From d8734b55d89f05901ba9a75349761a9c955d9243 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Fri, 22 Dec 2017 15:27:52 +0000 Subject: COMPMID-793 : Add graph intermediate representation Change-Id: Ic1685de4e19e0ac79669ef2da64e1dc96c7ea0bf Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/115248 Tested-by: Jenkins Reviewed-by: Anthony Barbier --- examples/SConscript | 8 +- examples/graph_googlenet.cpp | 34 ++--- examples/graph_inception_v3.cpp | 263 +++++++++++++++++++++---------------- examples/graph_inception_v4.cpp | 184 +++++++++++++------------- examples/graph_lenet.cpp | 22 ++-- examples/graph_mobilenet.cpp | 41 +++--- examples/graph_resnet50.cpp | 41 +++--- examples/graph_squeezenet.cpp | 32 ++--- examples/graph_squeezenet_v1_1.cpp | 29 ++-- examples/graph_vgg16.cpp | 31 ++--- examples/graph_vgg19.cpp | 25 ++-- 11 files changed, 378 insertions(+), 332 deletions(-) (limited to 'examples') diff --git a/examples/SConscript b/examples/SConscript index 9be9fa9d9a..80bce57316 100644 --- a/examples/SConscript +++ b/examples/SConscript @@ -57,15 +57,17 @@ if env['opencl'] and env['neon']: alias = examples_env.Alias(example, prog) Default(alias) if env['os'] == 'android': + Import('arm_compute_graph2_a') Import('arm_compute_graph_a') Import('arm_compute_core_a') Import('arm_compute_a') arm_compute_graph_libs = [ arm_compute_a, arm_compute_core_a, "OpenCL"] - graph_dependency = arm_compute_graph_a + graph_dependency = [arm_compute_graph_a, arm_compute_graph2_a] else: + Import('arm_compute_graph2_so') Import('arm_compute_graph_so') - arm_compute_graph_libs = ["arm_compute_graph", "arm_compute", "arm_compute_core"] - graph_dependency = arm_compute_graph_so + arm_compute_graph_libs = ["arm_compute_graph2", "arm_compute_graph", "arm_compute", "arm_compute_core"] + graph_dependency = [arm_compute_graph_so, arm_compute_graph2_so] graph_utils = examples_env.Object("../utils/GraphUtils.cpp") for file in Glob("./graph_*.cpp"): diff --git a/examples/graph_googlenet.cpp b/examples/graph_googlenet.cpp index de4afa29ea..d64512bb96 100644 --- a/examples/graph_googlenet.cpp +++ b/examples/graph_googlenet.cpp @@ -21,9 +21,7 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/graph/Graph.h" -#include "arm_compute/graph/Nodes.h" -#include "arm_compute/graph/SubGraph.h" +#include "arm_compute/graph2.h" #include "support/ToolchainSupport.h" #include "utils/GraphUtils.h" #include "utils/Utils.h" @@ -32,7 +30,7 @@ #include using namespace arm_compute::utils; -using namespace arm_compute::graph; +using namespace arm_compute::graph2::frontend; using namespace arm_compute::graph_utils; /** Example demonstrating how to implement Googlenet's network using the Compute Library's graph API @@ -54,9 +52,11 @@ public: std::unique_ptr preprocessor = arm_compute::support::cpp14::make_unique(mean_rgb); // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON - const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; - TargetHint target_hint = set_target_hint(int_target_hint); - ConvolutionMethodHint convolution_hint = ConvolutionMethodHint::GEMM; + const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; + Target target_hint = set_target_hint2(target); + ConvolutionMethod convolution_hint = ConvolutionMethod::GEMM; + bool enable_tuning = (target == 2); + bool enable_memory_management = true; // Parse arguments if(argc < 2) @@ -91,8 +91,8 @@ public: } graph << target_hint - << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32), - get_input_accessor(image, std::move(preprocessor))) + << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32), + get_input_accessor(image, std::move(preprocessor))) << ConvolutionLayer( 7U, 7U, 64U, get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_w.npy"), @@ -133,10 +133,10 @@ public: get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_w.npy"), get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_b.npy")) << SoftmaxLayer() - << Tensor(get_output_accessor(label, 5)); + << OutputLayer(get_output_accessor(label, 5)); - // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated - graph.graph_init(int_target_hint == 2); + // Finalize graph + graph.finalize(target_hint, enable_tuning, enable_memory_management); } void do_run() override { @@ -145,7 +145,7 @@ public: } private: - Graph graph{}; + Stream graph{ 0, "GoogleNet" }; BranchLayer get_inception_node(const std::string &data_path, std::string &¶m_path, unsigned int a_filt, @@ -154,7 +154,7 @@ private: unsigned int d_filt) { std::string total_path = "/cnn_data/googlenet_model/" + param_path + "/" + param_path + "_"; - SubGraph i_a; + SubStream i_a(graph); i_a << ConvolutionLayer( 1U, 1U, a_filt, get_weights_accessor(data_path, total_path + "1x1_w.npy"), @@ -162,7 +162,7 @@ private: PadStrideInfo(1, 1, 0, 0)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b; + 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"), @@ -176,7 +176,7 @@ private: PadStrideInfo(1, 1, 1, 1)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_c; + 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"), @@ -190,7 +190,7 @@ private: PadStrideInfo(1, 1, 2, 2)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_d; + SubStream i_d(graph); i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL))) << ConvolutionLayer( 1U, 1U, d_filt, diff --git a/examples/graph_inception_v3.cpp b/examples/graph_inception_v3.cpp index a10037be89..9bb51bad44 100644 --- a/examples/graph_inception_v3.cpp +++ b/examples/graph_inception_v3.cpp @@ -21,9 +21,7 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/graph/Graph.h" -#include "arm_compute/graph/Nodes.h" -#include "arm_compute/graph/SubGraph.h" +#include "arm_compute/graph2.h" #include "support/ToolchainSupport.h" #include "utils/GraphUtils.h" #include "utils/Utils.h" @@ -32,15 +30,15 @@ #include using namespace arm_compute::utils; -using namespace