "conv1/convolution") + right << ConvolutionLayer(3U, 3U, 64U, + get_weights_accessor(data_path, unit_path + "conv1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "conv1_biases.npy", weights_layout), + PadStrideInfo(1, 1, 1, 1)) + .set_name(unit_name + "conv1/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_beta.npy"), - 0.0000100099996416f) - .set_name(unit_name + "conv1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu") - - << ConvolutionLayer( - 3U, 3U, 64U, - get_weights_accessor(data_path, unit_path + "conv2_weights.npy", weights_layout), - get_weights_accessor(data_path, unit_path + "conv2_biases.npy", weights_layout), - PadStrideInfo(1, 1, 1, 1)) - .set_name(unit_name + "conv2/convolution") + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_beta.npy"), 0.0000100099996416f) + .set_name(unit_name + "conv1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "conv1/Relu") + + << ConvolutionLayer(3U, 3U, 64U, + get_weights_accessor(data_path, unit_path + "conv2_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "conv2_biases.npy", weights_layout), + PadStrideInfo(1, 1, 1, 1)) + .set_name(unit_name + "conv2/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_beta.npy"), - 0.0000100099996416f) - .set_name(unit_name + "conv2/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv2/Relu"); + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_beta.npy"), 0.0000100099996416f) + .set_name(unit_name + "conv2/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "conv2/Relu"); graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(unit_name + "add"); } diff --git a/examples/graph_resnet50.cpp b/examples/graph_resnet50.cpp index 0d3322c886..ba0f0d5fb6 100644 --- a/examples/graph_resnet50.cpp +++ b/examples/graph_resnet50.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -35,8 +36,7 @@ using namespace arm_compute::graph_utils; class GraphResNetV1_50Example : public Example { public: - GraphResNetV1_50Example() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ResNetV1_50") + GraphResNetV1_50Example() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ResNetV1_50") { } bool do_setup(int argc, char **argv) override @@ -49,7 +49,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -62,36 +62,40 @@ public: std::string data_path = common_params.data_path; // Create a preprocessor object - const std::array<float, 3> mean_rgb{ { 122.68f, 116.67f, 104.01f } }; - std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<CaffePreproccessor>(mean_rgb, - false /* Do not convert to BGR */); + const std::array<float, 3> mean_rgb{{122.68f, 116.67f, 104.01f}}; + std::unique_ptr<IPreprocessor> preprocessor = + std::make_unique<CaffePreproccessor>(mean_rgb, false /* Do not convert to BGR */); // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint - << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false /* Do not convert to BGR */)) + graph << common_params.target << common_params.fast_math_hint + << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), + false /* Do not convert to BGR */)) << ConvolutionLayer( - 7U, 7U, 64U, - get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 3, 3)) - .set_name("conv1/convolution") + 7U, 7U, 64U, + get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 3, 3)) + .set_name("conv1/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_beta.npy"), - 0.0000100099996416f) - .set_name("conv1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))).set_name("pool1/MaxPool"); + get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_beta.npy"), + 0.0000100099996416f) + .set_name("conv1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv1/Relu") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))) + .set_name("pool1/MaxPool"); add_residual_block(data_path, "block1", weights_layout, 64, 3, 2); add_residual_block(data_path, "block2", weights_layout, 128, 4, 2); @@ -100,13 +104,12 @@ public: graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("pool5") << ConvolutionLayer( - 1U, 1U, 1000U, - get_weights_accessor(data_path, "/cnn_data/resnet50_model/logits_weights.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/resnet50_model/logits_biases.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("logits/convolution") - << FlattenLayer().set_name("predictions/Reshape") - << SoftmaxLayer().set_name("predictions/Softmax") + 1U, 1U, 1000U, + get_weights_accessor(data_path, "/cnn_data/resnet50_model/logits_weights.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/resnet50_model/logits_biases.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("logits/convolution") + << FlattenLayer().set_name("predictions/Reshape") << SoftmaxLayer().set_name("predictions/Softmax") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph @@ -136,10 +139,14 @@ private: CommonGraphParams common_params; Stream graph; - void add_residual_block(const std::string &data_path, const std::string &name, DataLayout weights_layout, - unsigned int base_depth, unsigned int num_units, unsigned int stride) + void add_residual_block(const std::string &data_path, + const std::string &name, + DataLayout weights_layout, + unsigned int base_depth, + unsigned int num_units, + unsigned int stride) { - for(unsigned int i = 0; i < num_units; ++i) + for (unsigned int i = 0; i < num_units; ++i) { std::stringstream unit_path_ss; unit_path_ss << "/cnn_data/resnet50_model/" << name << "_unit_" << (i + 1) << "_bottleneck_v1_"; @@ -151,89 +158,90 @@ private: unsigned int middle_stride = 1; - if(i == (num_units - 1)) + if (i == (num_units - 1)) { middle_stride = stride; } SubStream right(graph); - right << ConvolutionLayer( - 1U, 1U, base_depth, - get_weights_accessor(data_path, unit_path + "conv1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "conv1/convolution") + right << ConvolutionLayer(1U, 1U, base_depth, + get_weights_accessor(data_path, unit_path + "conv1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "conv1/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_beta.npy"), - 0.0000100099996416f) - .set_name(unit_name + "conv1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu") + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_beta.npy"), 0.0000100099996416f) + .set_name(unit_name + "conv1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "conv1/Relu") - << ConvolutionLayer( - 3U, 3U, base_depth, - get_weights_accessor(data_path, unit_path + "conv2_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(middle_stride, middle_stride, 1, 1)) - .set_name(unit_name + "conv2/convolution") + << ConvolutionLayer(3U, 3U, base_depth, + get_weights_accessor(data_path, unit_path + "conv2_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(middle_stride, middle_stride, 1, 1)) + .set_name(unit_name + "conv2/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_beta.npy"), - 0.0000100099996416f) - .set_name(unit_name + "conv2/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu") + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_beta.npy"), 0.0000100099996416f) + .set_name(unit_name + "conv2/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "conv1/Relu") - << ConvolutionLayer( - 1U, 1U, base_depth * 4, - get_weights_accessor(data_path, unit_path + "conv3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "conv3/convolution") + << ConvolutionLayer(1U, 1U, base_depth * 4, + get_weights_accessor(data_path, unit_path + "conv3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "conv3/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_beta.npy"), - 0.0000100099996416f) - .set_name(unit_name + "conv2/BatchNorm"); + get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_beta.npy"), 0.0000100099996416f) + .set_name(unit_name + "conv2/BatchNorm"); - if(i == 0) + if (i == 0) { SubStream left(graph); left << ConvolutionLayer( - 1U, 1U, base_depth * 4, - get_weights_accessor(data_path, unit_path + "shortcut_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "shortcut/convolution") + 1U, 1U, base_depth * 4, + get_weights_accessor(data_path, unit_path + "shortcut_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "shortcut/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_beta.npy"), - 0.0000100099996416f) - .set_name(unit_name + "shortcut/BatchNorm"); + get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_beta.npy"), + 0.0000100099996416f) + .set_name(unit_name + "shortcut/BatchNorm"); - graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(unit_name + "add"); + graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add) + .set_name(unit_name + "add"); } - else if(middle_stride > 1) + else if (middle_stride > 1) { SubStream left(graph); - left << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 1, common_params.data_layout, PadStrideInfo(middle_stride, middle_stride, 0, 0), true)).set_name(unit_name + "shortcut/MaxPool"); + left << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 1, common_params.data_layout, + PadStrideInfo(middle_stride, middle_stride, 0, 0), true)) + .set_name(unit_name + "shortcut/MaxPool"); - graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(unit_name + "add"); + graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add) + .set_name(unit_name + "add"); } else { SubStream left(graph); - graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(unit_name + "add"); + graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add) + .set_name(unit_name + "add"); } - graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu"); + graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Relu"); } } }; diff --git a/examples/graph_resnet_v2_50.cpp b/examples/graph_resnet_v2_50.cpp index 6d5abb4f4b..48cf9b0b3c 100644 --- a/examples/graph_resnet_v2_50.cpp +++ b/examples/graph_resnet_v2_50.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -35,8 +36,7 @@ using namespace arm_compute::graph_utils; class GraphResNetV2_50Example : public Example { public: - GraphResNetV2_50Example() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ResNetV2_50") + GraphResNetV2_50Example() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ResNetV2_50") { } bool do_setup(int argc, char **argv) override @@ -49,7 +49,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -61,7 +61,7 @@ public: // Get trainable parameters data path std::string data_path = common_params.data_path; std::string model_path = "/cnn_data/resnet_v2_50_model/"; - if(!data_path.empty()) + if (!data_path.empty()) { data_path += model_path; } @@ -71,45 +71,42 @@ public: // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint - << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false /* Do not convert to BGR */)) - << ConvolutionLayer( - 7U, 7U, 64U, - get_weights_accessor(data_path, "conv1_weights.npy", weights_layout), - get_weights_accessor(data_path, "conv1_biases.npy", weights_layout), - PadStrideInfo(2, 2, 3, 3)) - .set_name("conv1/convolution") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))).set_name("pool1/MaxPool"); + graph << common_params.target << common_params.fast_math_hint + << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), + false /* Do not convert to BGR */)) + << ConvolutionLayer(7U, 7U, 64U, get_weights_accessor(data_path, "conv1_weights.npy", weights_layout), + get_weights_accessor(data_path, "conv1_biases.npy", weights_layout), + PadStrideInfo(2, 2, 3, 3)) + .set_name("conv1/convolution") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))) + .set_name("pool1/MaxPool"); add_residual_block(data_path, "block1", weights_layout, 64, 3, 2); add_residual_block(data_path, "block2", weights_layout, 128, 4, 2); add_residual_block(data_path, "block3", weights_layout, 256, 6, 2); add_residual_block(data_path, "block4", weights_layout, 512, 3, 1); - graph << BatchNormalizationLayer( - get_weights_accessor(data_path, "postnorm_moving_mean.npy"), - get_weights_accessor(data_path, "postnorm_moving_variance.npy"), - get_weights_accessor(data_path, "postnorm_gamma.npy"), - get_weights_accessor(data_path, "postnorm_beta.npy"), - 0.000009999999747378752f) - .set_name("postnorm/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("postnorm/Relu") + graph << BatchNormalizationLayer(get_weights_accessor(data_path, "postnorm_moving_mean.npy"), + get_weights_accessor(data_path, "postnorm_moving_variance.npy"), + get_weights_accessor(data_path, "postnorm_gamma.npy"), + get_weights_accessor(data_path, "postnorm_beta.npy"), 0.000009999999747378752f) + .set_name("postnorm/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("postnorm/Relu") << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("pool5") - << ConvolutionLayer( - 1U, 1U, 1001U, - get_weights_accessor(data_path, "logits_weights.npy", weights_layout), - get_weights_accessor(data_path, "logits_biases.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("logits/convolution") - << FlattenLayer().set_name("predictions/Reshape") - << SoftmaxLayer().set_name("predictions/Softmax") + << ConvolutionLayer(1U, 1U, 1001U, get_weights_accessor(data_path, "logits_weights.npy", weights_layout), + get_weights_accessor(data_path, "logits_biases.npy"), PadStrideInfo(1, 1, 0, 0)) + .set_name("logits/convolution") + << FlattenLayer().set_name("predictions/Reshape") << SoftmaxLayer().set_name("predictions/Softmax") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph @@ -139,10 +136,14 @@ private: CommonGraphParams common_params; Stream graph; - void add_residual_block(const std::string &data_path, const std::string &name, DataLayout weights_layout, - unsigned int base_depth, unsigned int num_units, unsigned int stride) + void add_residual_block(const std::string &data_path, + const std::string &name, + DataLayout weights_layout, + unsigned int base_depth, + unsigned int num_units, + unsigned int stride) { - for(unsigned int i = 0; i < num_units; ++i) + for (unsigned int i = 0; i < num_units; ++i) { // Generate unit names std::stringstream unit_path_ss; @@ -154,7 +155,8 @@ private: std::string unit_name = unit_name_ss.str(); const TensorShape last_shape = graph.graph().node(graph.tail_node())->output(0)->desc().shape; - unsigned int depth_in = last_shape[arm_compute::get_data_layout_dimension_index(common_params.data_layout, DataLayoutDimension::CHANNEL)]; + unsigned int depth_in = last_shape[arm_compute::get_data_layout_dimension_index( + common_params.data_layout, DataLayoutDimension::CHANNEL)]; unsigned int depth_out = base_depth * 4; // All units have stride 1 apart from last one @@ -162,73 +164,76 @@ private: // Preact SubStream preact(graph); - preact << BatchNormalizationLayer( - get_weights_accessor(data_path, unit_path + "preact_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "preact_moving_variance.npy"), - get_weights_accessor(data_path, unit_path + "preact_gamma.npy"), - get_weights_accessor(data_path, unit_path + "preact_beta.npy"), - 0.000009999999747378752f) - .set_name(unit_name + "preact/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "preact/Relu"); + preact << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "preact_moving_mean.npy"), + get_weights_accessor(data_path, unit_path + "preact_moving_variance.npy"), + get_weights_accessor(data_path, unit_path + "preact_gamma.npy"), + get_weights_accessor(data_path, unit_path + "preact_beta.npy"), + 0.000009999999747378752f) + .set_name(unit_name + "preact/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "preact/Relu"); // Create bottleneck path SubStream shortcut(graph); - if(depth_in == depth_out) + if (depth_in == depth_out) { - if(middle_stride != 1) + if (middle_stride != 1) { - shortcut << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 1, common_params.data_layout, PadStrideInfo(middle_stride, middle_stride, 0, 0), true)).set_name(unit_name + "shortcut/MaxPool"); + shortcut << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 1, common_params.data_layout, + PadStrideInfo(middle_stride, middle_stride, 0, 0), true)) + .set_name(unit_name + "shortcut/MaxPool"); } } else { shortcut.forward_tail(preact.tail_node()); shortcut << ConvolutionLayer( - 1U, 1U, depth_out, - get_weights_accessor(data_path, unit_path + "shortcut_weights.npy", weights_layout), - get_weights_accessor(data_path, unit_path + "shortcut_biases.npy", weights_layout), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "shortcut/convolution"); + 1U, 1U, depth_out, + get_weights_accessor(data_path, unit_path + "shortcut_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "shortcut_biases.npy", weights_layout), + PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "shortcut/convolution"); } // Create residual path SubStream residual(preact); - residual << ConvolutionLayer( - 1U, 1U, base_depth, - get_weights_accessor(data_path, unit_path + "conv1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "conv1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_beta.npy"), - 0.000009999999747378752f) - .set_name(unit_name + "conv1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu") - << ConvolutionLayer( - 3U, 3U, base_depth, - get_weights_accessor(data_path, unit_path + "conv2_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(middle_stride, middle_stride, 1, 1)) - .set_name(unit_name + "conv2/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_beta.npy"), - 0.000009999999747378752f) - .set_name(unit_name + "conv2/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu") - << ConvolutionLayer( - 1U, 1U, depth_out, - get_weights_accessor(data_path, unit_path + "conv3_weights.npy", weights_layout), - get_weights_accessor(data_path, unit_path + "conv3_biases.npy", weights_layout), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "conv3/convolution"); - - graph << EltwiseLayer(std::move(shortcut), std::move(residual), EltwiseOperation::Add).set_name(unit_name + "add"); + residual + << ConvolutionLayer(1U, 1U, base_depth, + get_weights_accessor(data_path, unit_path + "conv1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "conv1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_beta.npy"), + 0.000009999999747378752f) + .set_name(unit_name + "conv1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "conv1/Relu") + << ConvolutionLayer(3U, 3U, base_depth, + get_weights_accessor(data_path, unit_path + "conv2_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(middle_stride, middle_stride, 1, 1)) + .set_name(unit_name + "conv2/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_beta.npy"), + 0.000009999999747378752f) + .set_name(unit_name + "conv2/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "conv1/Relu") + << ConvolutionLayer(1U, 1U, depth_out, + get_weights_accessor(data_path, unit_path + "conv3_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "conv3_biases.npy", weights_layout), + PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "conv3/convolution"); + + graph << EltwiseLayer(std::move(shortcut), std::move(residual), EltwiseOperation::Add) + .set_name(unit_name + "add"); } } }; diff --git a/examples/graph_resnext50.cpp b/examples/graph_resnext50.cpp index 6378f6c741..12a1507c4c 100644 --- a/examples/graph_resnext50.cpp +++ b/examples/graph_resnext50.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -35,8 +36,7 @@ using namespace arm_compute::graph_utils; class GraphResNeXt50Example : public Example { public: - GraphResNeXt50Example() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ResNeXt50") + GraphResNeXt50Example() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ResNeXt50") { } bool do_setup(int argc, char **argv) override @@ -49,14 +49,15 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; } // Checks - ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph"); + ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), + "QASYMM8 not supported for this graph"); // Print parameter values std::cout << common_params << std::endl; @@ -66,28 +67,33 @@ public: // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint + graph << common_params.target << common_params.fast_math_hint << InputLayer(input_descriptor, get_input_accessor(common_params)) << ScaleLayer(get_weights_accessor(data_path, "/cnn_data/resnext50_model/bn_data_mul.npy"), get_weights_accessor(data_path, "/cnn_data/resnext50_model/bn_data_add.npy")) - .set_name("bn_data/Scale") + .set_name("bn_data/Scale") << ConvolutionLayer( - 7U, 7U, 64U, - get_weights_accessor(data_path, "/cnn_data/resnext50_model/conv0_weights.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/resnext50_model/conv0_biases.npy"), - PadStrideInfo(2, 2, 2, 3, 2, 3, DimensionRoundingType::FLOOR)) - .set_name("conv0/Convolution") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv0/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))).set_name("pool0"); - - add_residual_block(data_path, weights_layout, /*ofm*/ 256, /*stage*/ 1, /*num_unit*/ 3, /*stride_conv_unit1*/ 1); + 7U, 7U, 64U, + get_weights_accessor(data_path, "/cnn_data/resnext50_model/conv0_weights.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/resnext50_model/conv0_biases.npy"), + PadStrideInfo(2, 2, 2, 3, 2, 3, DimensionRoundingType::FLOOR)) + .set_name("conv0/Convolution") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv0/Relu") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))) + .set_name("pool0"); + + add_residual_block(data_path, weights_layout, /*ofm*/ 256, /*stage*/ 1, /*num_unit*/ 3, + /*stride_conv_unit1*/ 1); add_residual_block(data_path, weights_layout, 512, 2, 4, 2); add_residual_block(data_path, weights_layout, 1024, 3, 6, 2); add_residual_block(data_path, weights_layout, 2048, 4, 3, 2); @@ -121,10 +127,14 @@ private: CommonGraphParams common_params; Stream graph; - void add_residual_block(const std::string &data_path, DataLayout weights_layout, - unsigned int base_depth, unsigned int stage, unsigned int num_units, unsigned int stride_conv_unit1) + void add_residual_block(const std::string &data_path, + DataLayout weights_layout, + unsigned int base_depth, + unsigned int stage, + unsigned int num_units, + unsigned int stride_conv_unit1) { - for(unsigned int i = 0; i < num_units; ++i) + for (unsigned int i = 0; i < num_units; ++i) { std::stringstream unit_path_ss; unit_path_ss << "/cnn_data/resnext50_model/stage" << stage << "_unit" << (i + 1) << "_"; @@ -135,54 +145,55 @@ private: std::string unit_name = unit_name_ss.str(); PadStrideInfo pad_grouped_conv(1, 1, 1, 1); - if(i == 0) + if (i == 0) { - pad_grouped_conv = (stage == 1) ? PadStrideInfo(stride_conv_unit1, stride_conv_unit1, 1, 1) : PadStrideInfo(stride_conv_unit1, stride_conv_unit1, 0, 1, 0, 1, DimensionRoundingType::FLOOR); + pad_grouped_conv = (stage == 1) ? PadStrideInfo(stride_conv_unit1, stride_conv_unit1, 1, 1) + : PadStrideInfo(stride_conv_unit1, stride_conv_unit1, 0, 1, 0, 1, + DimensionRoundingType::FLOOR); } SubStream right(graph); - right << ConvolutionLayer( - 1U, 1U, base_depth / 2, - get_weights_accessor(data_path, unit_path + "conv1_weights.npy", weights_layout), - get_weights_accessor(data_path, unit_path + "conv1_biases.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "conv1/convolution") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu") - - << ConvolutionLayer( - 3U, 3U, base_depth / 2, - get_weights_accessor(data_path, unit_path + "conv2_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - pad_grouped_conv, 32) - .set_name(unit_name + "conv2/convolution") + right << ConvolutionLayer(1U, 1U, base_depth / 2, + get_weights_accessor(data_path, unit_path + "conv1_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "conv1_biases.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "conv1/convolution") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "conv1/Relu") + + << ConvolutionLayer(3U, 3U, base_depth / 2, + get_weights_accessor(data_path, unit_path + "conv2_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), pad_grouped_conv, + 32) + .set_name(unit_name + "conv2/convolution") << ScaleLayer(get_weights_accessor(data_path, unit_path + "bn2_mul.npy"), get_weights_accessor(data_path, unit_path + "bn2_add.npy")) - .set_name(unit_name + "conv1/Scale") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv2/Relu") + .set_name(unit_name + "conv1/Scale") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "conv2/Relu") - << ConvolutionLayer( - 1U, 1U, base_depth, - get_weights_accessor(data_path, unit_path + "conv3_weights.npy", weights_layout), - get_weights_accessor(data_path, unit_path + "conv3_biases.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "conv3/convolution"); + << ConvolutionLayer(1U, 1U, base_depth, + get_weights_accessor(data_path, unit_path + "conv3_weights.npy", weights_layout), + get_weights_accessor(data_path, unit_path + "conv3_biases.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "conv3/convolution"); SubStream left(graph); - if(i == 0) + if (i == 0) { - left << ConvolutionLayer( - 1U, 1U, base_depth, - get_weights_accessor(data_path, unit_path + "sc_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(stride_conv_unit1, stride_conv_unit1, 0, 0)) - .set_name(unit_name + "sc/convolution") + left << ConvolutionLayer(1U, 1U, base_depth, + get_weights_accessor(data_path, unit_path + "sc_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(stride_conv_unit1, stride_conv_unit1, 0, 0)) + .set_name(unit_name + "sc/convolution") << ScaleLayer(get_weights_accessor(data_path, unit_path + "sc_bn_mul.npy"), get_weights_accessor(data_path, unit_path + "sc_bn_add.npy")) - .set_name(unit_name + "sc/scale"); + .set_name(unit_name + "sc/scale"); } graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(unit_name + "add"); - graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu"); + graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Relu"); } } }; diff --git a/examples/graph_shufflenet.cpp b/examples/graph_shufflenet.cpp index 6e13c5eeb4..513d95884e 100644 --- a/examples/graph_shufflenet.cpp +++ b/examples/graph_shufflenet.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -35,8 +36,7 @@ using namespace arm_compute::graph_utils; class ShuffleNetExample : public Example { public: - ShuffleNetExample() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ShuffleNet") + ShuffleNetExample() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ShuffleNet") { } bool do_setup(int argc, char **argv) override @@ -49,20 +49,21 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; } // Set default layout if needed (Single kernel grouped convolution not yet supported int NHWC) - if(!common_opts.data_layout->is_set()) + if (!common_opts.data_layout->is_set()) { common_params.data_layout = DataLayout::NHWC; } // Checks - ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph"); + ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), + "QASYMM8 not supported for this graph"); // Print parameter values std::cout << common_params << std::endl; @@ -75,15 +76,17 @@ public: std::string data_path = common_params.data_path; // Add model path to data path - if(!data_path.empty()) + if (!data_path.empty()) { data_path += model_path; } // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; @@ -91,24 +94,22 @@ public: // Create preprocessor std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<TFPreproccessor>(0); - graph << common_params.target - << common_params.fast_math_hint - << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false /* Do not convert to BGR */)) - << ConvolutionLayer( - 3U, 3U, 24U, - get_weights_accessor(data_path, "conv3_0_w_0.npy", weights_layout), - get_weights_accessor(data_path, "conv3_0_b_0.npy", weights_layout), - PadStrideInfo(2, 2, 1, 1)) - .set_name("Conv1/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "conv3_0_bn_rm_0.npy"), - get_weights_accessor(data_path, "conv3_0_bn_riv_0.npy"), - get_weights_accessor(data_path, "conv3_0_bn_s_0.npy"), - get_weights_accessor(data_path, "conv3_0_bn_b_0.npy"), - 1e-5f) - .set_name("Conv1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv1/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 1, 1))).set_name("pool1/MaxPool"); + graph << common_params.target << common_params.fast_math_hint + << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), + false /* Do not convert to BGR */)) + << ConvolutionLayer(3U, 3U, 24U, get_weights_accessor(data_path, "conv3_0_w_0.npy", weights_layout), + get_weights_accessor(data_path, "conv3_0_b_0.npy", weights_layout), + PadStrideInfo(2, 2, 1, 1)) + .set_name("Conv1/convolution") + << BatchNormalizationLayer(get_weights_accessor(data_path, "conv3_0_bn_rm_0.npy"), + get_weights_accessor(data_path, "conv3_0_bn_riv_0.npy"), + get_weights_accessor(data_path, "conv3_0_bn_s_0.npy"), + get_weights_accessor(data_path, "conv3_0_bn_b_0.npy"), 1e-5f) + .set_name("Conv1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv1/Relu") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 1, 1))) + .set_name("pool1/MaxPool"); // Stage 2 add_residual_block(data_path, DataLayout::NCHW, 0U /* unit */, 112U /* depth */, 2U /* stride */); @@ -134,13 +135,10 @@ public: graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("predictions/AvgPool") << FlattenLayer().set_name("predictions/Reshape") - << FullyConnectedLayer( - 1000U, - get_weights_accessor(data_path, "pred_w_0.npy", weights_layout), - get_weights_accessor(data_path, "pred_b_0.npy")) - .set_name("predictions/FC") - << SoftmaxLayer().set_name("predictions/Softmax") - << OutputLayer(get_output_accessor(common_params, 5)); + << FullyConnectedLayer(1000U, get_weights_accessor(data_path, "pred_w_0.npy", weights_layout), + get_weights_accessor(data_path, "pred_b_0.npy")) + .set_name("predictions/FC") + << SoftmaxLayer().set_name("predictions/Softmax") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph GraphConfig config; @@ -167,8 +165,11 @@ private: CommonGraphParams common_params; Stream graph; - void add_residual_block(const std::string &data_path, DataLayout weights_layout, - unsigned int unit, unsigned int depth, unsigned int stride) + void add_residual_block(const std::string &data_path, + DataLayout weights_layout, + unsigned int unit, + unsigned int depth, + unsigned int stride) { PadStrideInfo dwc_info = PadStrideInfo(1, 1, 1, 1); const unsigned int gconv_id = unit * 2; @@ -181,63 +182,61 @@ private: SubStream left_ss(graph); SubStream right_ss(graph); - if(stride == 2) + if (stride == 2) { - right_ss << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, PadStrideInfo(2, 2, 1, 1))).set_name(unit_name + "/pool_1/AveragePool"); + right_ss << PoolingLayer( + PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, PadStrideInfo(2, 2, 1, 1))) + .set_name(unit_name + "/pool_1/AveragePool"); dwc_info = PadStrideInfo(2, 2, 1, 1); } - left_ss << ConvolutionLayer( - 1U, 1U, depth, - get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_w_0.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0), num_groups) - .set_name(unit_name + "/gconv1_" + gconv_id_name + "/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_rm_0.npy"), - get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_riv_0.npy"), - get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_s_0.npy"), - get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_b_0.npy"), - 1e-5f) - .set_name(unit_name + "/gconv1_" + gconv_id_name + "/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "/gconv1_" + gconv_id_name + "/Relu") - << ChannelShuffleLayer(num_groups).set_name(unit_name + "/shuffle_0/ChannelShufle") - << DepthwiseConvolutionLayer( - 3U, 3U, - get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_w_0.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - dwc_info) - .set_name(unit_name + "/gconv3_" + unit_id_name + "/depthwise") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_rm_0.npy"), - get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_riv_0.npy"), - get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_s_0.npy"), - get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_b_0.npy"), - 1e-5f) - .set_name(unit_name + "/gconv3_" + unit_id_name + "/BatchNorm") - << ConvolutionLayer( - 1U, 1U, depth, - get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_w_0.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0), num_groups) - .set_name(unit_name + "/gconv1_" + gconv_id_1_name + "/convolution") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_rm_0.npy"), - get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_riv_0.npy"), - get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_s_0.npy"), - get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_b_0.npy"), - 1e-5f) - .set_name(unit_name + "/gconv1_" + gconv_id_1_name + "/BatchNorm"); + left_ss + << ConvolutionLayer(1U, 1U, depth, + get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_w_0.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(1, 1, 0, 0), num_groups) + .set_name(unit_name + "/gconv1_" + gconv_id_name + "/convolution") + << BatchNormalizationLayer(get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_rm_0.npy"), + get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_riv_0.npy"), + get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_s_0.npy"), + get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_b_0.npy"), + 1e-5f) + .set_name(unit_name + "/gconv1_" + gconv_id_name + "/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "/gconv1_" + gconv_id_name + "/Relu") + << ChannelShuffleLayer(num_groups).set_name(unit_name + "/shuffle_0/ChannelShufle") + << DepthwiseConvolutionLayer( + 3U, 3U, get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_w_0.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), dwc_info) + .set_name(unit_name + "/gconv3_" + unit_id_name + "/depthwise") + << BatchNormalizationLayer(get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_rm_0.npy"), + get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_riv_0.npy"), + get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_s_0.npy"), + get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_b_0.npy"), 1e-5f) + .set_name(unit_name + "/gconv3_" + unit_id_name + "/BatchNorm") + << ConvolutionLayer( + 1U, 1U, depth, + get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_w_0.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0), num_groups) + .set_name(unit_name + "/gconv1_" + gconv_id_1_name + "/convolution") + << BatchNormalizationLayer(get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_rm_0.npy"), + get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_riv_0.npy"), + get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_s_0.npy"), + get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_b_0.npy"), + 1e-5f) + .set_name(unit_name + "/gconv1_" + gconv_id_1_name + "/BatchNorm"); - if(stride == 2) + if (stride == 2) { graph << ConcatLayer(std::move(left_ss), std::move(right_ss)).set_name(unit_name + "/Concat"); } else { - graph << EltwiseLayer(std::move(left_ss), std::move(right_ss), EltwiseOperation::Add).set_name(unit_name + "/Add"); + graph << EltwiseLayer(std::move(left_ss), std::move(right_ss), EltwiseOperation::Add) + .set_name(unit_name + "/Add"); } - graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "/Relu"); + graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "/Relu"); } }; diff --git a/examples/graph_squeezenet.cpp b/examples/graph_squeezenet.cpp index 3ea2fea38f..7d0528f805 100644 --- a/examples/graph_squeezenet.cpp +++ b/examples/graph_squeezenet.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -35,8 +36,7 @@ using namespace arm_compute::graph_utils; class GraphSqueezenetExample : public Example { public: - GraphSqueezenetExample() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "SqueezeNetV1") + GraphSqueezenetExample() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "SqueezeNetV1") { } bool do_setup(int argc, char **argv) override @@ -49,7 +49,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -62,104 +62,128 @@ public: std::string data_path = common_params.data_path; // Create a preprocessor object - const std::array<float, 3> mean_rgb{ { 122.68f, 116.67f, 104.01f } }; + const std::array<float, 3> mean_rgb{{122.68f, 116.67f, 104.01f}}; std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<CaffePreproccessor>(mean_rgb); // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint + graph << common_params.target << common_params.fast_math_hint << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor))) << ConvolutionLayer( - 7U, 7U, 96U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_b.npy"), - PadStrideInfo(2, 2, 0, 0)) - .set_name("conv1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu_conv1") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool1") + 7U, 7U, 96U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_b.npy"), + PadStrideInfo(2, 2, 0, 0)) + .set_name("conv1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("relu_conv1") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool1") << ConvolutionLayer( - 1U, 1U, 16U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire2/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire2/relu_squeeze1x1"); + 1U, 1U, 16U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire2/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire2/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire2", weights_layout, 64U, 64U).set_name("fire2/concat"); graph << ConvolutionLayer( - 1U, 1U, 16U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire3/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire3/relu_squeeze1x1"); + 1U, 1U, 16U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire3/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire3/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire3", weights_layout, 64U, 64U).set_name("fire3/concat"); graph << ConvolutionLayer( - 1U, 1U, 32U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire4/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire4/relu_squeeze1x1"); + 1U, 1U, 32U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire4/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire4/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire4", weights_layout, 128U, 128U).set_name("fire4/concat"); - graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool4") + graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool4") << ConvolutionLayer( - 1U, 1U, 32U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire5/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire5/relu_squeeze1x1"); + 1U, 1U, 32U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire5/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire5/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire5", weights_layout, 128U, 128U).set_name("fire5/concat"); graph << ConvolutionLayer( - 1U, 1U, 48U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire6/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire6/relu_squeeze1x1"); + 1U, 1U, 48U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire6/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire6/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire6", weights_layout, 192U, 192U).set_name("fire6/concat"); graph << ConvolutionLayer( - 1U, 1U, 48U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire7/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire7/relu_squeeze1x1"); + 1U, 1U, 48U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire7/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire7/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire7", weights_layout, 192U, 192U).set_name("fire7/concat"); graph << ConvolutionLayer( - 1U, 1U, 64U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire8/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire8/relu_squeeze1x1"); + 1U, 1U, 64U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire8/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire8/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire8", weights_layout, 256U, 256U).set_name("fire8/concat"); - graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool8") + graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool8") << ConvolutionLayer( - 1U, 1U, 64U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire9/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire9/relu_squeeze1x1"); + 1U, 1U, 64U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire9/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire9/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire9", weights_layout, 256U, 256U).set_name("fire9/concat"); graph << ConvolutionLayer( - 1U, 1U, 1000U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv10") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu_conv10") + 1U, 1U, 1000U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("conv10") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("relu_conv10") << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("pool10") - << FlattenLayer().set_name("flatten") - << SoftmaxLayer().set_name("prob") + << FlattenLayer().set_name("flatten") << SoftmaxLayer().set_name("prob") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph @@ -188,27 +212,30 @@ private: CommonGraphParams common_params; Stream graph; - ConcatLayer get_expand_fire_node(const std::string &data_path, std::string &¶m_path, DataLayout weights_layout, - unsigned int expand1_filt, unsigned int expand3_filt) + ConcatLayer get_expand_fire_node(const std::string &data_path, + std::string &¶m_path, + DataLayout weights_layout, + unsigned int expand1_filt, + unsigned int expand3_filt) { std::string total_path = "/cnn_data/squeezenet_v1.0_model/" + param_path + "_"; SubStream i_a(graph); - i_a << ConvolutionLayer( - 1U, 1U, expand1_filt, - get_weights_accessor(data_path, total_path + "expand1x1_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "expand1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/expand1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_expand1x1"); + i_a << ConvolutionLayer(1U, 1U, expand1_filt, + get_weights_accessor(data_path, total_path + "expand1x1_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "expand1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/expand1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_expand1x1"); SubStream i_b(graph); - i_b << ConvolutionLayer( - 3U, 3U, expand3_filt, - get_weights_accessor(data_path, total_path + "expand3x3_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "expand3x3_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name(param_path + "/expand3x3") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_expand3x3"); + i_b << ConvolutionLayer(3U, 3U, expand3_filt, + get_weights_accessor(data_path, total_path + "expand3x3_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "expand3x3_b.npy"), + PadStrideInfo(1, 1, 1, 1)) + .set_name(param_path + "/expand3x3") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_expand3x3"); return ConcatLayer(std::move(i_a), std::move(i_b)); } diff --git a/examples/graph_squeezenet_v1_1.cpp b/examples/graph_squeezenet_v1_1.cpp index 9cc183fbbd..ed0f692db2 100644 --- a/examples/graph_squeezenet_v1_1.cpp +++ b/examples/graph_squeezenet_v1_1.