arm_compute::graph; +using namespace arm_compute::graph2::frontend; using namespace arm_compute::graph_utils; /** Example demonstrating how to implement InceptionV3's network using the Compute Library's graph API * * @param[in] argc Number of arguments - * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels ) + * @param[in] argv Arguments ( [optional] Path to the weights folder, [optional] image, [optional] labels ) */ -class InceptionV3Example final : public Example +class InceptionV3Example : public Example { public: void do_setup(int argc, char **argv) override @@ -53,8 +51,10 @@ public: std::unique_ptr preprocessor = arm_compute::support::cpp14::make_unique(); // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON - const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; - TargetHint target_hint = set_target_hint(int_target_hint); + const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; + Target target_hint = set_target_hint2(target); + bool enable_tuning = (target == 2); + bool enable_memory_management = true; // Parse arguments if(argc < 2) @@ -88,8 +88,8 @@ public: label = argv[4]; } - graph << target_hint << Tensor(TensorInfo(TensorShape(299U, 299U, 3U, 1U), 1, DataType::F32), - get_input_accessor(image, std::move(preprocessor), false)) + graph << target_hint << InputLayer(TensorDescriptor(TensorShape(299U, 299U, 3U, 1U), DataType::F32), + get_input_accessor(image, std::move(preprocessor), false)) << ConvolutionLayer(3U, 3U, 32U, get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_1a_3x3_weights.npy"), @@ -100,7 +100,8 @@ public: "/cnn_data/inceptionv3_model/Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_1a_3x3_BatchNorm_beta.npy"), - 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(3U, 3U, 32U, get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2a_3x3_weights.npy"), @@ -111,7 +112,8 @@ public: "/cnn_data/inceptionv3_model/Conv2d_2a_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2a_3x3_BatchNorm_beta.npy"), - 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(3U, 3U, 64U, get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2b_3x3_weights.npy"), @@ -122,7 +124,8 @@ public: "/cnn_data/inceptionv3_model/Conv2d_2b_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2b_3x3_BatchNorm_beta.npy"), - 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) @@ -135,7 +138,8 @@ public: "/cnn_data/inceptionv3_model/Conv2d_3b_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_3b_1x1_BatchNorm_beta.npy"), - 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(3U, 3U, 192U, get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_4a_3x3_weights.npy"), @@ -146,7 +150,8 @@ public: "/cnn_data/inceptionv3_model/Conv2d_4a_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_4a_3x3_BatchNorm_beta.npy"), - 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) @@ -183,10 +188,10 @@ public: "/cnn_data/inceptionv3_model/Logits_Conv2d_1c_1x1_biases.npy"), PadStrideInfo(1, 1, 0, 0)) << ReshapeLayer(TensorShape(1001U)) << SoftmaxLayer() - << Tensor(get_output_accessor(label, 5)); + << OutputLayer(get_output_accessor(label, 5)); - // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated - graph.graph_init(int_target_hint == 2); + // Finalize graph + graph.finalize(target_hint, enable_tuning, enable_memory_management); } void do_run() override @@ -195,7 +200,7 @@ public: } private: - Graph graph{}; + Stream graph{ 0, "InceptionV3" }; private: BranchLayer get_inception_node_A(const std::string &data_path, std::string &¶m_path, @@ -216,7 +221,7 @@ private: conv_id1 = "_1_0c_"; } - SubGraph i_a; + 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"), @@ -227,9 +232,10 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b; + 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"), @@ -240,7 +246,8 @@ private: get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + "1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + "1x1_BatchNorm_beta.npy"), - 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 5U, 5U, std::get<1>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_weights.npy"), @@ -251,9 +258,10 @@ private: get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_BatchNorm_beta.npy"), - 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_c; + 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"), @@ -264,7 +272,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 3U, 3U, std::get<1>(c_filters), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy"), @@ -275,7 +284,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 3U, 3U, std::get<2>(c_filters), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_weights.npy"), @@ -286,9 +296,10 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_d; + SubStream i_d(graph); i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) << ConvolutionLayer( 1U, 1U, d_filt, @@ -300,7 +311,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)); } @@ -310,7 +322,7 @@ private: std::tuple b_filters) { std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_"; - SubGraph i_a; + 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"), @@ -321,9 +333,10 @@ private: get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_BatchNorm_beta.npy"), - 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b; + 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"), @@ -334,7 +347,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 3U, 3U, std::get<1>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy"), @@ -345,7 +359,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 3U, 3U, std::get<2>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_weights.npy"), @@ -356,12 +371,11 @@ private: get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_BatchNorm_beta.npy"), - 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_c; - i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) - // TODO (geopin01) : Remove once we understand why a single node graph does not run in CL - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f)); + SubStream i_c(graph); + i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))); return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c)); } @@ -373,7 +387,7 @@ private: unsigned int d_filt) { std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_"; - SubGraph i_a; + 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"), @@ -384,9 +398,10 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b; + 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"), @@ -397,7 +412,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 7U, 1U, std::get<1>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"), @@ -408,7 +424,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 1U, 7U, std::get<2>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"), @@ -419,9 +436,10 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_c; + 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"), @@ -432,7 +450,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 1U, 7U, std::get<1>(c_filters), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_weights.npy"), @@ -443,7 +462,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 7U, 1U, std::get<2>(c_filters), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_weights.npy"), @@ -454,7 +474,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 1U, 7U, std::get<3>(c_filters), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_weights.npy"), @@ -465,7 +486,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 7U, 1U, std::get<4>(c_filters), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_weights.npy"), @@ -476,9 +498,10 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_d; + SubStream i_d(graph); i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) << ConvolutionLayer( 1U, 1U, d_filt, @@ -490,7 +513,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)); } @@ -500,7 +524,7 @@ private: std::tuple b_filters) { std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_"; - SubGraph i_a; + 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"), @@ -511,7 +535,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 3U, 3U, std::get<1>(a_filters), get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"), @@ -522,9 +547,10 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b; + 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"), @@ -535,7 +561,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 7U, 1U, std::get<1>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"), @@ -546,7 +573,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 1U, 7U, std::get<2>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"), @@ -557,7 +585,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 3U, 3U, std::get<3>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy"), @@ -568,12 +597,11 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_c; - i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) - // TODO (geopin01) : Remove once we understand why a single node graph does not run in CL - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f)); + SubStream i_c(graph); + i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))); return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c)); } @@ -593,7 +621,7 @@ private: } std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_"; - SubGraph i_a; + 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"), @@ -604,9 +632,24 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::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"), + std::unique_ptr(nullptr), + PadStrideInfo(1, 1, 0, 0)) + << 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) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b1; + SubStream i_b1(static_cast(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"), @@ -617,9 +660,10 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b2; + SubStream i_b2(static_cast(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"), @@ -630,23 +674,39 @@ private: get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_BatchNorm_beta.npy"), - 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b; - i_b << ConvolutionLayer( - 1U, 1U, std::get<0>(b_filters), - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), + // Merge b1 and b2 + i_b << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_b1), std::move(i_b2)); + + 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"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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_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_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), - 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) - << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_b1), std::move(i_b2)); + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + << ConvolutionLayer( + 3U, 3U, std::get<1>(c_filters), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy"), + std::unique_ptr(nullptr), + PadStrideInfo(1, 1, 1, 1)) + << 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) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_c1; + SubStream i_c1(static_cast(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"), @@ -657,9 +717,10 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_c2; + SubStream i_c2(static_cast(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"), @@ -670,34 +731,13 @@ private: get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_beta.npy"), - 0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_c; - i_c << ConvolutionLayer( - 1U, 1U, std::get<0>(c_filters), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"), - std::unique_ptr(nullptr), - PadStrideInfo(1, 1, 0, 0)) - << 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) - << ConvolutionLayer( - 