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -35,8 +36,7 @@ using namespace arm_compute::graph_utils; class GraphSqueezenet_v1_1Example : public Example { public: - GraphSqueezenet_v1_1Example() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "SqueezeNetV1.1") + GraphSqueezenet_v1_1Example() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "SqueezeNetV1.1") { } bool do_setup(int argc, char **argv) override @@ -49,7 +49,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -62,104 +62,128 @@ public: std::string data_path = common_params.data_path; // Create a preprocessor object - const std::array<float, 3> mean_rgb{ { 122.68f, 116.67f, 104.01f } }; + const std::array<float, 3> mean_rgb{{122.68f, 116.67f, 104.01f}}; std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<CaffePreproccessor>(mean_rgb); // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(227U, 227U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(227U, 227U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint + graph << common_params.target << common_params.fast_math_hint << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor))) << ConvolutionLayer( - 3U, 3U, 64U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv1_b.npy"), - PadStrideInfo(2, 2, 0, 0)) - .set_name("conv1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu_conv1") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool1") + 3U, 3U, 64U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv1_b.npy"), + PadStrideInfo(2, 2, 0, 0)) + .set_name("conv1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("relu_conv1") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool1") << ConvolutionLayer( - 1U, 1U, 16U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire2_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire2_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire2/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire2/relu_squeeze1x1"); + 1U, 1U, 16U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire2_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire2_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire2/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire2/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire2", weights_layout, 64U, 64U).set_name("fire2/concat"); graph << ConvolutionLayer( - 1U, 1U, 16U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire3_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire3_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire3/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire3/relu_squeeze1x1"); + 1U, 1U, 16U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire3_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire3_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire3/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire3/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire3", weights_layout, 64U, 64U).set_name("fire3/concat"); - graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool3") + graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool3") << ConvolutionLayer( - 1U, 1U, 32U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire4_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire4_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire4/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire4/relu_squeeze1x1"); + 1U, 1U, 32U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire4_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire4_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire4/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire4/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire4", weights_layout, 128U, 128U).set_name("fire4/concat"); graph << ConvolutionLayer( - 1U, 1U, 32U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire5_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire5_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire5/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire5/relu_squeeze1x1"); + 1U, 1U, 32U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire5_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire5_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire5/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire5/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire5", weights_layout, 128U, 128U).set_name("fire5/concat"); - graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool5") + graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool5") << ConvolutionLayer( - 1U, 1U, 48U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire6_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire6_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire6/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire6/relu_squeeze1x1"); + 1U, 1U, 48U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire6_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire6_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire6/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire6/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire6", weights_layout, 192U, 192U).set_name("fire6/concat"); graph << ConvolutionLayer( - 1U, 1U, 48U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire7_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire7_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire7/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire7/relu_squeeze1x1"); + 1U, 1U, 48U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire7_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire7_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire7/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire7/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire7", weights_layout, 192U, 192U).set_name("fire7/concat"); graph << ConvolutionLayer( - 1U, 1U, 64U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire8_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire8_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire8/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire8/relu_squeeze1x1"); + 1U, 1U, 64U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire8_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire8_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire8/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire8/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire8", weights_layout, 256U, 256U).set_name("fire8/concat"); graph << ConvolutionLayer( - 1U, 1U, 64U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire9_squeeze1x1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire9_squeeze1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("fire9/squeeze1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire9/relu_squeeze1x1"); + 1U, 1U, 64U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire9_squeeze1x1_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire9_squeeze1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("fire9/squeeze1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("fire9/relu_squeeze1x1"); graph << get_expand_fire_node(data_path, "fire9", weights_layout, 256U, 256U).set_name("fire9/concat"); graph << ConvolutionLayer( - 1U, 1U, 1000U, - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv10_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv10_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv10") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu_conv10") + 1U, 1U, 1000U, + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv10_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv10_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("conv10") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("relu_conv10") << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("pool10") - << FlattenLayer().set_name("flatten") - << SoftmaxLayer().set_name("prob") + << FlattenLayer().set_name("flatten") << SoftmaxLayer().set_name("prob") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph @@ -188,27 +212,30 @@ private: CommonGraphParams common_params; Stream graph; - ConcatLayer get_expand_fire_node(const std::string &data_path, std::string &¶m_path, DataLayout weights_layout, - unsigned int expand1_filt, unsigned int expand3_filt) + ConcatLayer get_expand_fire_node(const std::string &data_path, + std::string &¶m_path, + DataLayout weights_layout, + unsigned int expand1_filt, + unsigned int expand3_filt) { std::string total_path = "/cnn_data/squeezenet_v1_1_model/" + param_path + "_"; SubStream i_a(graph); - i_a << ConvolutionLayer( - 1U, 1U, expand1_filt, - get_weights_accessor(data_path, total_path + "expand1x1_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "expand1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/expand1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_expand1x1"); + i_a << ConvolutionLayer(1U, 1U, expand1_filt, + get_weights_accessor(data_path, total_path + "expand1x1_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "expand1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/expand1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_expand1x1"); SubStream i_b(graph); - i_b << ConvolutionLayer( - 3U, 3U, expand3_filt, - get_weights_accessor(data_path, total_path + "expand3x3_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "expand3x3_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name(param_path + "/expand3x3") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_expand3x3"); + i_b << ConvolutionLayer(3U, 3U, expand3_filt, + get_weights_accessor(data_path, total_path + "expand3x3_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "expand3x3_b.npy"), + PadStrideInfo(1, 1, 1, 1)) + .set_name(param_path + "/expand3x3") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_expand3x3"); return ConcatLayer(std::move(i_a), std::move(i_b)); } diff --git a/examples/graph_srcnn955.cpp b/examples/graph_srcnn955.cpp index 855bbd848e..15a8b5d8ec 100644 --- a/examples/graph_srcnn955.cpp +++ b/examples/graph_srcnn955.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -36,7 +37,12 @@ class GraphSRCNN955Example : public Example { public: GraphSRCNN955Example() - : cmd_parser(), common_opts(cmd_parser), model_input_width(nullptr), model_input_height(nullptr), common_params(), graph(0, "SRCNN955") + : cmd_parser(), + common_opts(cmd_parser), + model_input_width(nullptr), + model_input_height(nullptr), + common_params(), + graph(0, "SRCNN955") { model_input_width = cmd_parser.add_option<SimpleOption<unsigned int>>("image-width", 300); model_input_height = cmd_parser.add_option<SimpleOption<unsigned int>>("image-height", 300); @@ -45,7 +51,7 @@ public: model_input_width->set_help("Input image width."); model_input_height->set_help("Input image height."); } - GraphSRCNN955Example(const GraphSRCNN955Example &) = delete; + GraphSRCNN955Example(const GraphSRCNN955Example &) = delete; GraphSRCNN955Example &operator=(const GraphSRCNN955Example &) = delete; ~GraphSRCNN955Example() override = default; bool do_setup(int argc, char **argv) override @@ -58,7 +64,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -81,36 +87,33 @@ public: std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<TFPreproccessor>(); // Create input descriptor - const TensorShape tensor_shape = permute_shape(TensorShape(image_width, image_height, 3U, common_params.batches), DataLayout::NCHW, common_params.data_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(image_width, image_height, 3U, common_params.batches), DataLayout::NCHW, + common_params.data_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint - << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false /* Do not convert to BGR */)) - << ConvolutionLayer( - 9U, 9U, 64U, - get_weights_accessor(data_path, "conv1_weights.npy", weights_layout), - get_weights_accessor(data_path, "conv1_biases.npy"), - PadStrideInfo(1, 1, 4, 4)) - .set_name("conv1/convolution") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1/Relu") - << ConvolutionLayer( - 5U, 5U, 32U, - get_weights_accessor(data_path, "conv2_weights.npy", weights_layout), - get_weights_accessor(data_path, "conv2_biases.npy"), - PadStrideInfo(1, 1, 2, 2)) - .set_name("conv2/convolution") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2/Relu") - << ConvolutionLayer( - 5U, 5U, 3U, - get_weights_accessor(data_path, "conv3_weights.npy", weights_layout), - get_weights_accessor(data_path, "conv3_biases.npy"), - PadStrideInfo(1, 1, 2, 2)) - .set_name("conv3/convolution") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3/Relu") + graph << common_params.target << common_params.fast_math_hint + << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), + false /* Do not convert to BGR */)) + << ConvolutionLayer(9U, 9U, 64U, get_weights_accessor(data_path, "conv1_weights.npy", weights_layout), + get_weights_accessor(data_path, "conv1_biases.npy"), PadStrideInfo(1, 1, 4, 4)) + .set_name("conv1/convolution") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv1/Relu") + << ConvolutionLayer(5U, 5U, 32U, get_weights_accessor(data_path, "conv2_weights.npy", weights_layout), + get_weights_accessor(data_path, "conv2_biases.npy"), PadStrideInfo(1, 1, 2, 2)) + .set_name("conv2/convolution") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv2/Relu") + << ConvolutionLayer(5U, 5U, 3U, get_weights_accessor(data_path, "conv3_weights.npy", weights_layout), + get_weights_accessor(data_path, "conv3_biases.npy"), PadStrideInfo(1, 1, 2, 2)) + .set_name("conv3/convolution") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv3/Relu") << OutputLayer(std::make_unique<DummyAccessor>(0)); // Finalize graph @@ -137,8 +140,8 @@ public: private: CommandLineParser cmd_parser; CommonGraphOptions common_opts; - SimpleOption<unsigned int> *model_input_width{ nullptr }; - SimpleOption<unsigned int> *model_input_height{ nullptr }; + SimpleOption<unsigned int> *model_input_width{nullptr}; + SimpleOption<unsigned int> *model_input_height{nullptr}; CommonGraphParams common_params; Stream graph; }; diff --git a/examples/graph_ssd_mobilenet.cpp b/examples/graph_ssd_mobilenet.cpp index 9fe7f5b454..5162fe6890 100644 --- a/examples/graph_ssd_mobilenet.cpp +++ b/examples/graph_ssd_mobilenet.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -36,23 +37,26 @@ using namespace arm_compute::graph_utils; class GraphSSDMobilenetExample : public Example { public: - GraphSSDMobilenetExample() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "MobileNetSSD") + GraphSSDMobilenetExample() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "MobileNetSSD") { // Add topk option keep_topk_opt = cmd_parser.add_option<SimpleOption<int>>("topk", 100); keep_topk_opt->set_help("Top k detections results per image. Used for data type F32."); // Add output option detection_boxes_opt = cmd_parser.add_option<SimpleOption<std::string>>("detection_boxes_opt", ""); - detection_boxes_opt->set_help("Filename containing the reference values for the graph output detection_boxes. Used for data type QASYMM8."); + detection_boxes_opt->set_help("Filename containing the reference values for the graph output detection_boxes. " + "Used for data type QASYMM8."); detection_classes_opt = cmd_parser.add_option<SimpleOption<std::string>>("detection_classes_opt", ""); - detection_classes_opt->set_help("Filename containing the reference values for the output detection_classes. Used for data type QASYMM8."); + detection_classes_opt->set_help( + "Filename containing the reference values for the output detection_classes. Used for data type QASYMM8."); detection_scores_opt = cmd_parser.add_option<SimpleOption<std::string>>("detection_scores_opt", ""); - detection_scores_opt->set_help("Filename containing the reference values for the output detection_scores. Used for data type QASYMM8."); + detection_scores_opt->set_help( + "Filename containing the reference values for the output detection_scores. Used for data type QASYMM8."); num_detections_opt = cmd_parser.add_option<SimpleOption<std::string>>("num_detections_opt", ""); - num_detections_opt->set_help("Filename containing the reference values for the output num_detections. Used with datatype QASYMM8."); + num_detections_opt->set_help( + "Filename containing the reference values for the output num_detections. Used with datatype QASYMM8."); } - GraphSSDMobilenetExample(const GraphSSDMobilenetExample &) = delete; + GraphSSDMobilenetExample(const GraphSSDMobilenetExample &) = delete; GraphSSDMobilenetExample &operator=(const GraphSSDMobilenetExample &) = delete; ~GraphSSDMobilenetExample() override = default; bool do_setup(int argc, char **argv) override @@ -65,7 +69,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -75,15 +79,16 @@ public: std::cout << common_params << std::endl; // Create input descriptor - const TensorShape tensor_shape = permute_shape(TensorShape(300, 300, 3U, 1U), DataLayout::NCHW, common_params.data_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(300, 300, 3U, 1U), DataLayout::NCHW, common_params.data_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); // Set graph hints - graph << common_params.target - << common_params.fast_math_hint; + graph << common_params.target << common_params.fast_math_hint; // Create core graph - if(arm_compute::is_data_type_float(common_params.data_type)) + if (arm_compute::is_data_type_float(common_params.data_type)) { create_graph_float(input_descriptor); } @@ -112,99 +117,98 @@ public: private: CommandLineParser cmd_parser; CommonGraphOptions common_opts; - SimpleOption<int> *keep_topk_opt{ nullptr }; + SimpleOption<int> *keep_topk_opt{nullptr}; CommonGraphParams common_params; Stream graph; - SimpleOption<std::string> *detection_boxes_opt{ nullptr }; - SimpleOption<std::string> *detection_classes_opt{ nullptr }; - SimpleOption<std::string> *detection_scores_opt{ nullptr }; - SimpleOption<std::string> *num_detections_opt{ nullptr }; - - ConcatLayer get_node_A_float(IStream &main_graph, const std::string &data_path, std::string &¶m_path, - unsigned int conv_filt, - PadStrideInfo dwc_pad_stride_info, PadStrideInfo conv_pad_stride_info) + SimpleOption<std::string> *detection_boxes_opt{nullptr}; + SimpleOption<std::string> *detection_classes_opt{nullptr}; + SimpleOption<std::string> *detection_scores_opt{nullptr}; + SimpleOption<std::string> *num_detections_opt{nullptr}; + + ConcatLayer get_node_A_float(IStream &main_graph, + const std::string &data_path, + std::string &¶m_path, + unsigned int conv_filt, + PadStrideInfo dwc_pad_stride_info, + PadStrideInfo conv_pad_stride_info) { const std::string total_path = param_path + "_"; SubStream sg(main_graph); - sg << DepthwiseConvolutionLayer( - 3U, 3U, - get_weights_accessor(data_path, total_path + "dw_w.npy"), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - dwc_pad_stride_info) - .set_name(param_path + "/dw") + sg << DepthwiseConvolutionLayer(3U, 3U, get_weights_accessor(data_path, total_path + "dw_w.npy"), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + dwc_pad_stride_info) + .set_name(param_path + "/dw") << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "dw_bn_mean.npy"), get_weights_accessor(data_path, total_path + "dw_bn_var.npy"), get_weights_accessor(data_path, total_path + "dw_scale_w.npy"), get_weights_accessor(data_path, total_path + "dw_scale_b.npy"), 0.00001f) - .set_name(param_path + "/dw/bn") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "dw/relu") - - << ConvolutionLayer( - 1U, 1U, conv_filt, - get_weights_accessor(data_path, total_path + "w.npy"), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - conv_pad_stride_info) - .set_name(param_path + "/pw") + .set_name(param_path + "/dw/bn") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "dw/relu") + + << ConvolutionLayer(1U, 1U, conv_filt, get_weights_accessor(data_path, total_path + "w.npy"), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), conv_pad_stride_info) + .set_name(param_path + "/pw") << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "bn_mean.npy"), get_weights_accessor(data_path, total_path + "bn_var.npy"), get_weights_accessor(data_path, total_path + "scale_w.npy"), get_weights_accessor(data_path, total_path + "scale_b.npy"), 0.00001f) - .set_name(param_path + "/pw/bn") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "pw/relu"); + .set_name(param_path + "/pw/bn") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "pw/relu"); return ConcatLayer(std::move(sg)); } - ConcatLayer get_node_B_float(IStream &main_graph, const std::string &data_path, std::string &¶m_path, - unsigned int conv_filt, - PadStrideInfo conv_pad_stride_info_1, PadStrideInfo conv_pad_stride_info_2) + ConcatLayer get_node_B_float(IStream &main_graph, + const std::string &data_path, + std::string &¶m_path, + unsigned int conv_filt, + PadStrideInfo conv_pad_stride_info_1, + PadStrideInfo conv_pad_stride_info_2) { const std::string total_path = param_path + "_"; SubStream sg(main_graph); - sg << ConvolutionLayer( - 1, 1, conv_filt / 2, - get_weights_accessor(data_path, total_path + "1_w.npy"), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - conv_pad_stride_info_1) - .set_name(total_path + "1/conv") + sg << ConvolutionLayer(1, 1, conv_filt / 2, get_weights_accessor(data_path, total_path + "1_w.npy"), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), conv_pad_stride_info_1) + .set_name(total_path + "1/conv") << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "1_bn_mean.npy"), get_weights_accessor(data_path, total_path + "1_bn_var.npy"), get_weights_accessor(data_path, total_path + "1_scale_w.npy"), get_weights_accessor(data_path, total_path + "1_scale_b.npy"), 0.00001f) - .set_name(total_path + "1/bn") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(total_path + "1/relu"); - - sg << ConvolutionLayer( - 3, 3, conv_filt, - get_weights_accessor(data_path, total_path + "2_w.npy"), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - conv_pad_stride_info_2) - .set_name(total_path + "2/conv") + .set_name(total_path + "1/bn") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(total_path + "1/relu"); + + sg << ConvolutionLayer(3, 3, conv_filt, get_weights_accessor(data_path, total_path + "2_w.npy"), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), conv_pad_stride_info_2) + .set_name(total_path + "2/conv") << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "2_bn_mean.npy"), get_weights_accessor(data_path, total_path + "2_bn_var.npy"), get_weights_accessor(data_path, total_path + "2_scale_w.npy"), get_weights_accessor(data_path, total_path + "2_scale_b.npy"), 0.00001f) - .set_name(total_path + "2/bn") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(total_path + "2/relu"); + .set_name(total_path + "2/bn") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(total_path + "2/relu"); return ConcatLayer(std::move(sg)); } - ConcatLayer get_node_C_float(IStream &main_graph, const std::string &data_path, std::string &¶m_path, - unsigned int conv_filt, PadStrideInfo conv_pad_stride_info) + ConcatLayer get_node_C_float(IStream &main_graph, + const std::string &data_path, + std::string &¶m_path, + unsigned int conv_filt, + PadStrideInfo conv_pad_stride_info) { const std::string total_path = param_path + "_"; SubStream sg(main_graph); - sg << ConvolutionLayer( - 1U, 1U, conv_filt, - get_weights_accessor(data_path, total_path + "w.npy"), - get_weights_accessor(data_path, total_path + "b.npy"), - conv_pad_stride_info) - .set_name(param_path + "/conv"); - if(common_params.data_layout == DataLayout::NCHW) + sg << ConvolutionLayer(1U, 1U, conv_filt, get_weights_accessor(data_path, total_path + "w.npy"), + get_weights_accessor(data_path, total_path + "b.npy"), conv_pad_stride_info) + .set_name(param_path + "/conv"); + if (common_params.data_layout == DataLayout::NCHW) { sg << PermuteLayer(PermutationVector(2U, 0U, 1U), DataLayout::NHWC).set_name(param_path + "/perm"); } @@ -216,62 +220,77 @@ private: void create_graph_float(TensorDescriptor &input_descriptor) { // Create a preprocessor object - const std::array<float, 3> mean_rgb{ { 127.5f, 127.5f, 127.5f } }; + const std::array<float, 3> mean_rgb{{127.5f, 127.5f, 127.5f}}; std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<CaffePreproccessor>(mean_rgb, true, 0.007843f); // Get trainable parameters data path std::string data_path = common_params.data_path; // Add model path to data path - if(!data_path.empty()) + if (!data_path.empty()) { data_path += "/cnn_data/ssd_mobilenet_model/"; } - graph << InputLayer(input_descriptor, - get_input_accessor(common_params, std::move(preprocessor))); + graph << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor))); SubStream conv_11(graph); - conv_11 << ConvolutionLayer( - 3U, 3U, 32U, - get_weights_accessor(data_path, "conv0_w.npy"), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 1, 1)) - .set_name("conv0"); + conv_11 << ConvolutionLayer(3U, 3U, 32U, get_weights_accessor(data_path, "conv0_w.npy"), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(2, 2, 1, 1)) + .set_name("conv0"); conv_11 << BatchNormalizationLayer(get_weights_accessor(data_path, "conv0_bn_mean.npy"), get_weights_accessor(data_path, "conv0_bn_var.npy"), get_weights_accessor(data_path, "conv0_scale_w.npy"), get_weights_accessor(data_path, "conv0_scale_b.npy"), 0.00001f) - .set_name("conv0/bn") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv0/relu"); - - conv_11 << get_node_A_float(conv_11, data_path, "conv1", 64, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0)); - conv_11 << get_node_A_float(conv_11, data_path, "conv2", 128, PadStrideInfo(2, 2, 1, 1), PadStrideInfo(1, 1, 0, 0)); - conv_11 << get_node_A_float(conv_11, data_path, "conv3", 128, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0)); - conv_11 << get_node_A_float(conv_11, data_path, "conv4", 256, PadStrideInfo(2, 2, 1, 1), PadStrideInfo(1, 1, 0, 0)); - conv_11 << get_node_A_float(conv_11, data_path, "conv5", 256, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0)); - conv_11 << get_node_A_float(conv_11, data_path, "conv6", 512, PadStrideInfo(2, 2, 1, 1), PadStrideInfo(1, 1, 0, 0)); - conv_11 << get_node_A_float(conv_11, data_path, "conv7", 512, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0)); - conv_11 << get_node_A_float(conv_11, data_path, "conv8", 512, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0)); - conv_11 << get_node_A_float(conv_11, data_path, "conv9", 512, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0)); - conv_11 << get_node_A_float(conv_11, data_path, "conv10", 512, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0)); - conv_11 << get_node_A_float(conv_11, data_path, "conv11", 512, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0)); + .set_name("conv0/bn") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv0/relu"); + + conv_11 << get_node_A_float(conv_11, data_path, "conv1", 64, PadStrideInfo(1, 1, 1, 1), + PadStrideInfo(1, 1, 0, 0)); + conv_11 << get_node_A_float(conv_11, data_path, "conv2", 128, PadStrideInfo(2, 2, 1, 1), + PadStrideInfo(1, 1, 0, 0)); + conv_11 << get_node_A_float(conv_11, data_path, "conv3", 128, PadStrideInfo(1, 1, 1, 1), + PadStrideInfo(1, 1, 0, 0)); + conv_11 << get_node_A_float(conv_11, data_path, "conv4", 256, PadStrideInfo(2, 2, 1, 1), + PadStrideInfo(1, 1, 0, 0)); + conv_11 << get_node_A_float(conv_11, data_path, "conv5", 256, PadStrideInfo(1, 1, 1, 1), + PadStrideInfo(1, 1, 0, 0)); + conv_11 << get_node_A_float(conv_11, data_path, "conv6", 512, PadStrideInfo(2, 2, 1, 1), + PadStrideInfo(1, 1, 0, 0)); + conv_11 << get_node_A_float(conv_11, data_path, "conv7", 512, PadStrideInfo(1, 1, 1, 1), + PadStrideInfo(1, 1, 0, 0)); + conv_11 << get_node_A_float(conv_11, data_path, "conv8", 512, PadStrideInfo(1, 1, 1, 1), + PadStrideInfo(1, 1, 0, 0)); + conv_11 << get_node_A_float(conv_11, data_path, "conv9", 512, PadStrideInfo(1, 1, 1, 1), + PadStrideInfo(1, 1, 0, 0)); + conv_11 << get_node_A_float(conv_11, data_path, "conv10", 512, PadStrideInfo(1, 1, 1, 1), + PadStrideInfo(1, 1, 0, 0)); + conv_11 << get_node_A_float(conv_11, data_path, "conv11", 512, PadStrideInfo(1, 1, 1, 1), + PadStrideInfo(1, 1, 0, 0)); SubStream conv_13(conv_11); - conv_13 << get_node_A_float(conv_11, data_path, "conv12", 1024, PadStrideInfo(2, 2, 1, 1), PadStrideInfo(1, 1, 0, 0)); - conv_13 << get_node_A_float(conv_13, data_path, "conv13", 1024, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0)); + conv_13 << get_node_A_float(conv_11, data_path, "conv12", 1024, PadStrideInfo(2, 2, 1, 1), + PadStrideInfo(1, 1, 0, 0)); + conv_13 << get_node_A_float(conv_13, data_path, "conv13", 1024, PadStrideInfo(1, 1, 1, 1), + PadStrideInfo(1, 1, 0, 0)); SubStream conv_14(conv_13); - conv_14 << get_node_B_float(conv_13, data_path, "conv14", 512, PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 2, 1, 1)); + conv_14 << get_node_B_float(conv_13, data_path, "conv14", 512, PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(2, 2, 1, 1)); SubStream conv_15(conv_14); - conv_15 << get_node_B_float(conv_14, data_path, "conv15", 256, PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 2, 1, 1)); + conv_15 << get_node_B_float(conv_14, data_path, "conv15", 256, PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(2, 2, 1, 1)); SubStream conv_16(conv_15); - conv_16 << get_node_B_float(conv_15, data_path, "conv16", 256, PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 2, 1, 1)); + conv_16 << get_node_B_float(conv_15, data_path, "conv16", 256, PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(2, 2, 1, 1)); SubStream conv_17(conv_16); - conv_17 << get_node_B_float(conv_16, data_path, "conv17", 128, PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 2, 1, 1)); + conv_17 << get_node_B_float(conv_16, data_path, "conv17", 128, PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(2, 2, 1, 1)); //mbox_loc SubStream conv_11_mbox_loc(conv_11); @@ -293,8 +312,9 @@ private: conv_17_2_mbox_loc << get_node_C_float(conv_17, data_path, "conv17_2_mbox_loc", 24, PadStrideInfo(1, 1, 0, 0)); SubStream mbox_loc(graph); - mbox_loc << ConcatLayer(std::move(conv_11_mbox_loc), std::move(conv_13_mbox_loc), conv_14_2_mbox_loc, std::move(conv_15_2_mbox_loc), - std::move(conv_16_2_mbox_loc), std::move(conv_17_2_mbox_loc)); + mbox_loc << ConcatLayer(std::move(conv_11_mbox_loc), std::move(conv_13_mbox_loc), conv_14_2_mbox_loc, + std::move(conv_15_2_mbox_loc), std::move(conv_16_2_mbox_loc), + std::move(conv_17_2_mbox_loc)); //mbox_conf SubStream conv_11_mbox_conf(conv_11); @@ -304,67 +324,79 @@ private: conv_13_mbox_conf << get_node_C_float(conv_13, data_path, "conv13_mbox_conf", 126, PadStrideInfo(1, 1, 0, 0)); SubStream conv_14_2_mbox_conf(conv_14); - conv_14_2_mbox_conf << get_node_C_float(conv_14, data_path, "conv14_2_mbox_conf", 126, PadStrideInfo(1, 1, 0, 0)); + conv_14_2_mbox_conf << get_node_C_float(conv_14, data_path, "conv14_2_mbox_conf", 126, + PadStrideInfo(1, 1, 0, 0)); SubStream conv_15_2_mbox_conf(conv_15); - conv_15_2_mbox_conf << get_node_C_float(conv_15, data_path, "conv15_2_mbox_conf", 126, PadStrideInfo(1, 1, 0, 0)); + conv_15_2_mbox_conf << get_node_C_float(conv_15, data_path, "conv15_2_mbox_conf", 126, + PadStrideInfo(1, 1, 0, 0)); SubStream conv_16_2_mbox_conf(conv_16); - conv_16_2_mbox_conf << get_node_C_float(conv_16, data_path, "conv16_2_mbox_conf", 126, PadStrideInfo(1, 1, 0, 0)); + conv_16_2_mbox_conf << get_node_C_float(conv_16, data_path, "conv16_2_mbox_conf", 126, + PadStrideInfo(1, 1, 0, 0)); SubStream conv_17_2_mbox_conf(conv_17); - conv_17_2_mbox_conf << get_node_C_float(conv_17, data_path, "conv17_2_mbox_conf", 126, PadStrideInfo(1, 1, 0, 0)); + conv_17_2_mbox_conf << get_node_C_float(conv_17, data_path, "conv17_2_mbox_conf", 126, + PadStrideInfo(1, 1, 0, 0)); SubStream mbox_conf(graph); - mbox_conf << ConcatLayer(std::move(conv_11_mbox_conf), std::move(conv_13_mbox_conf), std::move(conv_14_2_mbox_conf), - std::move(conv_15_2_mbox_conf), std::move(conv_16_2_mbox_conf), std::move(conv_17_2_mbox_conf)); + mbox_conf << ConcatLayer(std::move(conv_11_mbox_conf), std::move(conv_13_mbox_conf), + std::move(conv_14_2_mbox_conf), std::move(conv_15_2_mbox_conf), + std::move(conv_16_2_mbox_conf), std::move(conv_17_2_mbox_conf)); mbox_conf << ReshapeLayer(TensorShape(21U, 1917U)).set_name("mbox_conf/reshape"); mbox_conf << SoftmaxLayer().set_name("mbox_conf/softmax"); mbox_conf << FlattenLayer().set_name("mbox_conf/flat"); - const std::vector<float> priorbox_variances = { 0.1f, 0.1f, 0.2f, 0.2f }; + const std::vector<float> priorbox_variances = {0.1f, 0.1f, 0.2f, 0.2f}; const float priorbox_offset = 0.5f; - const std::vector<float> priorbox_aspect_ratios = { 2.f, 3.f }; + const std::vector<float> priorbox_aspect_ratios = {2.f, 3.f}; //mbox_priorbox branch SubStream conv_11_mbox_priorbox(conv_11); conv_11_mbox_priorbox << PriorBoxLayer(SubStream(graph), - PriorBoxLayerInfo({ 60.f }, priorbox_variances, priorbox_offset, true, false, {}, { 2.f })) - .set_name("conv11/priorbox"); + PriorBoxLayerInfo({60.f}, priorbox_variances, priorbox_offset, true, + false, {}, {2.f})) + .set_name("conv11/priorbox"); SubStream conv_13_mbox_priorbox(conv_13); conv_13_mbox_priorbox << PriorBoxLayer(SubStream(graph), - PriorBoxLayerInfo({ 105.f }, priorbox_variances, priorbox_offset, true, false, { 150.f }, priorbox_aspect_ratios)) - .set_name("conv13/priorbox"); + PriorBoxLayerInfo({105.f}, priorbox_variances, priorbox_offset, true, + false, {150.f}, priorbox_aspect_ratios)) + .set_name("conv13/priorbox"); SubStream conv_14_2_mbox_priorbox(conv_14); conv_14_2_mbox_priorbox << PriorBoxLayer(SubStream(graph), - PriorBoxLayerInfo({ 150.f }, priorbox_variances, priorbox_offset, true, false, { 195.f }, priorbox_aspect_ratios)) - .set_name("conv14/priorbox"); + PriorBoxLayerInfo({150.f}, priorbox_variances, priorbox_offset, true, + false, {195.f}, priorbox_aspect_ratios)) + .set_name("conv14/priorbox"); SubStream conv_15_2_mbox_priorbox(conv_15); conv_15_2_mbox_priorbox << PriorBoxLayer(SubStream(graph), - PriorBoxLayerInfo({ 195.f }, priorbox_variances, priorbox_offset, true, false, { 240.f }, priorbox_aspect_ratios)) - .set_name("conv15/priorbox"); + PriorBoxLayerInfo({195.f}, priorbox_variances, priorbox_offset, true, + false, {240.f}, priorbox_aspect_ratios)) + .set_name("conv15/priorbox"); SubStream conv_16_2_mbox_priorbox(conv_16); conv_16_2_mbox_priorbox << PriorBoxLayer(SubStream(graph), - PriorBoxLayerInfo({ 240.f }, priorbox_variances, priorbox_offset, true, false, { 285.f }, priorbox_aspect_ratios)) - .set_name("conv16/priorbox"); + PriorBoxLayerInfo({240.f}, priorbox_variances, priorbox_offset, true, + false, {285.f}, priorbox_aspect_ratios)) + .set_name("conv16/priorbox"); SubStream conv_17_2_mbox_priorbox(conv_17); conv_17_2_mbox_priorbox << PriorBoxLayer(SubStream(graph), - PriorBoxLayerInfo({ 285.f }, priorbox_variances, priorbox_offset, true, false, { 300.f }, priorbox_aspect_ratios)) - .set_name("conv17/priorbox"); + PriorBoxLayerInfo({285.f}, priorbox_variances, priorbox_offset, true, + false, {300.f}, priorbox_aspect_ratios)) + .set_name("conv17/priorbox"); SubStream mbox_priorbox(graph); mbox_priorbox << ConcatLayer( - (common_params.data_layout == DataLayout::NCHW) ? arm_compute::graph::descriptors::ConcatLayerDescriptor(DataLayoutDimension::WIDTH) : arm_compute::graph::descriptors::ConcatLayerDescriptor( - DataLayoutDimension::CHANNEL), - std::move(conv_11_mbox_priorbox), std::move(conv_13_mbox_priorbox), std::move(conv_14_2_mbox_priorbox), - std::move(conv_15_2_mbox_priorbox), std::move(conv_16_2_mbox_priorbox), std::move(conv_17_2_mbox_priorbox)); + (common_params.data_layout == DataLayout::NCHW) + ? arm_compute::graph::descriptors::ConcatLayerDescriptor(DataLayoutDimension::WIDTH) + : arm_compute::graph::descriptors::ConcatLayerDescriptor(DataLayoutDimension::CHANNEL), + std::move(conv_11_mbox_priorbox), std::move(conv_13_mbox_priorbox), std::move(conv_14_2_mbox_priorbox), + std::move(conv_15_2_mbox_priorbox), std::move(conv_16_2_mbox_priorbox), std::move(conv_17_2_mbox_priorbox)); const int num_classes = 21; const bool share_location = true; @@ -377,77 +409,85 @@ private: SubStream detection_ouput(mbox_loc); detection_ouput << DetectionOutputLayer(std::move(mbox_conf), std::move(mbox_priorbox), - DetectionOutputLayerInfo(num_classes, share_location, detection_type, keep_top_k, nms_threshold, top_k, label_id_background, conf_thrs)); - detection_ouput << OutputLayer(get_detection_output_accessor(common_params, { input_descriptor.shape })); + DetectionOutputLayerInfo(num_classes, share_location, detection_type, + keep_top_k, nms_threshold, top_k, + label_id_background, conf_thrs)); + detection_ouput << OutputLayer(get_detection_output_accessor(common_params, {input_descriptor.shape})); } - ConcatLayer get_node_A_qasymm(IStream &main_graph, const std::string &data_path, std::string &¶m_path, - unsigned int conv_filt, - PadStrideInfo dwc_pad_stride_info, PadStrideInfo conv_pad_stride_info, - std::pair<QuantizationInfo, QuantizationInfo> depth_quant_info, std::pair<QuantizationInfo, QuantizationInfo> point_quant_info) + ConcatLayer get_node_A_qasymm(IStream &main_graph, + const std::string &data_path, + std::string &¶m_path, + unsigned int conv_filt, + PadStrideInfo dwc_pad_stride_info, + PadStrideInfo conv_pad_stride_info, + std::pair<QuantizationInfo, QuantizationInfo> depth_quant_info, + std::pair<QuantizationInfo, QuantizationInfo> point_quant_info) { const std::string total_path = param_path + "_"; SubStream sg(main_graph); - sg << DepthwiseConvolutionLayer( - 3U, 3U, - get_weights_accessor(data_path, total_path + "dw_w.npy"), - get_weights_accessor(data_path, total_path + "dw_b.npy"), - dwc_pad_stride_info, 1, depth_quant_info.first, depth_quant_info.second) - .set_name(param_path + "/dw") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name(param_path + "/dw/relu6"); - - sg << ConvolutionLayer( - 1U, 1U, conv_filt, - get_weights_accessor(data_path, total_path + "w.npy"), - get_weights_accessor(data_path, total_path + "b.npy"), - conv_pad_stride_info, 1, point_quant_info.first, point_quant_info.second) - .set_name(param_path + "/pw") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name(param_path + "/pw/relu6"); + sg << DepthwiseConvolutionLayer(3U, 3U, get_weights_accessor(data_path, total_path + "dw_w.npy"), + get_weights_accessor(data_path, total_path + "dw_b.npy"), dwc_pad_stride_info, + 1, depth_quant_info.first, depth_quant_info.second) + .set_name(param_path + "/dw") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) + .set_name(param_path + "/dw/relu6"); + + sg << ConvolutionLayer(1U, 1U, conv_filt, get_weights_accessor(data_path, total_path + "w.npy"), + get_weights_accessor(data_path, total_path + "b.npy"), conv_pad_stride_info, 1, + point_quant_info.first, point_quant_info.second) + .set_name(param_path + "/pw") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) + .set_name(param_path + "/pw/relu6"); return ConcatLayer(std::move(sg)); } - ConcatLayer get_node_B_qasymm(IStream &main_graph, const std::string &data_path, std::string &¶m_path, - unsigned int conv_filt, - PadStrideInfo conv_pad_stride_info_1x1, PadStrideInfo conv_pad_stride_info_3x3, - const std::pair<QuantizationInfo, QuantizationInfo> quant_info_1x1, const std::pair<QuantizationInfo, QuantizationInfo> quant_info_3x3) + ConcatLayer get_node_B_qasymm(IStream &main_graph, + const std::string &data_path, + std::string &¶m_path, + unsigned int conv_filt, + PadStrideInfo conv_pad_stride_info_1x1, + PadStrideInfo conv_pad_stride_info_3x3, + const std::pair<QuantizationInfo, QuantizationInfo> quant_info_1x1, + const std::pair<QuantizationInfo, QuantizationInfo> quant_info_3x3) { const std::string total_path = param_path + "_"; SubStream sg(main_graph); - sg << ConvolutionLayer( - 1, 1, conv_filt / 2, - get_weights_accessor(data_path, total_path + "1x1_w.npy"), - get_weights_accessor(data_path, total_path + "1x1_b.npy"), - conv_pad_stride_info_1x1, 1, quant_info_1x1.first, quant_info_1x1.second) - .set_name(total_path + "1x1/conv") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name(total_path + "1x1/conv/relu6"); - - sg << ConvolutionLayer( - 3, 3, conv_filt, - get_weights_accessor(data_path, total_path + "3x3_w.npy"), - get_weights_accessor(data_path, total_path + "3x3_b.npy"), - conv_pad_stride_info_3x3, 1, quant_info_3x3.first, quant_info_3x3.second) - .set_name(total_path + "3x3/conv") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name(total_path + "3x3/conv/relu6"); + sg << ConvolutionLayer(1, 1, conv_filt / 2, get_weights_accessor(data_path, total_path + "1x1_w.npy"), + get_weights_accessor(data_path, total_path + "1x1_b.npy"), conv_pad_stride_info_1x1, 1, + quant_info_1x1.first, quant_info_1x1.second) + .set_name(total_path + "1x1/conv") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) + .set_name(total_path + "1x1/conv/relu6"); + + sg << ConvolutionLayer(3, 3, conv_filt, get_weights_accessor(data_path, total_path + "3x3_w.npy"), + get_weights_accessor(data_path, total_path + "3x3_b.npy"), conv_pad_stride_info_3x3, 1, + quant_info_3x3.first, quant_info_3x3.second) + .set_name(total_path + "3x3/conv") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) + .set_name(total_path + "3x3/conv/relu6"); return ConcatLayer(std::move(sg)); } - ConcatLayer get_node_C_qasymm(IStream &main_graph, const std::string &data_path, std::string &¶m_path, - unsigned int conv_filt, PadStrideInfo conv_pad_stride_info, - const std::pair<QuantizationInfo, QuantizationInfo> quant_info, TensorShape reshape_shape) + ConcatLayer get_node_C_qasymm(IStream &main_graph, + const std::string &data_path, + std::string &¶m_path, + unsigned int conv_filt, + PadStrideInfo conv_pad_stride_info, + const std::pair<QuantizationInfo, QuantizationInfo> quant_info, + TensorShape reshape_shape) { const std::string total_path = param_path + "_"; SubStream sg(main_graph); - sg << ConvolutionLayer( - 1U, 1U, conv_filt, - get_weights_accessor(data_path, total_path + "w.npy"), - get_weights_accessor(data_path, total_path + "b.npy"), - conv_pad_stride_info, 1, quant_info.first, quant_info.second) - .set_name(param_path + "/conv"); - if(common_params.data_layout == DataLayout::NCHW) + sg << ConvolutionLayer(1U, 1U, conv_filt, get_weights_accessor(data_path, total_path + "w.npy"), + get_weights_accessor(data_path, total_path + "b.npy"), conv_pad_stride_info, 1, + quant_info.first, quant_info.second) + .set_name(param_path + "/conv"); + if (common_params.data_layout == DataLayout::NCHW) { sg << PermuteLayer(PermutationVector(2U, 0U, 1U), DataLayout::NHWC); } @@ -462,57 +502,59 @@ private: std::string data_path = common_params.data_path; // Add model path to data path - if(!data_path.empty()) + if (!data_path.empty()) { data_path += "/cnn_data/ssd_mobilenet_qasymm8_model/"; } // Quantization info are saved as pair for each (pointwise/depthwise) convolution layer: <weight_quant_info, output_quant_info> - const std::vector<std::pair<QuantizationInfo, QuantizationInfo>> conv_quant_info = - { - { QuantizationInfo(0.03624850884079933f, 163), QuantizationInfo(0.22219789028167725f, 113) }, // conv0 - { QuantizationInfo(0.0028752065263688564f, 113), QuantizationInfo(0.05433657020330429f, 128) }, // conv13_2_1_1 - { QuantizationInfo(0.0014862528769299388f, 125), QuantizationInfo(0.05037643015384674f, 131) }, // conv13_2_3_3 - { QuantizationInfo(0.00233650766313076f, 113), QuantizationInfo(0.04468846693634987f, 126) }, // conv13_3_1_1 - { QuantizationInfo(0.002501056529581547f, 120), QuantizationInfo(0.06026708707213402f, 111) }, // conv13_3_3_3 - { QuantizationInfo(0.002896666992455721f, 121), QuantizationInfo(0.037775348871946335f, 117) }, // conv13_4_1_1 - { QuantizationInfo(0.0023875406477600336f, 122), QuantizationInfo(0.03881589323282242f, 108) }, // conv13_4_3_3 - { QuantizationInfo(0.0022081052884459496f, 77), QuantizationInfo(0.025450613349676132f, 125) }, // conv13_5_1_1 - { QuantizationInfo(0.00604657270014286f, 121), QuantizationInfo(0.033533502370119095f, 109) } // conv13_5_3_3 + const std::vector<std::pair<QuantizationInfo, QuantizationInfo>> conv_quant_info = { + {QuantizationInfo(0.03624850884079933f, 163), QuantizationInfo(0.22219789028167725f, 113)}, // conv0 + {QuantizationInfo(0.0028752065263688564f, 113), + QuantizationInfo(0.05433657020330429f, 128)}, // conv13_2_1_1 + {QuantizationInfo(0.0014862528769299388f, 125), + QuantizationInfo(0.05037643015384674f, 131)}, // conv13_2_3_3 + {QuantizationInfo(0.00233650766313076f, 113), QuantizationInfo(0.04468846693634987f, 126)}, // conv13_3_1_1 + {QuantizationInfo(0.002501056529581547f, 120), QuantizationInfo(0.06026708707213402f, 111)}, // conv13_3_3_3 + {QuantizationInfo(0.002896666992455721f, 121), + QuantizationInfo(0.037775348871946335f, 117)}, // conv13_4_1_1 + {QuantizationInfo(0.0023875406477600336f, 122), + QuantizationInfo(0.03881589323282242f, 108)}, // conv13_4_3_3 + {QuantizationInfo(0.0022081052884459496f, 77), + QuantizationInfo(0.025450613349676132f, 125)}, // conv13_5_1_1 + {QuantizationInfo(0.00604657270014286f, 121), QuantizationInfo(0.033533502370119095f, 109)} // conv13_5_3_3 }; - const std::vector<std::pair<QuantizationInfo, QuantizationInfo>> depth_quant_info = - { - { QuantizationInfo(0.03408717364072f, 131), QuantizationInfo(0.29286590218544006f, 108) }, // dwsc1 - { QuantizationInfo(0.027518004179000854f, 107), QuantizationInfo(0.20796941220760345, 117) }, // dwsc2 - { QuantizationInfo(0.052489638328552246f, 85), QuantizationInfo(0.4303881824016571f, 142) }, // dwsc3 - { QuantizationInfo(0.016570359468460083f, 79), QuantizationInfo(0.10512150079011917f, 116) }, // dwsc4 - { QuantizationInfo(0.060739465057849884f, 65), QuantizationInfo(0.15331414341926575f, 94) }, // dwsc5 - { QuantizationInfo(0.01324534136801958f, 124), QuantizationInfo(0.13010895252227783f, 153) }, // dwsc6 - { QuantizationInfo(0.032326459884643555f, 124), QuantizationInfo(0.11565316468477249, 156) }, // dwsc7 - { QuantizationInfo(0.029948478564620018f, 155), QuantizationInfo(0.11413891613483429f, 146) }, // dwsc8 - { QuantizationInfo(0.028054025024175644f, 129), QuantizationInfo(0.1142905130982399f, 140) }, // dwsc9 - { QuantizationInfo(0.025204822421073914f, 129), QuantizationInfo(0.14668069779872894f, 149) }, // dwsc10 - { QuantizationInfo(0.019332280382514f, 110), QuantizationInfo(0.1480235457420349f, 91) }, // dwsc11 - { QuantizationInfo(0.0319712869822979f, 88), QuantizationInfo(0.10424695909023285f, 117) }, // dwsc12 - { QuantizationInfo(0.04378943517804146f, 164), QuantizationInfo(0.23176774382591248f, 138) } // dwsc13 + const std::vector<std::pair<QuantizationInfo, QuantizationInfo>> depth_quant_info = { + {QuantizationInfo(0.03408717364072f, 131), QuantizationInfo(0.29286590218544006f, 108)}, // dwsc1 + {QuantizationInfo(0.027518004179000854f, 107), QuantizationInfo(0.20796941220760345, 117)}, // dwsc2 + {QuantizationInfo(0.052489638328552246f, 85), QuantizationInfo(0.4303881824016571f, 142)}, // dwsc3 + {QuantizationInfo(0.016570359468460083f, 79), QuantizationInfo(0.10512150079011917f, 116)}, // dwsc4 + {QuantizationInfo(0.060739465057849884f, 65), QuantizationInfo(0.15331414341926575f, 94)}, // dwsc5 + {QuantizationInfo(0.01324534136801958f, 124), QuantizationInfo(0.13010895252227783f, 153)}, // dwsc6 + {QuantizationInfo(0.032326459884643555f, 124), QuantizationInfo(0.11565316468477249, 156)}, // dwsc7 + {QuantizationInfo(0.029948478564620018f, 155), QuantizationInfo(0.11413891613483429f, 146)}, // dwsc8 + {QuantizationInfo(0.028054025024175644f, 129), QuantizationInfo(0.1142905130982399f, 140)}, // dwsc9 + {QuantizationInfo(0.025204822421073914f, 129), QuantizationInfo(0.14668069779872894f, 149)}, // dwsc10 + {QuantizationInfo(0.019332280382514f, 110), QuantizationInfo(0.1480235457420349f, 91)}, // dwsc11 + {QuantizationInfo(0.0319712869822979f, 88), QuantizationInfo(0.10424695909023285f, 117)}, // dwsc12 + {QuantizationInfo(0.04378943517804146f, 164), QuantizationInfo(0.23176774382591248f, 138)} // dwsc13 }; - const std::vector<std::pair<QuantizationInfo, QuantizationInfo>> point_quant_info = - { - { QuantizationInfo(0.028777318075299263f, 144), QuantizationInfo(0.2663874328136444f, 121) }, // pw1 - { QuantizationInfo(0.015796702355146408f, 127), QuantizationInfo(0.1739964485168457f, 111) }, // pw2 - { QuantizationInfo(0.009349990636110306f, 127), QuantizationInfo(0.1805974692106247f, 104) }, // pw3 - { QuantizationInfo(0.012920888140797615f, 106), QuantizationInfo(0.1205204650759697f, 100) }, // pw4 - { QuantizationInfo(0.008119508624076843f, 145), QuantizationInfo(0.12272439152002335f, 97) }, // pw5 - { QuantizationInfo(0.0070041813887655735f, 115), QuantizationInfo(0.0947074219584465f, 101) }, // pw6 - { QuantizationInfo(0.004827278666198254f, 115), QuantizationInfo(0.0842885747551918f, 110) }, // pw7 - { QuantizationInfo(0.004755120258778334f, 128), QuantizationInfo(0.08283159881830215f, 116) }, // pw8 - { QuantizationInfo(0.007527193054556847f, 142), QuantizationInfo(0.12555131316184998f, 137) }, // pw9 - { QuantizationInfo(0.006050156895071268f, 109), QuantizationInfo(0.10871313512325287f, 124) }, // pw10 - { QuantizationInfo(0.00490700313821435f, 127), QuantizationInfo(0.10364262014627457f, 140) }, // pw11 - { QuantizationInfo(0.006063731852918863, 124), QuantizationInfo(0.11241862177848816f, 125) }, // pw12 - { QuantizationInfo(0.007901716977357864f, 139), QuantizationInfo(0.49889302253723145f, 141) } // pw13 + const std::vector<std::pair<QuantizationInfo, QuantizationInfo>> point_quant_info = { + {QuantizationInfo(0.028777318075299263f, 144), QuantizationInfo(0.2663874328136444f, 121)}, // pw1 + {QuantizationInfo(0.015796702355146408f, 127), QuantizationInfo(0.1739964485168457f, 111)}, // pw2 + {QuantizationInfo(0.009349990636110306f, 127), QuantizationInfo(0.1805974692106247f, 104)}, // pw3 + {QuantizationInfo(0.012920888140797615f, 106), QuantizationInfo(0.1205204650759697f, 100)}, // pw4 + {QuantizationInfo(0.008119508624076843f, 145), QuantizationInfo(0.12272439152002335f, 97)}, // pw5 + {QuantizationInfo(0.0070041813887655735f, 115), QuantizationInfo(0.0947074219584465f, 101)}, // pw6 + {QuantizationInfo(0.004827278666198254f, 115), QuantizationInfo(0.0842885747551918f, 110)}, // pw7 + {QuantizationInfo(0.004755120258778334f, 128), QuantizationInfo(0.08283159881830215f, 116)}, // pw8 + {QuantizationInfo(0.007527193054556847f, 142), QuantizationInfo(0.12555131316184998f, 137)}, // pw9 + {QuantizationInfo(0.006050156895071268f, 109), QuantizationInfo(0.10871313512325287f, 124)}, // pw10 + {QuantizationInfo(0.00490700313821435f, 127), QuantizationInfo(0.10364262014627457f, 140)}, // pw11 + {QuantizationInfo(0.006063731852918863, 124), QuantizationInfo(0.11241862177848816f, 125)}, // pw12 + {QuantizationInfo(0.007901716977357864f, 139), QuantizationInfo(0.49889302253723145f, 141)} // pw13 }; // Quantization info taken from the TfLite SSD MobileNet example @@ -520,114 +562,154 @@ private: // Create core graph graph << InputLayer(input_descriptor.set_quantization_info(in_quant_info), get_weights_accessor(data_path, common_params.image, DataLayout::NHWC)); - graph << ConvolutionLayer( - 3U, 3U, 32U, - get_weights_accessor(data_path, "conv0_w.npy"), - get_weights_accessor(data_path, "conv0_b.npy"), - PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::CEIL), 1, conv_quant_info.at(0).first, conv_quant_info.at(0).second) - .set_name("conv0"); - graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name("conv0/relu"); - graph << get_node_A_qasymm(graph, data_path, "conv1", 64U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(0), - point_quant_info.at(0)); - graph << get_node_A_qasymm(graph, data_path, "conv2", 128U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::CEIL), PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(1), - point_quant_info.at(1)); - graph << get_node_A_qasymm(graph, data_path, "conv3", 128U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(2), - point_quant_info.at(2)); - graph << get_node_A_qasymm(graph, data_path, "conv4", 256U, PadStrideInfo(2U, 2U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(3), - point_quant_info.at(3)); - graph << get_node_A_qasymm(graph, data_path, "conv5", 256U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(4), - point_quant_info.at(4)); - graph << get_node_A_qasymm(graph, data_path, "conv6", 512U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::CEIL), PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(5), - point_quant_info.at(5)); - graph << get_node_A_qasymm(graph, data_path, "conv7", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(6), - point_quant_info.at(6)); - graph << get_node_A_qasymm(graph, data_path, "conv8", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(7), - point_quant_info.at(7)); - graph << get_node_A_qasymm(graph, data_path, "conv9", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(8), - point_quant_info.at(8)); - graph << get_node_A_qasymm(graph, data_path, "conv10", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(9), - point_quant_info.at(9)); - graph << get_node_A_qasymm(graph, data_path, "conv11", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(10), - point_quant_info.at(10)); + graph << ConvolutionLayer(3U, 3U, 32U, get_weights_accessor(data_path, "conv0_w.npy"), + get_weights_accessor(data_path, "conv0_b.npy"), + PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::CEIL), 1, + conv_quant_info.at(0).first, conv_quant_info.at(0).second) + .set_name("conv0"); + graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) + .set_name("conv0/relu"); + graph << get_node_A_qasymm(graph, data_path, "conv1", 64U, + PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), + PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(0), point_quant_info.at(0)); + graph << get_node_A_qasymm(graph, data_path, "conv2", 128U, + PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::CEIL), + PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(1), point_quant_info.at(1)); + graph << get_node_A_qasymm(graph, data_path, "conv3", 128U, + PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), + PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(2), point_quant_info.at(2)); + graph << get_node_A_qasymm(graph, data_path, "conv4", 256U, + PadStrideInfo(2U, 2U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), + PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(3), point_quant_info.at(3)); + graph << get_node_A_qasymm(graph, data_path, "conv5", 256U, + PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), + PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(4), point_quant_info.at(4)); + graph << get_node_A_qasymm(graph, data_path, "conv6", 512U, + PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::CEIL), + PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(5), point_quant_info.at(5)); + graph << get_node_A_qasymm(graph, data_path, "conv7", 512U, + PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), + PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(6), point_quant_info.at(6)); + graph << get_node_A_qasymm(graph, data_path, "conv8", 512U, + PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), + PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(7), point_quant_info.at(7)); + graph << get_node_A_qasymm(graph, data_path, "conv9", 512U, + PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), + PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(8), point_quant_info.at(8)); + graph << get_node_A_qasymm(graph, data_path, "conv10", 512U, + PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), + PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(9), point_quant_info.at(9)); + graph << get_node_A_qasymm(graph, data_path, "conv11", 512U, + PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), + PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(10), point_quant_info.at(10)); SubStream conv_13(graph); - conv_13 << get_node_A_qasymm(graph, data_path, "conv12", 1024U, PadStrideInfo(2U, 2U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(11), - point_quant_info.at(11)); - conv_13 << get_node_A_qasymm(conv_13, data_path, "conv13", 1024U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(12), - point_quant_info.at(12)); + conv_13 << get_node_A_qasymm(graph, data_path, "conv12", 1024U, + PadStrideInfo(2U, 2U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), + PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(11), point_quant_info.at(11)); + conv_13 << get_node_A_qasymm(conv_13, data_path, "conv13", 1024U, + PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), + PadStrideInfo(1U, 1U, 0U, 0U), depth_quant_info.at(12), point_quant_info.at(12)); SubStream conv_14(conv_13); - conv_14 << get_node_B_qasymm(conv_13, data_path, "conv13_2", 512U, PadStrideInfo(1U, 1U, 0U, 0U), PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::CEIL), conv_quant_info.at(1), - conv_quant_info.at(2)); + conv_14 << get_node_B_qasymm(conv_13, data_path, "conv13_2", 512U, PadStrideInfo(1U, 1U, 0U, 0U), + PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::CEIL), + conv_quant_info.at(1), conv_quant_info.at(2)); SubStream conv_15(conv_14); - conv_15 << get_node_B_qasymm(conv_14, data_path, "conv13_3", 256U, PadStrideInfo(1U, 1U, 0U, 0U), PadStrideInfo(2U, 2U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), conv_quant_info.at(3), - conv_quant_info.at(4)); + conv_15 << get_node_B_qasymm(conv_14, data_path, "conv13_3", 256U, PadStrideInfo(1U, 1U, 0U, 0U), + PadStrideInfo(2U, 2U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), + conv_quant_info.at(3), conv_quant_info.at(4)); SubStream conv_16(conv_15); - conv_16 << get_node_B_qasymm(conv_15, data_path, "conv13_4", 256U, PadStrideInfo(1U, 1U, 0U, 0U), PadStrideInfo(2U, 2U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), conv_quant_info.at(5), - conv_quant_info.at(6)); + conv_16 << get_node_B_qasymm(conv_15, data_path, "conv13_4", 256U, PadStrideInfo(1U, 1U, 0U, 0U), + PadStrideInfo(2U, 2U, 1U, 1U, 1U, 1U, DimensionRoundingType::CEIL), + conv_quant_info.at(5), conv_quant_info.at(6)); SubStream conv_17(conv_16); - conv_17 << get_node_B_qasymm(conv_16, data_path, "conv13_5", 128U, PadStrideInfo(1U, 1U, 0U, 0U), PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::CEIL), conv_quant_info.at(7), - conv_quant_info.at(8)); + conv_17 << get_node_B_qasymm(conv_16, data_path, "conv13_5", 128U, PadStrideInfo(1U, 1U, 0U, 0U), + PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::CEIL), + conv_quant_info.at(7), conv_quant_info.at(8)); // box_predictor - const std::vector<std::pair<QuantizationInfo, QuantizationInfo>> box_enc_pred_quant_info = - { - { QuantizationInfo(0.005202020984143019f, 136), QuantizationInfo(0.08655580133199692f, 183) }, // boxpredictor0_bep - { QuantizationInfo(0.003121797926723957f, 132), QuantizationInfo(0.03218776360154152f, 140) }, // boxpredictor1_bep - { QuantizationInfo(0.002995674265548587f, 130), QuantizationInfo(0.029072262346744537f, 125) }, // boxpredictor2_bep - { QuantizationInfo(0.0023131705820560455f, 130), QuantizationInfo(0.026488754898309708f, 127) }, // boxpredictor3_bep - { QuantizationInfo(0.0013905081432312727f, 132), QuantizationInfo(0.0199890099465847f, 137) }, // boxpredictor4_bep - { QuantizationInfo(0.00216794665902853f, 121), QuantizationInfo(0.019798893481492996f, 151) } // boxpredictor5_bep + const std::vector<std::pair<QuantizationInfo, QuantizationInfo>> box_enc_pred_quant_info = { + {QuantizationInfo(0.005202020984143019f, 136), + QuantizationInfo(0.08655580133199692f, 183)}, // boxpredictor0_bep + {QuantizationInfo(0.003121797926723957f, 132), + QuantizationInfo(0.03218776360154152f, 140)}, // boxpredictor1_bep + {QuantizationInfo(0.002995674265548587f, 130), + QuantizationInfo(0.029072262346744537f, 125)}, // boxpredictor2_bep + {QuantizationInfo(0.0023131705820560455f, 130), + QuantizationInfo(0.026488754898309708f, 127)}, // boxpredictor3_bep + {QuantizationInfo(0.0013905081432312727f, 132), + QuantizationInfo(0.0199890099465847f, 137)}, // boxpredictor4_bep + {QuantizationInfo(0.00216794665902853f, 121), + QuantizationInfo(0.019798893481492996f, 151)} // boxpredictor5_bep }; const std::vector<TensorShape> box_reshape = // NHWC - { - TensorShape(4U, 1U, 1083U), // boxpredictor0_bep_reshape - TensorShape(4U, 1U, 600U), // boxpredictor1_bep_reshape - TensorShape(4U, 1U, 150U), // boxpredictor2_bep_reshape - TensorShape(4U, 1U, 54U), // boxpredictor3_bep_reshape - TensorShape(4U, 1U, 24U), // boxpredictor4_bep_reshape - TensorShape(4U, 1U, 6U) // boxpredictor5_bep_reshape - }; + { + TensorShape(4U, 1U, 1083U), // boxpredictor0_bep_reshape + TensorShape(4U, 1U, 600U), // boxpredictor1_bep_reshape + TensorShape(4U, 1U, 150U), // boxpredictor2_bep_reshape + TensorShape(4U, 1U, 54U), // boxpredictor3_bep_reshape + TensorShape(4U, 1U, 24U), // boxpredictor4_bep_reshape + TensorShape(4U, 1U, 6U) // boxpredictor5_bep_reshape + }; SubStream conv_11_box_enc_pre(graph); - conv_11_box_enc_pre << get_node_C_qasymm(graph, data_path, "BoxPredictor_0_BEP", 12U, PadStrideInfo(1U, 1U, 0U, 0U), box_enc_pred_quant_info.at(0), box_reshape.at(0)); + conv_11_box_enc_pre << get_node_C_qasymm(graph, data_path, "BoxPredictor_0_BEP", 12U, + PadStrideInfo(1U, 1U, 0U, 0U), box_enc_pred_quant_info.at(0), + box_reshape.at(0)); SubStream conv_13_box_enc_pre(conv_13); - conv_13_box_enc_pre << get_node_C_qasymm(conv_13, data_path, "BoxPredictor_1_BEP", 24U, PadStrideInfo(1U, 1U, 0U, 0U), box_enc_pred_quant_info.at(1), box_reshape.at(1)); + conv_13_box_enc_pre << get_node_C_qasymm(conv_13, data_path, "BoxPredictor_1_BEP", 24U, + PadStrideInfo(1U, 1U, 0U, 0U), box_enc_pred_quant_info.at(1), + box_reshape.at(1)); SubStream conv_14_2_box_enc_pre(conv_14); - conv_14_2_box_enc_pre << get_node_C_qasymm(conv_14, data_path, "BoxPredictor_2_BEP", 24U, PadStrideInfo(1U, 1U, 0U, 0U), box_enc_pred_quant_info.at(2), box_reshape.at(2)); + conv_14_2_box_enc_pre << get_node_C_qasymm(conv_14, data_path, "BoxPredictor_2_BEP", 24U, + PadStrideInfo(1U, 1U, 0U, 0U), box_enc_pred_quant_info.at(2), + box_reshape.at(2)); SubStream conv_15_2_box_enc_pre(conv_15); - conv_15_2_box_enc_pre << get_node_C_qasymm(conv_15, data_path, "BoxPredictor_3_BEP", 24U, PadStrideInfo(1U, 1U, 0U, 0U), box_enc_pred_quant_info.at(3), box_reshape.at(3)); + conv_15_2_box_enc_pre << get_node_C_qasymm(conv_15, data_path, "BoxPredictor_3_BEP", 24U, + PadStrideInfo(1U, 1U, 0U, 0U), box_enc_pred_quant_info.at(3), + box_reshape.at(3)); SubStream conv_16_2_box_enc_pre(conv_16); - conv_16_2_box_enc_pre << get_node_C_qasymm(conv_16, data_path, "BoxPredictor_4_BEP", 24U, PadStrideInfo(1U, 1U, 0U, 0U), box_enc_pred_quant_info.at(4), box_reshape.at(4)); + conv_16_2_box_enc_pre << get_node_C_qasymm(conv_16, data_path, "BoxPredictor_4_BEP", 24U, + PadStrideInfo(1U, 1U, 0U, 0U), box_enc_pred_quant_info.at(4), + box_reshape.at(4)); SubStream conv_17_2_box_enc_pre(conv_17); - conv_17_2_box_enc_pre << get_node_C_qasymm(conv_17, data_path, "BoxPredictor_5_BEP", 24U, PadStrideInfo(1U, 1U, 0U, 0U), box_enc_pred_quant_info.at(5), box_reshape.at(5)); + conv_17_2_box_enc_pre << get_node_C_qasymm(conv_17, data_path, "BoxPredictor_5_BEP", 24U, + PadStrideInfo(1U, 1U, 0U, 0U), box_enc_pred_quant_info.at(5), + box_reshape.at(5)); SubStream box_enc_pre(graph); const QuantizationInfo bep_concate_qinfo = QuantizationInfo(0.08655580133199692f, 183); - box_enc_pre << ConcatLayer(arm_compute::graph::descriptors::ConcatLayerDescriptor(DataLayoutDimension::HEIGHT, bep_concate_qinfo), - std::move(conv_11_box_enc_pre), std::move(conv_13_box_enc_pre), conv_14_2_box_enc_pre, std::move(conv_15_2_box_enc_pre), + box_enc_pre << ConcatLayer(arm_compute::graph::descriptors::ConcatLayerDescriptor(DataLayoutDimension::HEIGHT, + bep_concate_qinfo), + std::move(conv_11_box_enc_pre), std::move(conv_13_box_enc_pre), + conv_14_2_box_enc_pre, std::move(conv_15_2_box_enc_pre), std::move(conv_16_2_box_enc_pre), std::move(conv_17_2_box_enc_pre)) - .set_name("BoxPredictor/concat"); + .set_name("BoxPredictor/concat"); box_enc_pre << ReshapeLayer(TensorShape(4U, 1917U)).set_name("BoxPredictor/reshape"); // class_predictor - const std::vector<std::pair<QuantizationInfo, QuantizationInfo>> class_pred_quant_info = - { - { QuantizationInfo(0.002744135679677129f, 125), QuantizationInfo(0.05746262148022652f, 234) }, // boxpredictor0_cp - { QuantizationInfo(0.0024326108396053314f, 80), QuantizationInfo(0.03764628246426582f, 217) }, // boxpredictor1_cp - { QuantizationInfo(0.0013898586621508002f, 141), QuantizationInfo(0.034081317484378815f, 214) }, // boxpredictor2_cp - { QuantizationInfo(0.0014176908880472183f, 133), QuantizationInfo(0.033889178186655045f, 215) }, // boxpredictor3_cp - { QuantizationInfo(0.001090311910957098f, 125), QuantizationInfo(0.02646234817802906f, 230) }, // boxpredictor4_cp - { QuantizationInfo(0.001134163816459477f, 115), QuantizationInfo(0.026926767081022263f, 218) } // boxpredictor5_cp + const std::vector<std::pair<QuantizationInfo, QuantizationInfo>> class_pred_quant_info = { + {QuantizationInfo(0.002744135679677129f, 125), + QuantizationInfo(0.05746262148022652f, 234)}, // boxpredictor0_cp + {QuantizationInfo(0.0024326108396053314f, 80), + QuantizationInfo(0.03764628246426582f, 217)}, // boxpredictor1_cp + {QuantizationInfo(0.0013898586621508002f, 141), + QuantizationInfo(0.034081317484378815f, 214)}, // boxpredictor2_cp + {QuantizationInfo(0.0014176908880472183f, 133), + QuantizationInfo(0.033889178186655045f, 215)}, // boxpredictor3_cp + {QuantizationInfo(0.001090311910957098f, 125), + QuantizationInfo(0.02646234817802906f, 230)}, // boxpredictor4_cp + {QuantizationInfo(0.001134163816459477f, 115), + QuantizationInfo(0.026926767081022263f, 218)} // boxpredictor5_cp }; - const std::vector<TensorShape> class_reshape = - { + const std::vector<TensorShape> class_reshape = { TensorShape(91U, 1083U), // boxpredictor0_cp_reshape TensorShape(91U, 600U), // boxpredictor1_cp_reshape TensorShape(91U, 150U), // boxpredictor2_cp_reshape @@ -637,60 +719,80 @@ private: }; SubStream conv_11_class_pre(graph); - conv_11_class_pre << get_node_C_qasymm(graph, data_path, "BoxPredictor_0_CP", 273U, PadStrideInfo(1U, 1U, 0U, 0U), class_pred_quant_info.at(0), class_reshape.at(0)); + conv_11_class_pre << get_node_C_qasymm(graph, data_path, "BoxPredictor_0_CP", 273U, + PadStrideInfo(1U, 1U, 0U, 0U), class_pred_quant_info.at(0), + class_reshape.at(0)); SubStream conv_13_class_pre(conv_13); - conv_13_class_pre << get_node_C_qasymm(conv_13, data_path, "BoxPredictor_1_CP", 546U, PadStrideInfo(1U, 1U, 0U, 0U), class_pred_quant_info.at(1), class_reshape.at(1)); + conv_13_class_pre << get_node_C_qasymm(conv_13, data_path, "BoxPredictor_1_CP", 546U, + PadStrideInfo(1U, 1U, 0U, 0U), class_pred_quant_info.at(1), + class_reshape.at(1)); SubStream conv_14_2_class_pre(conv_14); - conv_14_2_class_pre << get_node_C_qasymm(conv_14, data_path, "BoxPredictor_2_CP", 546U, PadStrideInfo(1U, 1U, 0U, 0U), class_pred_quant_info.at(2), class_reshape.at(2)); + conv_14_2_class_pre << get_node_C_qasymm(conv_14, data_path, "BoxPredictor_2_CP", 546U, + PadStrideInfo(1U, 1U, 0U, 0U), class_pred_quant_info.at(2), + class_reshape.at(2)); SubStream conv_15_2_class_pre(conv_15); - conv_15_2_class_pre << get_node_C_qasymm(conv_15, data_path, "BoxPredictor_3_CP", 546U, PadStrideInfo(1U, 1U, 0U, 0U), class_pred_quant_info.at(3), class_reshape.at(3)); + conv_15_2_class_pre << get_node_C_qasymm(conv_15, data_path, "BoxPredictor_3_CP", 546U, + PadStrideInfo(1U, 1U, 0U, 0U), class_pred_quant_info.at(3), + class_reshape.at(3)); SubStream conv_16_2_class_pre(conv_16); - conv_16_2_class_pre << get_node_C_qasymm(conv_16, data_path, "BoxPredictor_4_CP", 546U, PadStrideInfo(1U, 1U, 0U, 0U), class_pred_quant_info.at(4), class_reshape.at(4)); + conv_16_2_class_pre << get_node_C_qasymm(conv_16, data_path, "BoxPredictor_4_CP", 546U, + PadStrideInfo(1U, 1U, 0U, 0U), class_pred_quant_info.at(4), + class_reshape.at(4)); SubStream conv_17_2_class_pre(conv_17); - conv_17_2_class_pre << get_node_C_qasymm(conv_17, data_path, "BoxPredictor_5_CP", 546U, PadStrideInfo(1U, 1U, 0U, 0U), class_pred_quant_info.at(5), class_reshape.at(5)); + conv_17_2_class_pre << get_node_C_qasymm(conv_17, data_path, "BoxPredictor_5_CP", 546U, + PadStrideInfo(1U, 1U, 0U, 0U), class_pred_quant_info.at(5), + class_reshape.at(5)); const QuantizationInfo cp_concate_qinfo = QuantizationInfo(0.0584389753639698f, 230); SubStream class_pred(graph); - class_pred << ConcatLayer( - arm_compute::graph::descriptors::ConcatLayerDescriptor(DataLayoutDimension::WIDTH, cp_concate_qinfo), - std::move(conv_11_class_pre), std::move(conv_13_class_pre), std::move(conv_14_2_class_pre), - std::move(conv_15_2_class_pre), std::move(conv_16_2_class_pre), std::move(conv_17_2_class_pre)) - .set_name("ClassPrediction/concat"); + class_pred << ConcatLayer(arm_compute::graph::descriptors::ConcatLayerDescriptor(DataLayoutDimension::WIDTH, + cp_concate_qinfo), + std::move(conv_11_class_pre), std::move(conv_13_class_pre), + std::move(conv_14_2_class_pre), std::move(conv_15_2_class_pre), + std::move(conv_16_2_class_pre), std::move(conv_17_2_class_pre)) + .set_name("ClassPrediction/concat"); const QuantizationInfo logistic_out_qinfo = QuantizationInfo(0.00390625f, 0); - class_pred << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC), logistic_out_qinfo).set_name("ClassPrediction/logistic"); - - const int max_detections = 10; - const int max_classes_per_detection = 1; - const float nms_score_threshold = 0.30000001192092896f; - const float nms_iou_threshold = 0.6000000238418579f; - const int num_classes = 90; - const float x_scale = 10.f; - const float y_scale = 10.f; - const float h_scale = 5.f; - const float w_scale = 5.f; - std::array<float, 4> scales = { y_scale, x_scale, w_scale, h_scale }; - const QuantizationInfo anchors_qinfo = QuantizationInfo(0.006453060545027256f, 0); + class_pred << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC), + logistic_out_qinfo) + .set_name("ClassPrediction/logistic"); + + const int max_detections = 10; + const int max_classes_per_detection = 1; + const float nms_score_threshold = 0.30000001192092896f; + const float nms_iou_threshold = 0.6000000238418579f; + const int num_classes = 90; + const float x_scale = 10.f; + const float y_scale = 10.f; + const float h_scale = 5.f; + const float w_scale = 5.f; + std::array<float, 4> scales = {y_scale, x_scale, w_scale, h_scale}; + const QuantizationInfo anchors_qinfo = QuantizationInfo(0.006453060545027256f, 0); SubStream detection_ouput(box_enc_pre); detection_ouput << DetectionPostProcessLayer(std::move(class_pred), - DetectionPostProcessLayerInfo(max_detections, max_classes_per_detection, nms_score_threshold, nms_iou_threshold, num_classes, scales), + DetectionPostProcessLayerInfo( + max_detections, max_classes_per_detection, nms_score_threshold, + nms_iou_threshold, num_classes, scales), get_weights_accessor(data_path, "anchors.npy"), anchors_qinfo) - .set_name("DetectionPostProcess"); + .set_name("DetectionPostProcess"); SubStream ouput_0(detection_ouput); - ouput_0 << OutputLayer(get_npy_output_accessor(detection_boxes_opt->value(), TensorShape(4U, 10U), DataType::F32), 0); + ouput_0 << OutputLayer( + get_npy_output_accessor(detection_boxes_opt->value(), TensorShape(4U, 10U), DataType::F32), 0); SubStream ouput_1(detection_ouput); - ouput_1 << OutputLayer(get_npy_output_accessor(detection_classes_opt->value(), TensorShape(10U), DataType::F32), 1); + ouput_1 << OutputLayer(get_npy_output_accessor(detection_classes_opt->value(), TensorShape(10U), DataType::F32), + 1); SubStream ouput_2(detection_ouput); - ouput_2 << OutputLayer(get_npy_output_accessor(detection_scores_opt->value(), TensorShape(10U), DataType::F32), 2); + ouput_2 << OutputLayer(get_npy_output_accessor(detection_scores_opt->value(), TensorShape(10U), DataType::F32), + 2); SubStream ouput_3(detection_ouput); ouput_3 << OutputLayer(get_npy_output_accessor(num_detections_opt->value(), TensorShape(1U), DataType::F32), 3); diff --git a/examples/graph_vgg16.cpp b/examples/graph_vgg16.cpp index fcfe6ef50d..72ac9694b1 100644 --- a/examples/graph_vgg16.cpp +++ b/examples/graph_vgg16.