3U, 3U, std::get<1>(c_filters), - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy"), - std::unique_ptr(nullptr), - PadStrideInfo(1, 1, 1, 1)) - << 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) - << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_c1), std::move(i_c2)); + // Merge i_c1 and i_c2 + i_c << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_c1), std::move(i_c2)); - SubGraph i_d; + SubStream i_d(graph); i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) << ConvolutionLayer( 1U, 1U, d_filt, @@ -709,7 +749,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, 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 f004b41fb0..d9f6156fb2 100644 --- a/examples/graph_inception_v4.cpp +++ b/examples/graph_inception_v4.cpp @@ -21,9 +21,7 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/graph/Graph.h" -#include "arm_compute/graph/Nodes.h" -#include "arm_compute/graph/SubGraph.h" +#include "arm_compute/graph2.h" #include "support/ToolchainSupport.h" #include "utils/GraphUtils.h" #include "utils/Utils.h" @@ -32,7 +30,7 @@ #include using namespace arm_compute::utils; -using namespace arm_compute::graph; +using namespace arm_compute::graph2::frontend; using namespace arm_compute::graph_utils; /** Example demonstrating how to implement InceptionV4's network using the Compute Library's graph API @@ -52,9 +50,11 @@ public: // Create a preprocessor object std::unique_ptr preprocessor = arm_compute::support::cpp14::make_unique(); - // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON - const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; - TargetHint target_hint = set_target_hint(int_target_hint); + // Set target. 0 (NEON), 1 (OpenCL). By default it is NEON + const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; + Target target_hint = set_target_hint2(target); + bool enable_tuning = (target == 2); + bool enable_memory_management = true; // Parse arguments if(argc < 2) @@ -88,8 +88,8 @@ public: label = argv[4]; } - graph << target_hint << Tensor(TensorInfo(TensorShape(299U, 299U, 3U, 1U), 1, DataType::F32), - get_input_accessor(image, std::move(preprocessor), false)) + graph << target_hint << InputLayer(TensorDescriptor(TensorShape(299U, 299U, 3U, 1U), DataType::F32), + get_input_accessor(image, std::move(preprocessor), false)) // Conv2d_1a_3x3 << ConvolutionLayer(3U, 3U, 32U, @@ -153,10 +153,10 @@ public: get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_weights.npy"), get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_biases.npy")) << SoftmaxLayer() - << Tensor(get_output_accessor(label, 5)); + << OutputLayer(get_output_accessor(label, 5)); - // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated - graph.graph_init(int_target_hint == 2); + // Finalize graph + graph.finalize(target_hint, enable_tuning, enable_memory_management); } void do_run() override @@ -165,19 +165,17 @@ public: } private: - Graph graph{}; + Stream graph{ 0, "InceptionV4" }; private: BranchLayer get_mixed_3a(const std::string &data_path) { std::string total_path = "/cnn_data/inceptionv4_model/Mixed_3a_"; - SubGraph i_a; - i_a << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)) - // TODO (geopin01) : Remove once we understand why a single node graph does not run in CL - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f)); + SubStream i_a(graph); + i_a << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)); - SubGraph i_b; + SubStream i_b(graph); i_b << ConvolutionLayer(3U, 3U, 96U, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(2, 2, 0, 0)) @@ -195,7 +193,7 @@ private: { std::string total_path = "/cnn_data/inceptionv4_model/Mixed_4a_"; - SubGraph i_a; + SubStream i_a(graph); i_a << ConvolutionLayer(1U, 1U, 64U, get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) @@ -215,7 +213,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b; + SubStream i_b(graph); i_b << ConvolutionLayer(1U, 1U, 64U, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) @@ -260,7 +258,7 @@ private: { std::string total_path = "/cnn_data/inceptionv4_model/Mixed_5a_"; - SubGraph i_a; + SubStream i_a(graph); i_a << ConvolutionLayer(3U, 3U, 192U, get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(2, 2, 0, 0)) @@ -271,10 +269,8 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b; - i_b << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)) - // TODO (geopin01) : Remove once we understand why a single node graph does not run in CL - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f)); + SubStream i_b(graph); + i_b << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)); return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b)); } @@ -283,7 +279,7 @@ private: { std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_"; - SubGraph i_a; + SubStream i_a(graph); i_a << ConvolutionLayer(1U, 1U, 96U, get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) @@ -294,7 +290,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b; + SubStream i_b(graph); i_b << ConvolutionLayer(1U, 1U, 64U, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) @@ -314,7 +310,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_c; + SubStream i_c(graph); i_c << ConvolutionLayer(1U, 1U, 64U, get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) @@ -343,7 +339,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_d; + SubStream i_d(graph); i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) << ConvolutionLayer(1U, 1U, 96U, get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"), @@ -362,7 +358,7 @@ private: { std::string total_path = "/cnn_data/inceptionv4_model/Mixed_6a_"; - SubGraph i_a; + SubStream i_a(graph); i_a << ConvolutionLayer(3U, 3U, 384U, get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(2, 2, 0, 0)) @@ -373,7 +369,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b; + SubStream i_b(graph); i_b << ConvolutionLayer(1U, 1U, 192U, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) @@ -402,10 +398,9 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_c; - i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)) - // TODO (geopin01) : Remove once we understand why a single node graph does not run in CL - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f)); + SubStream i_c(graph); + i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)); + return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c)); } @@ -413,7 +408,7 @@ private: { std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_"; - SubGraph i_a; + SubStream i_a(graph); i_a << ConvolutionLayer(1U, 1U, 384U, get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) @@ -424,7 +419,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b; + SubStream i_b(graph); i_b << ConvolutionLayer(1U, 1U, 192U, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) @@ -453,7 +448,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_c; + SubStream i_c(graph); i_c << ConvolutionLayer(1U, 1U, 192U, get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) @@ -500,7 +495,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_d; + SubStream i_d(graph); i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) << ConvolutionLayer(1U, 1U, 128U, get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"), @@ -519,7 +514,7 @@ private: { std::string total_path = "/cnn_data/inceptionv4_model/Mixed_7a_"; - SubGraph i_a; + SubStream i_a(graph); i_a << ConvolutionLayer(1U, 1U, 192U, get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) @@ -539,7 +534,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b; + SubStream i_b(graph); i_b << ConvolutionLayer(1U, 1U, 256U, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) @@ -577,10 +572,9 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_c; - i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)) - // TODO (geopin01) : Remove once we understand why a single node graph does not run in CL - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f)); + SubStream i_c(graph); + i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)); + return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c)); } @@ -588,7 +582,7 @@ private: { std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_"; - SubGraph i_a; + SubStream i_a(graph); i_a << ConvolutionLayer(1U, 1U, 256U, get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) @@ -599,35 +593,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b1; - i_b1 << ConvolutionLayer( - 3U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_weights.npy"), - std::unique_ptr(nullptr), - PadStrideInfo(1, 1, 1, 0)) - << 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) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - - SubGraph i_b2; - i_b2 << ConvolutionLayer( - 1U, 3U, 256U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_weights.npy"), - std::unique_ptr(nullptr), - PadStrideInfo(1, 1, 0, 1)) - << 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) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - - SubGraph i_b; + SubStream i_b(graph); i_b << ConvolutionLayer( 1U, 1U, 384U, get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), @@ -639,38 +605,40 @@ private: get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) - << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_b1), std::move(i_b2)); + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_c1; - i_c1 << ConvolutionLayer( + SubStream i_b1(static_cast(i_b)); + i_b1 << ConvolutionLayer( 3U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 1, 0)) << 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_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_2_Conv2d_0d_1x3_BatchNorm_beta.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"), 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_c2; - i_c2 << ConvolutionLayer( + SubStream i_b2(static_cast(i_b)); + i_b2 << ConvolutionLayer( 1U, 3U, 256U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 1)) << 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_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_2_Conv2d_0e_3x1_BatchNorm_beta.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_beta.npy"), 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_c; + // Merge b1 and b2 + i_b << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_b1), std::move(i_b2)); + + SubStream i_c(graph); i_c << ConvolutionLayer( 1U, 1U, 384U, get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"), @@ -706,10 +674,40 @@ private: get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) - << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_c1), std::move(i_c2)); + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + + SubStream i_c1(static_cast(i_c)); + i_c1 << ConvolutionLayer( + 3U, 1U, 256U, + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_weights.npy"), + std::unique_ptr(nullptr), + PadStrideInfo(1, 1, 1, 0)) + << 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) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + + SubStream i_c2(static_cast(i_c)); + i_c2 << ConvolutionLayer( + 1U, 3U, 256U, + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_weights.npy"), + std::unique_ptr(nullptr), + PadStrideInfo(1, 1, 0, 1)) + << 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) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + + // Merge i_c1 and i_c2 + i_c << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_c1), std::move(i_c2)); - SubGraph i_d; + SubStream i_d(graph); i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) << ConvolutionLayer(1U, 1U, 256U, get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"), diff --git a/examples/graph_lenet.cpp b/examples/graph_lenet.cpp index 61bc7bd3bf..e4b8effe5d 100644 --- a/examples/graph_lenet.cpp +++ b/examples/graph_lenet.cpp @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/graph/Graph.h" -#include "arm_compute/graph/Nodes.h" +#include "arm_compute/graph2.h" + #include "support/ToolchainSupport.h" #include "utils/GraphUtils.h" #include "utils/Utils.h" @@ -30,7 +30,7 @@ #include using namespace arm_compute::utils; -using namespace arm_compute::graph; +using namespace arm_compute::graph2::frontend; using namespace arm_compute::graph_utils; /** Example demonstrating how to implement LeNet's network using the Compute Library's graph API @@ -47,8 +47,10 @@ public: unsigned int batches = 4; /** Number of batches */ // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON - const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; - TargetHint target_hint = set_target_hint(int_target_hint); + const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; + Target target_hint = set_target_hint2(target); + bool enable_tuning = (target == 2); + bool enable_memory_management = true; // Parse arguments if(argc < 2) @@ -78,7 +80,7 @@ public: //conv1 << pool1 << conv2 << pool2 << fc1 << act1 << fc2 << smx graph << target_hint - << Tensor(TensorInfo(TensorShape(28U, 28U, 1U, batches), 1, DataType::F32), DummyAccessor()) + << InputLayer(TensorDescriptor(TensorShape(28U, 28U, 1U, batches), DataType::F32), get_input_accessor("")) << ConvolutionLayer( 5U, 5U, 20U, get_weights_accessor(data_path, "/cnn_data/lenet_model/conv1_w.npy"), @@ -101,10 +103,10 @@ public: get_weights_accessor(data_path, "/cnn_data/lenet_model/ip2_w.npy"), get_weights_accessor(data_path, "/cnn_data/lenet_model/ip2_b.npy")) << SoftmaxLayer() - << Tensor(DummyAccessor(0)); + << OutputLayer(get_output_accessor("")); - // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated - graph.graph_init(int_target_hint == 2); + // Finalize graph + graph.finalize(target_hint, enable_tuning, enable_memory_management); } void do_run() override { @@ -113,7 +115,7 @@ public: } private: - Graph graph{}; + Stream graph{ 0, "LeNet" }; }; /** Main program for LeNet diff --git a/examples/graph_mobilenet.cpp b/examples/graph_mobilenet.cpp index 1a930dd950..4d01055c50 100644 --- a/examples/graph_mobilenet.cpp +++ b/examples/graph_mobilenet.cpp @@ -21,8 +21,7 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/graph/Graph.h" -#include "arm_compute/graph/Nodes.h" +#include "arm_compute/graph2.h" #include "support/ToolchainSupport.h" #include "utils/GraphUtils.h" #include "utils/Utils.h" @@ -30,7 +29,7 @@ #include using namespace arm_compute::utils; -using namespace arm_compute::graph; +using namespace arm_compute::graph2::frontend; using namespace arm_compute::graph_utils; /** Example demonstrating how to implement MobileNet's network using the Compute Library's graph API @@ -51,9 +50,12 @@ public: std::unique_ptr preprocessor = arm_compute::support::cpp14::make_unique(); // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON - const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; - TargetHint target_hint = set_target_hint(int_target_hint); - ConvolutionMethodHint convolution_hint = ConvolutionMethodHint::GEMM; + const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; + Target target_hint = set_target_hint2(target); + ConvolutionMethod convolution_hint = ConvolutionMethod::GEMM; + DepthwiseConvolutionMethod depthwise_convolution_hint = DepthwiseConvolutionMethod::OPTIMIZED_3x3; + bool enable_tuning = (target == 2); + bool enable_memory_management = true; // Set model to execute. 0 (MobileNetV1_1.0_224), 1 (MobileNetV1_0.75_160) int model_id = (argc > 2) ? std::strtol(argv[2], nullptr, 10) : 0; @@ -109,8 +111,9 @@ public: graph << target_hint << convolution_hint - << Tensor(TensorInfo(TensorShape(spatial_size, spatial_size, 3U, 1U), 1, DataType::F32), - get_input_accessor(image, std::move(preprocessor), false)) + << depthwise_convolution_hint + << InputLayer(TensorDescriptor(TensorShape(spatial_size, spatial_size, 3U, 1U), DataType::F32), + get_input_accessor(image, std::move(preprocessor), false)) << ConvolutionLayer( 3U, 3U, 32U * depth_scale, get_weights_accessor(data_path, "Conv2d_0_weights.npy"), @@ -121,7 +124,8 @@ public: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) << get_dwsc_node(data_path, "Conv2d_1", 64 * depth_scale, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0)) << get_dwsc_node(data_path, "Conv2d_2", 128 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)) << get_dwsc_node(data_path, "Conv2d_3", 128 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)) @@ -143,10 +147,10 @@ public: PadStrideInfo(1, 1, 0, 0)) << ReshapeLayer(TensorShape(1001U)) << SoftmaxLayer() - << Tensor(get_output_accessor(label, 5)); + << OutputLayer(get_output_accessor(label, 5)); - // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated - graph.graph_init(int_target_hint == 2); + // Finalize graph + graph.finalize(target_hint, enable_tuning, enable_memory_management); } void do_run() override { @@ -155,26 +159,26 @@ public: } private: - Graph graph{}; + Stream graph{ 0, "MobileNetV1" }; BranchLayer get_dwsc_node(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 + "_"; - SubGraph sg; + SubStream sg(graph); sg << DepthwiseConvolutionLayer( 3U, 3U, get_weights_accessor(data_path, total_path + "depthwise_depthwise_weights.npy"), std::unique_ptr(nullptr), - dwc_pad_stride_info, - true) + dwc_pad_stride_info) << 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) << ConvolutionLayer( 1U, 1U, conv_filt, get_weights_accessor(data_path, total_path + "pointwise_weights.npy"), @@ -185,7 +189,8 @@ private: 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, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)); + 0.001f) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)); return BranchLayer(std::move(sg)); } diff --git a/examples/graph_resnet50.cpp b/examples/graph_resnet50.cpp