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -35,8 +36,7 @@ using namespace arm_compute::graph_utils; class GraphVGG16Example : public Example { public: - GraphVGG16Example() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "VGG16") + GraphVGG16Example() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "VGG16") { } bool do_setup(int argc, char **argv) override @@ -49,7 +49,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -62,153 +62,143 @@ public: std::string data_path = common_params.data_path; // Create a preprocessor object - const std::array<float, 3> mean_rgb{ { 123.68f, 116.779f, 103.939f } }; + const std::array<float, 3> mean_rgb{{123.68f, 116.779f, 103.939f}}; std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<CaffePreproccessor>(mean_rgb); // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; // Create graph - graph << common_params.target - << common_params.fast_math_hint - << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor))) - // Layer 1 - << ConvolutionLayer( - 3U, 3U, 64U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_1_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv1_1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1_1/Relu") - // Layer 2 - << ConvolutionLayer( - 3U, 3U, 64U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_2_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv1_2") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1_2/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool1") - // Layer 3 - << ConvolutionLayer( - 3U, 3U, 128U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_1_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2_1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2_1/Relu") - // Layer 4 - << ConvolutionLayer( - 3U, 3U, 128U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_2_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2_2") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2_2/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool2") - // Layer 5 - << ConvolutionLayer( - 3U, 3U, 256U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_1_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv3_1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_1/Relu") - // Layer 6 - << ConvolutionLayer( - 3U, 3U, 256U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_2_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv3_2") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_2/Relu") - // Layer 7 - << ConvolutionLayer( - 3U, 3U, 256U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_3_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_3_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv3_3") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_3/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool3") - // Layer 8 - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_1_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv4_1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_1/Relu") - // Layer 9 - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_2_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv4_2") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_2/Relu") - // Layer 10 - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_3_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_3_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv4_3") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_3/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool4") - // Layer 11 - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_1_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv5_1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_1/Relu") - // Layer 12 - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_2_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv5_2") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_2/Relu") - // Layer 13 - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_3_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_3_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv5_3") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_3/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool5") - // Layer 14 - << FullyConnectedLayer( - 4096U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc6_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc6_b.npy")) - .set_name("fc6") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu") - // Layer 15 - << FullyConnectedLayer( - 4096U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc7_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc7_b.npy")) - .set_name("fc7") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu_1") - // Layer 16 - << FullyConnectedLayer( - 1000U, - get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc8_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc8_b.npy")) - .set_name("fc8") - // Softmax - << SoftmaxLayer().set_name("prob") - << OutputLayer(get_output_accessor(common_params, 5)); + graph + << common_params.target << common_params.fast_math_hint + << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor))) + // Layer 1 + << ConvolutionLayer( + 3U, 3U, 64U, get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_1_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv1_1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv1_1/Relu") + // Layer 2 + << ConvolutionLayer( + 3U, 3U, 64U, get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_2_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_2_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv1_2") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv1_2/Relu") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool1") + // Layer 3 + << ConvolutionLayer( + 3U, 3U, 128U, get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_1_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2_1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv2_1/Relu") + // Layer 4 + << ConvolutionLayer( + 3U, 3U, 128U, get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_2_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_2_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2_2") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv2_2/Relu") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool2") + // Layer 5 + << ConvolutionLayer( + 3U, 3U, 256U, get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_1_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv3_1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv3_1/Relu") + // Layer 6 + << ConvolutionLayer( + 3U, 3U, 256U, get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_2_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_2_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv3_2") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv3_2/Relu") + // Layer 7 + << ConvolutionLayer( + 3U, 3U, 256U, get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_3_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_3_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv3_3") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv3_3/Relu") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool3") + // Layer 8 + << ConvolutionLayer( + 3U, 3U, 512U, get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_1_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv4_1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv4_1/Relu") + // Layer 9 + << ConvolutionLayer( + 3U, 3U, 512U, get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_2_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_2_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv4_2") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv4_2/Relu") + // Layer 10 + << ConvolutionLayer( + 3U, 3U, 512U, get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_3_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_3_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv4_3") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv4_3/Relu") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool4") + // Layer 11 + << ConvolutionLayer( + 3U, 3U, 512U, get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_1_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv5_1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv5_1/Relu") + // Layer 12 + << ConvolutionLayer( + 3U, 3U, 512U, get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_2_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_2_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv5_2") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv5_2/Relu") + // Layer 13 + << ConvolutionLayer( + 3U, 3U, 512U, get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_3_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_3_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv5_3") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv5_3/Relu") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool5") + // Layer 14 + << FullyConnectedLayer(4096U, + get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc6_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc6_b.npy")) + .set_name("fc6") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu") + // Layer 15 + << FullyConnectedLayer(4096U, + get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc7_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc7_b.npy")) + .set_name("fc7") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu_1") + // Layer 16 + << FullyConnectedLayer(1000U, + get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc8_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc8_b.npy")) + .set_name("fc8") + // Softmax + << SoftmaxLayer().set_name("prob") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph GraphConfig config; diff --git a/examples/graph_vgg19.cpp b/examples/graph_vgg19.cpp index efc0bcce19..9293544655 100644 --- a/examples/graph_vgg19.cpp +++ b/examples/graph_vgg19.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -34,8 +35,7 @@ using namespace arm_compute::graph_utils; class GraphVGG19Example : public Example { public: - GraphVGG19Example() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "VGG19") + GraphVGG19Example() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "VGG19") { } bool do_setup(int argc, char **argv) override @@ -48,7 +48,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -61,165 +61,152 @@ public: std::string data_path = common_params.data_path; // Create a preprocessor object - const std::array<float, 3> mean_rgb{ { 123.68f, 116.779f, 103.939f } }; + const std::array<float, 3> mean_rgb{{123.68f, 116.779f, 103.939f}}; std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<CaffePreproccessor>(mean_rgb); // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint - << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor))) - // Layer 1 - << ConvolutionLayer( - 3U, 3U, 64U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_1_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv1_1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1_1/Relu") - << ConvolutionLayer( - 3U, 3U, 64U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_2_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv1_2") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1_2/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool1") - // Layer 2 - << ConvolutionLayer( - 3U, 3U, 128U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_1_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2_1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2_1/Relu") - << ConvolutionLayer( - 3U, 3U, 128U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_2_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2_2") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2_2/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool2") - // Layer 3 - << ConvolutionLayer( - 3U, 3U, 256U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_1_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv3_1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_1/Relu") - << ConvolutionLayer( - 3U, 3U, 256U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_2_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv3_2") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_2/Relu") - << ConvolutionLayer( - 3U, 3U, 256U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_3_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_3_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv3_3") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_3/Relu") - << ConvolutionLayer( - 3U, 3U, 256U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_4_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_4_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv3_4") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_4/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool3") - // Layer 4 - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_1_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv4_1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_1/Relu") - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_2_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv4_2") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_2/Relu") - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_3_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_3_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv4_3") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_3/Relu") - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_4_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_4_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv4_4") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_4/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool4") - // Layer 5 - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_1_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv5_1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_1/Relu") - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_2_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv5_2") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_2/Relu") - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_3_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_3_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv5_3") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_3/Relu") - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_4_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_4_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv5_4") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_4/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool5") - // Layer 6 - << FullyConnectedLayer( - 4096U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc6_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc6_b.npy")) - .set_name("fc6") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu") - // Layer 7 - << FullyConnectedLayer( - 4096U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc7_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc7_b.npy")) - .set_name("fc7") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu_1") - // Layer 8 - << FullyConnectedLayer( - 1000U, - get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc8_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc8_b.npy")) - .set_name("fc8") - // Softmax - << SoftmaxLayer().set_name("prob") - << OutputLayer(get_output_accessor(common_params, 5)); + graph + << common_params.target << common_params.fast_math_hint + << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor))) + // Layer 1 + << ConvolutionLayer( + 3U, 3U, 64U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_1_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv1_1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv1_1/Relu") + << ConvolutionLayer( + 3U, 3U, 64U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_2_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_2_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv1_2") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv1_2/Relu") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool1") + // Layer 2 + << ConvolutionLayer( + 3U, 3U, 128U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_1_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2_1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv2_1/Relu") + << ConvolutionLayer( + 3U, 3U, 128U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_2_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_2_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2_2") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv2_2/Relu") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool2") + // Layer 3 + << ConvolutionLayer( + 3U, 3U, 256U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_1_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv3_1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv3_1/Relu") + << ConvolutionLayer( + 3U, 3U, 256U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_2_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_2_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv3_2") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv3_2/Relu") + << ConvolutionLayer( + 3U, 3U, 256U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_3_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_3_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv3_3") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv3_3/Relu") + << ConvolutionLayer( + 3U, 3U, 256U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_4_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_4_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv3_4") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv3_4/Relu") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool3") + // Layer 4 + << ConvolutionLayer( + 3U, 3U, 512U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_1_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv4_1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv4_1/Relu") + << ConvolutionLayer( + 3U, 3U, 512U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_2_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_2_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv4_2") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv4_2/Relu") + << ConvolutionLayer( + 3U, 3U, 512U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_3_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_3_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv4_3") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv4_3/Relu") + << ConvolutionLayer( + 3U, 3U, 512U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_4_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_4_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv4_4") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv4_4/Relu") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool4") + // Layer 5 + << ConvolutionLayer( + 3U, 3U, 512U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_1_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv5_1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv5_1/Relu") + << ConvolutionLayer( + 3U, 3U, 512U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_2_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_2_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv5_2") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv5_2/Relu") + << ConvolutionLayer( + 3U, 3U, 512U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_3_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_3_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv5_3") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv5_3/Relu") + << ConvolutionLayer( + 3U, 3U, 512U, get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_4_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_4_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv5_4") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv5_4/Relu") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool5") + // Layer 6 + << FullyConnectedLayer(4096U, + get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc6_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc6_b.npy")) + .set_name("fc6") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu") + // Layer 7 + << FullyConnectedLayer(4096U, + get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc7_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc7_b.npy")) + .set_name("fc7") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu_1") + // Layer 8 + << FullyConnectedLayer(1000U, + get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc8_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc8_b.npy")) + .set_name("fc8") + // Softmax + << SoftmaxLayer().set_name("prob") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph GraphConfig config; diff --git a/examples/graph_vgg_vdsr.cpp b/examples/graph_vgg_vdsr.cpp index 3fe28e0fed..a6cd337f82 100644 --- a/examples/graph_vgg_vdsr.cpp +++ b/examples/graph_vgg_vdsr.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -36,8 +37,7 @@ using namespace arm_compute::graph_utils; class GraphVDSRExample : public Example { public: - GraphVDSRExample() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "VDSR") + GraphVDSRExample() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "VDSR") { model_input_width = cmd_parser.add_option<SimpleOption<unsigned int>>("image-width", 192); model_input_height = cmd_parser.add_option<SimpleOption<unsigned int>>("image-height", 192); @@ -46,7 +46,7 @@ public: model_input_width->set_help("Input image width."); model_input_height->set_help("Input image height."); } - GraphVDSRExample(const GraphVDSRExample &) = delete; + GraphVDSRExample(const GraphVDSRExample &) = delete; GraphVDSRExample &operator=(const GraphVDSRExample &) = delete; ~GraphVDSRExample() override = default; bool do_setup(int argc, char **argv) override @@ -59,7 +59,7 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; @@ -82,15 +82,17 @@ public: std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<TFPreproccessor>(); // Create input descriptor - const TensorShape tensor_shape = permute_shape(TensorShape(image_width, image_height, 1U, common_params.batches), DataLayout::NCHW, common_params.data_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(image_width, image_height, 1U, common_params.batches), DataLayout::NCHW, + common_params.data_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; // Note: Quantization info are random and used only for benchmarking purposes - graph << common_params.target - << common_params.fast_math_hint + graph << common_params.target << common_params.fast_math_hint << InputLayer(input_descriptor.set_quantization_info(QuantizationInfo(0.0078125f, 128)), get_input_accessor(common_params, std::move(preprocessor), false)); @@ -98,37 +100,34 @@ public: SubStream right(graph); // Layer 1 - right << ConvolutionLayer( - 3U, 3U, 64U, - get_weights_accessor(data_path, "conv0_w.npy", weights_layout), - get_weights_accessor(data_path, "conv0_b.npy"), - PadStrideInfo(1, 1, 1, 1), 1, QuantizationInfo(0.031778190285f, 156), QuantizationInfo(0.0784313753247f, 128)) - .set_name("conv0") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv0/Relu"); + right << ConvolutionLayer(3U, 3U, 64U, get_weights_accessor(data_path, "conv0_w.npy", weights_layout), + get_weights_accessor(data_path, "conv0_b.npy"), PadStrideInfo(1, 1, 1, 1), 1, + QuantizationInfo(0.031778190285f, 156), QuantizationInfo(0.0784313753247f, 128)) + .set_name("conv0") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv0/Relu"); // Rest 17 layers - for(unsigned int i = 1; i < 19; ++i) + for (unsigned int i = 1; i < 19; ++i) { const std::string conv_w_path = "conv" + arm_compute::support::cpp11::to_string(i) + "_w.npy"; const std::string conv_b_path = "conv" + arm_compute::support::cpp11::to_string(i) + "_b.npy"; const std::string conv_name = "conv" + arm_compute::support::cpp11::to_string(i); - right << ConvolutionLayer( - 3U, 3U, 64U, - get_weights_accessor(data_path, conv_w_path, weights_layout), - get_weights_accessor(data_path, conv_b_path), - PadStrideInfo(1, 1, 1, 1), 1, QuantizationInfo(0.015851572156f, 93)) - .set_name(conv_name) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(conv_name + "/Relu"); + right << ConvolutionLayer(3U, 3U, 64U, get_weights_accessor(data_path, conv_w_path, weights_layout), + get_weights_accessor(data_path, conv_b_path), PadStrideInfo(1, 1, 1, 1), 1, + QuantizationInfo(0.015851572156f, 93)) + .set_name(conv_name) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(conv_name + "/Relu"); } // Final layer - right << ConvolutionLayer( - 3U, 3U, 1U, - get_weights_accessor(data_path, "conv20_w.npy", weights_layout), - get_weights_accessor(data_path, "conv20_b.npy"), - PadStrideInfo(1, 1, 1, 1), 1, QuantizationInfo(0.015851572156f, 93)) - .set_name("conv20") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv20/Relu"); + right << ConvolutionLayer(3U, 3U, 1U, get_weights_accessor(data_path, "conv20_w.npy", weights_layout), + get_weights_accessor(data_path, "conv20_b.npy"), PadStrideInfo(1, 1, 1, 1), 1, + QuantizationInfo(0.015851572156f, 93)) + .set_name("conv20") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv20/Relu"); // Add residual to input graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name("add") @@ -157,8 +156,8 @@ public: private: CommandLineParser cmd_parser; CommonGraphOptions common_opts; - SimpleOption<unsigned int> *model_input_width{ nullptr }; - SimpleOption<unsigned int> *model_input_height{ nullptr }; + SimpleOption<unsigned int> *model_input_width{nullptr}; + SimpleOption<unsigned int> *model_input_height{nullptr}; CommonGraphParams common_params; Stream graph; }; diff --git a/examples/graph_yolov3.cpp b/examples/graph_yolov3.cpp index 3c8ddbffd8..5c8d3426ec 100644 --- a/examples/graph_yolov3.cpp +++ b/examples/graph_yolov3.