index e4d31f98d7..90debb4293 100644 --- a/examples/graph_resnet50.cpp +++ b/examples/graph_resnet50.cpp @@ -21,8 +21,7 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/graph/Graph.h" -#include "arm_compute/graph/Nodes.h" +#include "arm_compute/graph2.h" #include "support/ToolchainSupport.h" #include "utils/GraphUtils.h" #include "utils/Utils.h" @@ -30,7 +29,7 @@ #include using namespace arm_compute::utils; -using namespace arm_compute::graph; +using namespace arm_compute::graph2::frontend; using namespace arm_compute::graph_utils; /** Example demonstrating how to implement Microsoft's ResNet50 network using the Compute Library's graph API @@ -53,8 +52,10 @@ public: false /* Do not convert to BGR */); // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON - const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; - TargetHint target_hint = set_target_hint(int_target_hint); + const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; + Target target_hint = set_target_hint2(target); + bool enable_tuning = (target == 2); + bool enable_memory_management = true; // Parse arguments if(argc < 2) @@ -89,8 +90,8 @@ public: } graph << target_hint - << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32), - get_input_accessor(image, std::move(preprocessor), false /* Do not convert to BGR */)) + << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32), + get_input_accessor(image, 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"), @@ -118,11 +119,12 @@ public: PadStrideInfo(1, 1, 0, 0)) << FlattenLayer() << SoftmaxLayer() - << Tensor(get_output_accessor(label, 5)); + << OutputLayer(get_output_accessor(label, 5)); - // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated - graph.graph_init(int_target_hint == 2); + // Finalize graph + graph.finalize(target_hint, enable_tuning, enable_memory_management); } + void do_run() override { // Run graph @@ -130,7 +132,7 @@ public: } private: - Graph graph{}; + Stream graph{ 0, "ResNet50" }; void add_residual_block(const std::string &data_path, const std::string &name, unsigned int base_depth, unsigned int num_units, unsigned int stride) { @@ -147,7 +149,7 @@ private: middle_stride = stride; } - SubGraph right; + SubStream right(graph); right << ConvolutionLayer( 1U, 1U, base_depth, get_weights_accessor(data_path, unit_name + "conv1_weights.npy"), @@ -188,7 +190,7 @@ private: if(i == 0) { - SubGraph left; + SubStream left(graph); left << ConvolutionLayer( 1U, 1U, base_depth * 4, get_weights_accessor(data_path, unit_name + "shortcut_weights.npy"), @@ -201,20 +203,19 @@ private: get_weights_accessor(data_path, unit_name + "shortcut_BatchNorm_beta.npy"), 0.0000100099996416f); - graph << ResidualLayer(std::move(left), std::move(right)); + graph << BranchLayer(BranchMergeMethod::ADD, std::move(left), std::move(right)); } else if(middle_stride > 1) { - SubGraph left; - left << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 1, PadStrideInfo(middle_stride, middle_stride, 0, 0), true)) - // TODO (alegil01) : Remove once we understand why a single node graph does not run in CL - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f)); + SubStream left(graph); + left << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 1, PadStrideInfo(middle_stride, middle_stride, 0, 0), true)); - graph << ResidualLayer(std::move(left), std::move(right)); + graph << BranchLayer(BranchMergeMethod::ADD, std::move(left), std::move(right)); } else { - graph << ResidualLayer(std::move(right)); + SubStream left(graph); + graph << BranchLayer(BranchMergeMethod::ADD, std::move(left), std::move(right)); } graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); diff --git a/examples/graph_squeezenet.cpp b/examples/graph_squeezenet.cpp index d0c823a11c..b4e00a451b 100644 --- a/examples/graph_squeezenet.cpp +++ b/examples/graph_squeezenet.cpp @@ -21,9 +21,7 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/graph/Graph.h" -#include "arm_compute/graph/Nodes.h" -#include "arm_compute/graph/SubGraph.h" +#include "arm_compute/graph2.h" #include "support/ToolchainSupport.h" #include "utils/GraphUtils.h" #include "utils/Utils.h" @@ -32,14 +30,10 @@ #include using namespace arm_compute::utils; -using namespace arm_compute::graph; +using namespace arm_compute::graph2::frontend; using namespace arm_compute::graph_utils; using namespace arm_compute::logging; -namespace -{ -} // namespace - /** Example demonstrating how to implement Squeezenet's network using the Compute Library's graph API * * @param[in] argc Number of arguments @@ -59,8 +53,10 @@ public: std::unique_ptr preprocessor = arm_compute::support::cpp14::make_unique(mean_rgb); // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON - const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; - TargetHint target_hint = set_target_hint(int_target_hint); + const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; + Target target_hint = set_target_hint2(target); + bool enable_tuning = (target == 2); + bool enable_memory_management = true; // Parse arguments if(argc < 2) @@ -95,8 +91,8 @@ public: } graph << target_hint - << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32), - get_input_accessor(image, std::move(preprocessor))) + << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32), + get_input_accessor(image, std::move(preprocessor))) << ConvolutionLayer( 7U, 7U, 96U, get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_w.npy"), @@ -171,10 +167,10 @@ public: << PoolingLayer(PoolingLayerInfo(PoolingType::AVG)) << FlattenLayer() << SoftmaxLayer() - << Tensor(get_output_accessor(label, 5)); + << OutputLayer(get_output_accessor(label, 5)); - // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated - graph.graph_init(int_target_hint == 2); + // Finalize graph + graph.finalize(target_hint, enable_tuning, enable_memory_management); } void