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -35,8 +36,7 @@ using namespace arm_compute::graph_utils; class GraphYOLOv3Example : public Example { public: - GraphYOLOv3Example() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "YOLOv3") + GraphYOLOv3Example() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "YOLOv3") { } @@ -50,14 +50,15 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; } // Checks - ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph"); + ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), + "QASYMM8 not supported for this graph"); // Print parameter values std::cout << common_params << std::endl; @@ -69,331 +70,322 @@ public: std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<TFPreproccessor>(0.f); // Create input descriptor - const TensorShape tensor_shape = permute_shape(TensorShape(608U, 608U, 3U, 1U), DataLayout::NCHW, common_params.data_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(608U, 608U, 3U, 1U), DataLayout::NCHW, common_params.data_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint + graph << common_params.target << common_params.fast_math_hint << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false)); std::pair<SubStream, SubStream> intermediate_layers = darknet53(data_path, weights_layout); - graph << ConvolutionLayer( - 1U, 1U, 512U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_53_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2d_53") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_beta.npy"), - 0.000001f) - .set_name("conv2d_53/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_53/LeakyRelu") - << ConvolutionLayer( - 3U, 3U, 1024U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_54_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2d_54") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_beta.npy"), - 0.000001f) - .set_name("conv2d_54/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_54/LeakyRelu") - << ConvolutionLayer( - 1U, 1U, 512U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_55_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2d_55") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_beta.npy"), - 0.000001f) - .set_name("conv2d_55/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_55/LeakyRelu") - << ConvolutionLayer( - 3U, 3U, 1024U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_56_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2d_56") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_beta.npy"), - 0.000001f) - .set_name("conv2d_56/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_56/LeakyRelu") - << ConvolutionLayer( - 1U, 1U, 512U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_57_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2d_57") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_beta.npy"), - 0.000001f) - .set_name("conv2d_57/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_57/LeakyRelu"); + graph + << ConvolutionLayer( + 1U, 1U, 512U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_53_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2d_53") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_beta.npy"), 0.000001f) + .set_name("conv2d_53/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_53/LeakyRelu") + << ConvolutionLayer( + 3U, 3U, 1024U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_54_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2d_54") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_beta.npy"), 0.000001f) + .set_name("conv2d_54/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_54/LeakyRelu") + << ConvolutionLayer( + 1U, 1U, 512U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_55_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2d_55") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_beta.npy"), 0.000001f) + .set_name("conv2d_55/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_55/LeakyRelu") + << ConvolutionLayer( + 3U, 3U, 1024U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_56_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2d_56") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_beta.npy"), 0.000001f) + .set_name("conv2d_56/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_56/LeakyRelu") + << ConvolutionLayer( + 1U, 1U, 512U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_57_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2d_57") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_beta.npy"), 0.000001f) + .set_name("conv2d_57/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_57/LeakyRelu"); SubStream route_1(graph); - graph << ConvolutionLayer( - 3U, 3U, 1024U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_58_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2d_58") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_beta.npy"), - 0.000001f) - .set_name("conv2d_58/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_58/LeakyRelu") - << ConvolutionLayer( - 1U, 1U, 255U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_59_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_59_b.npy", weights_layout), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2d_59") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f)).set_name("conv2d_59/Linear") - << YOLOLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.1f)).set_name("Yolo1") - << OutputLayer(get_output_accessor(common_params, 5)); + graph + << ConvolutionLayer( + 3U, 3U, 1024U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_58_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2d_58") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_beta.npy"), 0.000001f) + .set_name("conv2d_58/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_58/LeakyRelu") + << ConvolutionLayer( + 1U, 1U, 255U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_59_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_59_b.npy", weights_layout), + PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2d_59") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f)) + .set_name("conv2d_59/Linear") + << YOLOLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.1f)).set_name("Yolo1") + << OutputLayer(get_output_accessor(common_params, 5)); route_1 << ConvolutionLayer( - 1U, 1U, 256U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_60_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2d_60") + 1U, 1U, 256U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_60_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2d_60") << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_beta.npy"), - 0.000001f) - .set_name("conv2d_59/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_60/LeakyRelu") + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_beta.npy"), + 0.000001f) + .set_name("conv2d_59/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_60/LeakyRelu") << ResizeLayer(InterpolationPolicy::NEAREST_NEIGHBOR, 2, 2).set_name("Upsample_60"); SubStream concat_1(route_1); - concat_1 << ConcatLayer(std::move(route_1), std::move(intermediate_layers.second)).set_name("Route1") - << ConvolutionLayer( - 1U, 1U, 256U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_61_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2d_61") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_beta.npy"), - 0.000001f) - .set_name("conv2d_60/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_61/LeakyRelu") - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_62_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2d_62") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_beta.npy"), - 0.000001f) - .set_name("conv2d_61/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_62/LeakyRelu") - << ConvolutionLayer( - 1U, 1U, 256U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_63_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2d_63") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_beta.npy"), - 0.000001f) - .set_name("conv2d_62/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_63/LeakyRelu") - << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_64_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2d_64") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_beta.npy"), - 0.000001f) - .set_name("conv2d_63/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_64/LeakyRelu") - << ConvolutionLayer( - 1U, 1U, 256U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_65_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2d_65") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_beta.npy"), - 0.000001f) - .set_name("conv2d_65/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_65/LeakyRelu"); + concat_1 + << ConcatLayer(std::move(route_1), std::move(intermediate_layers.second)).set_name("Route1") + << ConvolutionLayer( + 1U, 1U, 256U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_61_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2d_61") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_beta.npy"), 0.000001f) + .set_name("conv2d_60/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_61/LeakyRelu") + << ConvolutionLayer( + 3U, 3U, 512U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_62_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2d_62") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_beta.npy"), 0.000001f) + .set_name("conv2d_61/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_62/LeakyRelu") + << ConvolutionLayer( + 1U, 1U, 256U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_63_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2d_63") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_beta.npy"), 0.000001f) + .set_name("conv2d_62/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_63/LeakyRelu") + << ConvolutionLayer( + 3U, 3U, 512U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_64_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2d_64") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_beta.npy"), 0.000001f) + .set_name("conv2d_63/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_64/LeakyRelu") + << ConvolutionLayer( + 1U, 1U, 256U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_65_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2d_65") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_beta.npy"), 0.000001f) + .set_name("conv2d_65/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_65/LeakyRelu"); SubStream route_2(concat_1); - concat_1 << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_66_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2d_66") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_beta.npy"), - 0.000001f) - .set_name("conv2d_65/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_66/LeakyRelu") - << ConvolutionLayer( - 1U, 1U, 255U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_67_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_67_b.npy", weights_layout), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2d_67") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f)).set_name("conv2d_67/Linear") - << YOLOLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.1f)).set_name("Yolo2") - << OutputLayer(get_output_accessor(common_params, 5)); + concat_1 + << ConvolutionLayer( + 3U, 3U, 512U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_66_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2d_66") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_beta.npy"), 0.000001f) + .set_name("conv2d_65/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_66/LeakyRelu") + << ConvolutionLayer( + 1U, 1U, 255U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_67_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_67_b.npy", weights_layout), + PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2d_67") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f)) + .set_name("conv2d_67/Linear") + << YOLOLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.1f)).set_name("Yolo2") + << OutputLayer(get_output_accessor(common_params, 5)); route_2 << ConvolutionLayer( - 1U, 1U, 128U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_68_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2d_68") + 1U, 1U, 128U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_68_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2d_68") << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_beta.npy"), - 0.000001f) - .set_name("conv2d_66/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_68/LeakyRelu") + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_beta.npy"), + 0.000001f) + .set_name("conv2d_66/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_68/LeakyRelu") << ResizeLayer(InterpolationPolicy::NEAREST_NEIGHBOR, 2, 2).set_name("Upsample_68"); SubStream concat_2(route_2); - concat_2 << ConcatLayer(std::move(route_2), std::move(intermediate_layers.first)).set_name("Route2") - << ConvolutionLayer( - 1U, 1U, 128U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_69_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2d_69") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_beta.npy"), - 0.000001f) - .set_name("conv2d_67/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_69/LeakyRelu") - << ConvolutionLayer( - 3U, 3U, 256U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_70_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2d_70") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_beta.npy"), - 0.000001f) - .set_name("conv2d_68/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_70/LeakyRelu") - << ConvolutionLayer( - 1U, 1U, 128U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_71_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2d_71") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_beta.npy"), - 0.000001f) - .set_name("conv2d_69/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_71/LeakyRelu") - << ConvolutionLayer( - 3U, 3U, 256U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_72_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2d_72") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_beta.npy"), - 0.000001f) - .set_name("conv2d_70/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_72/LeakyRelu") - << ConvolutionLayer( - 1U, 1U, 128U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_73_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2d_73") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_beta.npy"), - 0.000001f) - .set_name("conv2d_71/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_73/LeakyRelu") - << ConvolutionLayer( - 3U, 3U, 256U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_74_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2d_74") - << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_beta.npy"), - 0.000001f) - .set_name("conv2d_72/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_74/LeakyRelu") - << ConvolutionLayer( - 1U, 1U, 255U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_75_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_75_b.npy", weights_layout), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2d_75") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f)).set_name("conv2d_75/Linear") - << YOLOLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.1f)).set_name("Yolo3") - << OutputLayer(get_output_accessor(common_params, 5)); + concat_2 + << ConcatLayer(std::move(route_2), std::move(intermediate_layers.first)).set_name("Route2") + << ConvolutionLayer( + 1U, 1U, 128U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_69_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2d_69") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_beta.npy"), 0.000001f) + .set_name("conv2d_67/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_69/LeakyRelu") + << ConvolutionLayer( + 3U, 3U, 256U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_70_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2d_70") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_beta.npy"), 0.000001f) + .set_name("conv2d_68/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_70/LeakyRelu") + << ConvolutionLayer( + 1U, 1U, 128U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_71_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2d_71") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_beta.npy"), 0.000001f) + .set_name("conv2d_69/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_71/LeakyRelu") + << ConvolutionLayer( + 3U, 3U, 256U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_72_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2d_72") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_beta.npy"), 0.000001f) + .set_name("conv2d_70/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_72/LeakyRelu") + << ConvolutionLayer( + 1U, 1U, 128U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_73_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2d_73") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_beta.npy"), 0.000001f) + .set_name("conv2d_71/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_73/LeakyRelu") + << ConvolutionLayer( + 3U, 3U, 256U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_74_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2d_74") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_beta.npy"), 0.000001f) + .set_name("conv2d_72/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_74/LeakyRelu") + << ConvolutionLayer( + 1U, 1U, 255U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_75_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_75_b.npy", weights_layout), + PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2d_75") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f)) + .set_name("conv2d_75/Linear") + << YOLOLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.1f)).set_name("Yolo3") + << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph GraphConfig config; @@ -422,64 +414,64 @@ private: std::pair<SubStream, SubStream> darknet53(const std::string &data_path, DataLayout weights_layout) { graph << ConvolutionLayer( - 3U, 3U, 32U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_1_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2d_1/Conv2D") + 3U, 3U, 32U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_1_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2d_1/Conv2D") << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_beta.npy"), - 0.000001f) - .set_name("conv2d_1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_1/LeakyRelu") + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_beta.npy"), + 0.000001f) + .set_name("conv2d_1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_1/LeakyRelu") << ConvolutionLayer( - 3U, 3U, 64U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_2_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 1, 1)) - .set_name("conv2d_2/Conv2D") + 3U, 3U, 64U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_2_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 1, 1)) + .set_name("conv2d_2/Conv2D") << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_beta.npy"), - 0.000001f) - .set_name("conv2d_2/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_2/LeakyRelu"); + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_beta.npy"), + 0.000001f) + .set_name("conv2d_2/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_2/LeakyRelu"); darknet53_block(data_path, "3", weights_layout, 32U); graph << ConvolutionLayer( - 3U, 3U, 128U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_5_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 1, 1)) - .set_name("conv2d_5/Conv2D") + 3U, 3U, 128U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_5_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 1, 1)) + .set_name("conv2d_5/Conv2D") << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_beta.npy"), - 0.000001f) - .set_name("conv2d_5/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_5/LeakyRelu"); + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_beta.npy"), + 0.000001f) + .set_name("conv2d_5/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_5/LeakyRelu"); darknet53_block(data_path, "6", weights_layout, 64U); darknet53_block(data_path, "8", weights_layout, 64U); graph << ConvolutionLayer( - 3U, 3U, 256U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_10_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 1, 1)) - .set_name("conv2d_10/Conv2D") + 3U, 3U, 256U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_10_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 1, 1)) + .set_name("conv2d_10/Conv2D") << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_beta.npy"), - 0.000001f) - .set_name("conv2d_10/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_10/LeakyRelu"); + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_beta.npy"), + 0.000001f) + .set_name("conv2d_10/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_10/LeakyRelu"); darknet53_block(data_path, "11", weights_layout, 128U); darknet53_block(data_path, "13", weights_layout, 128U); darknet53_block(data_path, "15", weights_layout, 128U); @@ -490,19 +482,19 @@ private: darknet53_block(data_path, "25", weights_layout, 128U); SubStream layer_36(graph); graph << ConvolutionLayer( - 3U, 3U, 512U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_27_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 1, 1)) - .set_name("conv2d_27/Conv2D") + 3U, 3U, 512U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_27_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 1, 1)) + .set_name("conv2d_27/Conv2D") << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_beta.npy"), - 0.000001f) - .set_name("conv2d_27/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_27/LeakyRelu"); + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_beta.npy"), + 0.000001f) + .set_name("conv2d_27/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_27/LeakyRelu"); darknet53_block(data_path, "28", weights_layout, 256U); darknet53_block(data_path, "30", weights_layout, 256U); darknet53_block(data_path, "32", weights_layout, 256U); @@ -513,19 +505,19 @@ private: darknet53_block(data_path, "42", weights_layout, 256U); SubStream layer_61(graph); graph << ConvolutionLayer( - 3U, 3U, 1024U, - get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_44_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 1, 1)) - .set_name("conv2d_44/Conv2D") + 3U, 3U, 1024U, + get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_44_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 1, 1)) + .set_name("conv2d_44/Conv2D") << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_var.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_beta.npy"), - 0.000001f) - .set_name("conv2d_44/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_44/LeakyRelu"); + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_var.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_beta.npy"), + 0.000001f) + .set_name("conv2d_44/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_44/LeakyRelu"); darknet53_block(data_path, "45", weights_layout, 512U); darknet53_block(data_path, "47", weights_layout, 512U); darknet53_block(data_path, "49", weights_layout, 512U); @@ -534,43 +526,48 @@ private: return std::pair<SubStream, SubStream>(layer_36, layer_61); } - void darknet53_block(const std::string &data_path, std::string &¶m_path, DataLayout weights_layout, - unsigned int filter_size) + void darknet53_block(const std::string &data_path, + std::string &¶m_path, + DataLayout weights_layout, + unsigned int filter_size) { - std::string total_path = "/cnn_data/yolov3_model/"; - std::string param_path2 = arm_compute::support::cpp11::to_string(arm_compute::support::cpp11::stoi(param_path) + 1); - SubStream i_a(graph); - SubStream i_b(graph); + std::string total_path = "/cnn_data/yolov3_model/"; + std::string param_path2 = + arm_compute::support::cpp11::to_string(arm_compute::support::cpp11::stoi(param_path) + 1); + SubStream i_a(graph); + SubStream i_b(graph); i_a << ConvolutionLayer( - 1U, 1U, filter_size, - get_weights_accessor(data_path, total_path + "conv2d_" + param_path + "_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2d_" + param_path + "/Conv2D") + 1U, 1U, filter_size, + get_weights_accessor(data_path, total_path + "conv2d_" + param_path + "_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2d_" + param_path + "/Conv2D") << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_mean.npy"), - get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_var.npy"), - get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_gamma.npy"), - get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_beta.npy"), - 0.000001f) - .set_name("conv2d_" + param_path + "/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_" + param_path + "/LeakyRelu") + get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_mean.npy"), + get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_var.npy"), + get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_gamma.npy"), + get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_beta.npy"), + 0.000001f) + .set_name("conv2d_" + param_path + "/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_" + param_path + "/LeakyRelu") << ConvolutionLayer( - 3U, 3U, filter_size * 2, - get_weights_accessor(data_path, total_path + "conv2d_" + param_path2 + "_w.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2d_" + param_path2 + "/Conv2D") + 3U, 3U, filter_size * 2, + get_weights_accessor(data_path, total_path + "conv2d_" + param_path2 + "_w.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2d_" + param_path2 + "/Conv2D") << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_mean.npy"), - get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_var.npy"), - get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_gamma.npy"), - get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_beta.npy"), - 0.000001f) - .set_name("conv2d_" + param_path2 + "/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_" + param_path2 + "/LeakyRelu"); + get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_mean.npy"), + get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_var.npy"), + get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_gamma.npy"), + get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_beta.npy"), + 0.000001f) + .set_name("conv2d_" + param_path2 + "/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)) + .set_name("conv2d_" + param_path2 + "/LeakyRelu"); - graph << EltwiseLayer(std::move(i_a), std::move(i_b), EltwiseOperation::Add).set_name("").set_name("add_" + param_path + "_" + param_path2); + graph << EltwiseLayer(std::move(i_a), std::move(i_b), EltwiseOperation::Add) + .set_name("") + .set_name("add_" + param_path + "_" + param_path2); } }; diff --git a/examples/neon_cnn.cpp b/examples/neon_cnn.cpp index 5ecf055e60..1f7a1ea6ca 100644 --- a/examples/neon_cnn.cpp +++ b/examples/neon_cnn.cpp @@ -21,13 +21,13 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/runtime/NEON/NEFunctions.h" - #include "arm_compute/core/Types.h" #include "arm_compute/runtime/Allocator.h" #include "arm_compute/runtime/BlobLifetimeManager.h" #include "arm_compute/runtime/MemoryManagerOnDemand.h" +#include "arm_compute/runtime/NEON/NEFunctions.h" #include "arm_compute/runtime/PoolManager.h" + #include "utils/Utils.h" using namespace arm_compute; @@ -43,12 +43,13 @@ public: // Create memory manager components // We need 2 memory managers: 1 for handling the tensors within the functions (mm_layers) and 1 for handling the input and output tensors of the functions (mm_transitions)) - auto lifetime_mgr0 = std::make_shared<BlobLifetimeManager>(); // Create lifetime manager - auto lifetime_mgr1 = std::make_shared<BlobLifetimeManager>(); // Create lifetime manager - auto pool_mgr0 = std::make_shared<PoolManager>(); // Create pool manager - auto pool_mgr1 = std::make_shared<PoolManager>(); // Create pool manager - auto mm_layers = std::make_shared<MemoryManagerOnDemand>(lifetime_mgr0, pool_mgr0); // Create the memory manager - auto mm_transitions = std::make_shared<MemoryManagerOnDemand>(lifetime_mgr1, pool_mgr1); // Create the memory manager + auto lifetime_mgr0 = std::make_shared<BlobLifetimeManager>(); // Create lifetime manager + auto lifetime_mgr1 = std::make_shared<BlobLifetimeManager>(); // Create lifetime manager + auto pool_mgr0 = std::make_shared<PoolManager>(); // Create pool manager + auto pool_mgr1 = std::make_shared<PoolManager>(); // Create pool manager + auto mm_layers = std::make_shared<MemoryManagerOnDemand>(lifetime_mgr0, pool_mgr0); // Create the memory manager + auto mm_transitions = + std::make_shared<MemoryManagerOnDemand>(lifetime_mgr1, pool_mgr1); // Create the memory manager // The weights and biases tensors should be initialized with the values inferred with the training @@ -116,7 +117,8 @@ public: // Initialize tensor of fc0 constexpr unsigned int num_labels = 128; - const TensorShape weights_shape_fc0(out_shape_pool1.x() * out_shape_pool1.y() * out_shape_pool1.z(), num_labels); + const TensorShape weights_shape_fc0(out_shape_pool1.x() * out_shape_pool1.y() * out_shape_pool1.z(), + num_labels); const TensorShape biases_shape_fc0(num_labels); const TensorShape out_shape_fc0(num_labels); @@ -138,22 +140,28 @@ public: /* [Configure functions] */ // in:32x32x1: 5x5 convolution, 8 output features maps (OFM) - conv0->configure(&src, &weights0, &biases0, &out_conv0, PadStrideInfo(1 /* stride_x */, 1 /* stride_y */, 2 /* pad_x */, 2 /* pad_y */)); + conv0->configure(&src, &weights0, &biases0, &out_conv0, + PadStrideInfo(1 /* stride_x */, 1 /* stride_y */, 2 /* pad_x */, 2 /* pad_y */)); // in:32x32x8, out:32x32x8, Activation function: relu act0.configure(&out_conv0, &out_act0, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); // in:32x32x8, out:16x16x8 (2x2 pooling), Pool type function: Max - pool0.configure(&out_act0, &out_pool0, PoolingLayerInfo(PoolingType::MAX, 2, data_layout, PadStrideInfo(2 /* stride_x */, 2 /* stride_y */))); + pool0.configure( + &out_act0, &out_pool0, + PoolingLayerInfo(PoolingType::MAX, 2, data_layout, PadStrideInfo(2 /* stride_x */, 2 /* stride_y */))); // in:16x16x8: 3x3 convolution, 16 output features maps (OFM) - conv1->configure(&out_pool0, &weights1, &biases1, &out_conv1, PadStrideInfo(1 /* stride_x */, 1 /* stride_y */, 1 /* pad_x */, 1 /* pad_y */)); + conv1->configure(&out_pool0, &weights1, &biases1, &out_conv1, + PadStrideInfo(1 /* stride_x */, 1 /* stride_y */, 1 /* pad_x */, 1 /* pad_y */)); // in:16x16x16, out:16x16x16, Activation function: relu act1.configure(&out_conv1, &out_act1, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); // in:16x16x16, out:8x8x16 (2x2 pooling), Pool type function: Average - pool1.configure(&out_act1, &out_pool1, PoolingLayerInfo(PoolingType::AVG, 2, data_layout, PadStrideInfo(2 /* stride_x */, 2 /* stride_y */))); + pool1.configure( + &out_act1, &out_pool1, + PoolingLayerInfo(PoolingType::AVG, 2, data_layout, PadStrideInfo(2 /* stride_x */, 2 /* stride_y */))); // in:8x8x16, out:128 fc0->configure(&out_pool1, &weights2, &biases2, &out_fc0); diff --git a/examples/neon_copy_objects.cpp b/examples/neon_copy_objects.cpp index b060b09759..6e9ebcaad5 100644 --- a/examples/neon_copy_objects.cpp +++ b/examples/neon_copy_objects.cpp @@ -22,9 +22,9 @@ * SOFTWARE. */ +#include "arm_compute/core/Types.h" #include "arm_compute/runtime/NEON/NEFunctions.h" -#include "arm_compute/core/Types.h" #include "utils/Utils.h" #include <cstring> @@ -50,11 +50,11 @@ public: dst_data = new float[width * height * batch]; // Fill src_data with pseudo(meaningless) values: - for(unsigned int b = 0; b < batch; b++) + for (unsigned int b = 0; b < batch; b++) { - for(unsigned int h = 0; h < height; h++) + for (unsigned int h = 0; h < height; h++) { - for(unsigned int w = 0; w < width; w++) + for (unsigned int w = 0; w < width; w++) { src_data[b * (width * height) + h * width + w] = static_cast<float>(100 * b + 10 * h + w); } @@ -78,9 +78,12 @@ public: Window input_window; input_window.use_tensor_dimensions(input.info()->tensor_shape()); std::cout << " Dimensions of the input's iterator:\n"; - std::cout << " X = [start=" << input_window.x().start() << ", end=" << input_window.x().end() << ", step=" << input_window.x().step() << "]\n"; - std::cout << " Y = [start=" << input_window.y().start() << ", end=" << input_window.y().end() << ", step=" << input_window.y().step() << "]\n"; - std::cout << " Z = [start=" << input_window.z().start() << ", end=" << input_window.z().end() << ", step=" << input_window.z().step() << "]\n"; + std::cout << " X = [start=" << input_window.x().start() << ", end=" << input_window.x().end() + << ", step=" << input_window.x().step() << "]\n"; + std::cout << " Y = [start=" << input_window.y().start() << ", end=" << input_window.y().end() + << ", step=" << input_window.y().step() << "]\n"; + std::cout << " Z = [start=" << input_window.z().start() << ", end=" << input_window.z().end() + << ", step=" << input_window.z().step() << "]\n"; // Create an iterator: Iterator input_it(&input, input_window); @@ -98,20 +101,28 @@ public: // } // } // Except it works for an arbitrary number of dimensions - execute_window_loop(input_window, [&](const Coordinates & id) - { - std::cout << "Setting item [" << id.x() << "," << id.y() << "," << id.z() << "]\n"; - *reinterpret_cast<float *>(input_it.ptr()) = src_data[id.z() * (width * height) + id.y() * width + id.x()]; - }, - input_it); + execute_window_loop( + input_window, + [&](const Coordinates &id) + { + std::cout << "Setting item [" << id.x() << "," << id.y() << "," << id.z() << "]\n"; + *reinterpret_cast<float *>(input_it.ptr()) = + src_data[id.z() * (width * height) + id.y() * width + id.x()]; + }, + input_it); // More efficient way: create an iterator to iterate through each row (instead of each element) of the output tensor: Window output_window; - output_window.use_tensor_dimensions(output.info()->tensor_shape(), /* first_dimension =*/Window::DimY); // Iterate through the rows (not each element) + output_window.use_tensor_dimensions( + output.info()->tensor_shape(), + /* first_dimension =*/Window::DimY); // Iterate through the rows (not each element) std::cout << " Dimensions of the output's iterator:\n"; - std::cout << " X = [start=" << output_window.x().start() << ", end=" << output_window.x().end() << ", step=" << output_window.x().step() << "]\n"; - std::cout << " Y = [start=" << output_window.y().start() << ", end=" << output_window.y().end() << ", step=" << output_window.y().step() << "]\n"; - std::cout << " Z = [start=" << output_window.z().start() << ", end=" << output_window.z().end() << ", step=" << output_window.z().step() << "]\n"; + std::cout << " X = [start=" << output_window.x().start() << ", end=" << output_window.x().end() + << ", step=" << output_window.x().step() << "]\n"; + std::cout << " Y = [start=" << output_window.y().start() << ", end=" << output_window.y().end() + << ", step=" << output_window.y().step() << "]\n"; + std::cout << " Z = [start=" << output_window.z().start() << ", end=" << output_window.z().end() + << ", step=" << output_window.z().step() << "]\n"; // Create an iterator: Iterator output_it(&output, output_window); @@ -126,13 +137,15 @@ public: // } // } // Except it works for an arbitrary number of dimensions - execute_window_loop(output_window, [&](const Coordinates & id) - { - std::cout << "Copying one row starting from [" << id.x() << "," << id.y() << "," << id.z() << "]\n"; - // Copy one whole row: - memcpy(dst_data + id.z() * (width * height) + id.y() * width, output_it.ptr(), width * sizeof(float)); - }, - output_it); + execute_window_loop( + output_window, + [&](const Coordinates &id) + { + std::cout << "Copying one row starting from [" << id.x() << "," << id.y() << "," << id.z() << "]\n"; + // Copy one whole row: + memcpy(dst_data + id.z() * (width * height) + id.y() * width, output_it.ptr(), width * sizeof(float)); + }, + output_it); /** [Copy objects example] */ diff --git a/examples/neon_gemm_qasymm8.cpp b/examples/neon_gemm_qasymm8.cpp index f6f5dc1026..3aaad02f8a 100644 --- a/examples/neon_gemm_qasymm8.cpp +++ b/examples/neon_gemm_qasymm8.cpp @@ -22,10 +22,11 @@ * SOFTWARE. */ #include "arm_compute/core/Types.h" -#include "arm_compute/core/WindowIterator.h" #include "arm_compute/core/utils/quantization/AsymmHelpers.h" +#include "arm_compute/core/WindowIterator.h" #include "arm_compute/runtime/NEON/NEFunctions.h" #include "arm_compute/runtime/NEON/NEScheduler.h" + #include "support/ToolchainSupport.h" #include "utils/Utils.h" @@ -38,7 +39,7 @@ using namespace utils; void find_min_max(int size, const float *data, float *min, float *max) { *min = *max = data[0]; - for(int i = 0; i < size; i++) + for (int i = 0; i < size; i++) { const float val = data[i]; *min = std::min(*min, val); @@ -66,11 +67,11 @@ QuantizationInfo choose_quantization_params(float min, float max) // But we need to nudge the zero_point to an integer (exact quantized value) std::uint8_t zero_point_nudged = 0; - if(zero_point_real < qmin) + if (zero_point_real < qmin) { zero_point_nudged = qmin; } - else if(zero_point_real > qmax) + else if (zero_point_real > qmax) { zero_point_nudged = qmax; } @@ -85,7 +86,7 @@ QuantizationInfo choose_quantization_params(float min, float max) void quantize_values(int size, qasymm8_t *output, float *input, const QuantizationInfo qinfo) { - for(int i = 0; i < size; i++) + for (int i = 0; i < size; i++) { output[i] = quantize_qasymm8(input[i], qinfo); } @@ -108,7 +109,7 @@ int main(int argc, char **argv) bool default_input = true; // Parse args - if(argc < 3) /* case default matrix sizes */ + if (argc < 3) /* case default matrix sizes */ { // Print help std::cout << "Usage: ./build/neon_gemm_qasymm8 M N K\n"; @@ -144,23 +145,23 @@ int main(int argc, char **argv) // Fill in: one is the identity matrix, other is sequential values // src1: Identity matrix - for(size_t i = 0; i < M * K; i++) + for (size_t i = 0; i < M * K; i++) { src1_ptr[i] = 0; } - for(size_t i = 0; i < M; i++) + for (size_t i = 0; i < M; i++) { src1_ptr[i * K + i] = 1.0f; } // src2: Sequential values matrix - for(size_t i = 0; i < K * N; i++) + for (size_t i = 0; i < K * N; i++) { src2_ptr[i] = i * 1.123f; } // Otherwise if M, N, K is given, fill in with random values - if(!default_input) + if (!default_input) { fill_random_tensor(src1, 0.f, 1.f); fill_random_tensor(src2, 0.f, 1.f); @@ -223,7 +224,7 @@ int main(int argc, char **argv) NEGEMMLowpOutputStage gemmlowp_output_stage; int output_multiplier; int output_shift; - float multiplier = (src1_qinfo.uniform().scale * src2_qinfo.uniform().scale) / dst0_qinfo.uniform().scale; + float multiplier = (src1_qinfo.uniform().scale * src2_qinfo.uniform().scale) / dst0_qinfo.uniform().scale; quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); std::cout << "(q_multiplier, q_shift) = (" << output_multiplier << ", " << output_shift << ")\n\n"; diff --git a/examples/neon_permute.cpp b/examples/neon_permute.cpp index 49848de4ea..76ba079430 100644 --- a/examples/neon_permute.cpp +++ b/examples/neon_permute.cpp @@ -21,9 +21,9 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ +#include "arm_compute/core/Types.h" #include "arm_compute/runtime/NEON/NEFunctions.h" -#include "arm_compute/core/Types.h" #include "utils/Utils.h" using namespace arm_compute; @@ -85,11 +85,13 @@ private: window.use_tensor_dimensions(reference.info()->tensor_shape()); Iterator ref_it(&reference, window); Iterator res_it(&result, window); - execute_window_loop(window, [&](const Coordinates &) - { - assert(*reinterpret_cast<unsigned char *>(ref_it.ptr()) == *reinterpret_cast<unsigned char *>(res_it.ptr())); - }, - ref_it, res_it); + execute_window_loop( + window, + [&](const Coordinates &) { + assert(*reinterpret_cast<unsigned char *>(ref_it.ptr()) == + *reinterpret_cast<unsigned char *>(res_it.ptr())); + }, + ref_it, res_it); } void fill_tensor(Tensor &tensor) @@ -98,11 +100,9 @@ private: window.use_tensor_dimensions(tensor.info()->tensor_shape()); Iterator tensor_it(&tensor, window); unsigned char val(0); - execute_window_loop(window, [&](const Coordinates &) - { - *reinterpret_cast<unsigned char *>(tensor_it.ptr()) = val++; - }, - tensor_it); + execute_window_loop( + window, [&](const Coordinates &) { *reinterpret_cast<unsigned char *>(tensor_it.ptr()) = val++; }, + tensor_it); } void init_tensor(const TensorShape shape, Tensor &tensor, DataType type, DataLayout layout) { diff --git a/examples/neon_scale.cpp b/examples/neon_scale.cpp index f120ea7f96..28590bd861 100644 --- a/examples/neon_scale.cpp +++ b/examples/neon_scale.cpp @@ -21,9 +21,9 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ +#include "arm_compute/core/Types.h" #include "arm_compute/runtime/NEON/NEFunctions.h" -#include "arm_compute/core/Types.h" #include "utils/ImageLoader.h" #include "utils/Utils.h" @@ -37,7 +37,7 @@ public: { PPMLoader ppm; - if(argc < 2) + if (argc < 2) { // Print help std::cout << "Usage: ./build/neon_scale[input_image.ppm]\n\n"; @@ -60,20 +60,16 @@ public: dst.allocator()->init(dst_tensor_info); // Configure Scale function object: - scale.configure(&src, &dst, ScaleKernelInfo{ - InterpolationPolicy::NEAREST_NEIGHBOR, - BorderMode::UNDEFINED, - PixelValue(), - SamplingPolicy::CENTER, - false - }); + scale.configure(&src, &dst, + ScaleKernelInfo{InterpolationPolicy::NEAREST_NEIGHBOR, BorderMode::UNDEFINED, PixelValue(), + SamplingPolicy::CENTER, false}); // Allocate all the images src.allocator()->allocate(); dst.allocator()->allocate(); // Fill the input image with the content of the PPM image if a filename was provided: - if(ppm.is_open()) + if (ppm.is_open()) { ppm.fill_image(src); output_filename = std::string(argv[1]) + "_out.ppm"; @@ -89,7 +85,7 @@ public: void do_teardown() override { // Save the result to file: - if(!output_filename.empty()) + if (!output_filename.empty()) { save_to_ppm(dst, output_filename); // save_to_ppm maps and unmaps the image to store as PPM } diff --git a/examples/neon_sgemm.cpp b/examples/neon_sgemm.cpp index 07696bd622..8cda65a400 100644 --- a/examples/neon_sgemm.cpp +++ b/examples/neon_sgemm.cpp @@ -24,6 +24,7 @@ #include "arm_compute/core/Types.h" #include "arm_compute/runtime/NEON/NEFunctions.h" #include "arm_compute/runtime/NEON/NEScheduler.h" + #include "utils/Utils.h" #include <cstdlib> @@ -43,15 +44,16 @@ public: beta = 0.0f; std::ifstream stream; - if(argc > 1) + if (argc > 1) { stream.open(argv[1], std::fstream::in); } - if(argc < 3 || (argc < 4 && stream.bad())) + if (argc < 3 || (argc < 4 && stream.bad())) { // Print help - std::cout << "Usage: 1) ./build/neon_sgemm input_matrix_1.npy input_matrix_2.npy [input_matrix_3.npy] [alpha = 1] [beta = 0]\n"; + std::cout << "Usage: 1) ./build/neon_sgemm input_matrix_1.npy input_matrix_2.npy [input_matrix_3.npy] " + "[alpha = 1] [beta = 0]\n"; std::cout << " 2) ./build/neon_sgemm M N K [alpha = 1.0f] [beta = 0.0f]\n\n"; std::cout << "Too few or no input_matrices provided. Using M=7, N=3, K=5, alpha=1.0f and beta=0.0f\n\n"; @@ -61,29 +63,29 @@ public: } else { - if(stream.good()) /* case file1.npy file2.npy [file3.npy] [alpha = 1.0f] [beta = 0.0f] */ + if (stream.good()) /* case file1.npy file2.npy [file3.npy] [alpha = 1.0f] [beta = 0.0f] */ { npy0.open(argv[1]); npy0.init_tensor(src0, DataType::F32); npy1.open(argv[2]); npy1.init_tensor(src1, DataType::F32); - if(argc > 3) + if (argc > 3) { stream.close(); stream.clear(); stream.open(argv[3], std::fstream::in); - if(stream.good()) /* case with third file */ + if (stream.good()) /* case with third file */ { npy2.open(argv[3]); npy2.init_tensor(src2, DataType::F32); - if(argc > 4) + if (argc > 4) { // Convert string to float alpha = strtof(argv[4], nullptr); - if(argc > 5) + if (argc > 5) { // Convert string to float beta = strtof(argv[5], nullptr); @@ -94,7 +96,7 @@ public: { alpha = strtof(argv[3], nullptr); - if(argc > 4) + if (argc > 4) { beta = strtof(argv[4], nullptr); } @@ -111,11 +113,11 @@ public: src1.allocator()->init(TensorInfo(TensorShape(N, K), 1, DataType::F32)); src2.allocator()->init(TensorInfo(TensorShape(N, M), 1, DataType::F32)); - if(argc > 4) + if (argc > 4) { alpha = strtof(argv[4], nullptr); - if(argc > 5) + if (argc > 5) { beta = strtof(argv[5], nullptr); } @@ -134,7 +136,7 @@ public: dst.allocator()->allocate(); // Fill the input images with either the data provided or random data - if(npy0.is_open()) + if (npy0.is_open()) { npy0.fill_tensor(src0); npy1.fill_tensor(src1); @@ -142,7 +144,7 @@ public: output_filename = "sgemm_out.npy"; is_fortran = npy0.is_fortran(); - if(npy2.is_open()) + if (npy2.is_open()) { src2.allocator()->allocate(); npy2.fill_tensor(src2); @@ -169,7 +171,7 @@ public: } void do_teardown() override { - if(!output_filename.empty()) /* Save to .npy file */ + if (!output_filename.empty()) /* Save to .npy file */ { save_to_npy(dst, output_filename, is_fortran); } |