do_run() override { @@ -183,12 +179,12 @@ public: } private: - Graph graph{}; + Stream graph{ 0, "SqueezeNetV1" }; BranchLayer get_expand_fire_node(const std::string &data_path, std::string &¶m_path, unsigned int expand1_filt, unsigned int expand3_filt) { std::string total_path = "/cnn_data/squeezenet_v1.0_model/" + param_path + "_"; - SubGraph i_a; + SubStream i_a(graph); i_a << ConvolutionLayer( 1U, 1U, expand1_filt, get_weights_accessor(data_path, total_path + "expand1x1_w.npy"), @@ -196,7 +192,7 @@ private: PadStrideInfo(1, 1, 0, 0)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b; + SubStream i_b(graph); i_b << ConvolutionLayer( 3U, 3U, expand3_filt, get_weights_accessor(data_path, total_path + "expand3x3_w.npy"), diff --git a/examples/graph_squeezenet_v1_1.cpp b/examples/graph_squeezenet_v1_1.cpp index 189cc027fd..4ebfd3fe8e 100644 --- a/examples/graph_squeezenet_v1_1.cpp +++ b/examples/graph_squeezenet_v1_1.cpp @@ -21,9 +21,7 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/graph/Graph.h" -#include "arm_compute/graph/Nodes.h" -#include "arm_compute/graph/SubGraph.h" +#include "arm_compute/graph2.h" #include "support/ToolchainSupport.h" #include "utils/GraphUtils.h" #include "utils/Utils.h" @@ -32,9 +30,8 @@ #include using namespace arm_compute::utils; -using namespace arm_compute::graph; +using namespace arm_compute::graph2::frontend; using namespace arm_compute::graph_utils; -using namespace arm_compute::logging; namespace { @@ -59,8 +56,10 @@ public: std::unique_ptr preprocessor = arm_compute::support::cpp14::make_unique(mean_rgb); // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON - const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; - TargetHint target_hint = set_target_hint(int_target_hint); + const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; + Target target_hint = set_target_hint2(target); + bool enable_tuning = (target == 2); + bool enable_memory_management = true; // Parse arguments if(argc < 2) @@ -95,8 +94,8 @@ public: } graph << target_hint - << Tensor(TensorInfo(TensorShape(227U, 227U, 3U, 1U), 1, DataType::F32), - get_input_accessor(image, std::move(preprocessor))) + << InputLayer(TensorDescriptor(TensorShape(227U, 227U, 3U, 1U), DataType::F32), + get_input_accessor(image, std::move(preprocessor))) << ConvolutionLayer( 3U, 3U, 64U, get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv1_w.npy"), @@ -171,10 +170,10 @@ public: << PoolingLayer(PoolingLayerInfo(PoolingType::AVG)) << FlattenLayer() << SoftmaxLayer() - << Tensor(get_output_accessor(label, 5)); + << OutputLayer(get_output_accessor(label, 5)); - // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated - graph.graph_init(int_target_hint == 2); + // Finalize graph + graph.finalize(target_hint, enable_tuning, enable_memory_management); } void do_run() override { @@ -183,12 +182,12 @@ public: } private: - Graph graph{}; + Stream graph{ 0, "SqueezeNetV1.1" }; BranchLayer get_expand_fire_node(const std::string &data_path, std::string &¶m_path, unsigned int expand1_filt, unsigned int expand3_filt) { std::string total_path = "/cnn_data/squeezenet_v1_1_model/" + param_path + "_"; - SubGraph i_a; + SubStream i_a(graph); i_a << ConvolutionLayer( 1U, 1U, expand1_filt, get_weights_accessor(data_path, total_path + "expand1x1_w.npy"), @@ -196,7 +195,7 @@ private: PadStrideInfo(1, 1, 0, 0)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - SubGraph i_b; + SubStream i_b(graph); i_b << ConvolutionLayer( 3U, 3U, expand3_filt, get_weights_accessor(data_path, total_path + "expand3x3_w.npy"), diff --git a/examples/graph_vgg16.cpp b/examples/graph_vgg16.cpp index c8cc5b2558..faaf579047 100644 --- a/examples/graph_vgg16.cpp +++ b/examples/graph_vgg16.cpp @@ -21,8 +21,7 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/graph/Graph.h" -#include "arm_compute/graph/Nodes.h" +#include "arm_compute/graph2.h" #include "support/ToolchainSupport.h" #include "utils/GraphUtils.h" #include "utils/Utils.h" @@ -30,7 +29,7 @@ #include using namespace arm_compute::utils; -using namespace arm_compute::graph; +using namespace arm_compute::graph2::frontend; using namespace arm_compute::graph_utils; namespace @@ -41,9 +40,9 @@ namespace * * @return The convolution layer hint */ -ConvolutionMethodHint convolution_hint_vgg16(size_t size_in_bytes) +ConvolutionMethod convolution_hint_vgg16(size_t size_in_bytes) { - return ((get_mem_free_from_meminfo() * 1024) >= size_in_bytes) ? ConvolutionMethodHint::GEMM : ConvolutionMethodHint::DIRECT; + return ((get_mem_free_from_meminfo() * 1024) >= size_in_bytes) ? ConvolutionMethod::GEMM : ConvolutionMethod::DIRECT; } } // namespace @@ -66,12 +65,14 @@ public: std::unique_ptr preprocessor = arm_compute::support::cpp14::make_unique(mean_rgb); // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON - const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; - TargetHint target_hint = set_target_hint(int_target_hint); + const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; + Target target_hint = set_target_hint2(target); + bool enable_tuning = (target == 2); + bool enable_memory_management = true; // Check if we can use GEMM-based convolutions evaluating if the platform has at least 1.8 GB of available memory - const size_t memory_required = 1932735283L; - ConvolutionMethodHint convolution_hint = convolution_hint_vgg16(memory_required); + const size_t memory_required = 1932735283L; + ConvolutionMethod convolution_hint = convolution_hint_vgg16(memory_required); // Parse arguments if(argc < 2) @@ -107,8 +108,8 @@ public: graph << target_hint << convolution_hint - << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32), - get_input_accessor(image, std::move(preprocessor))) + << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32), + get_input_accessor(image, std::move(preprocessor))) // Layer 1 << ConvolutionLayer( 3U, 3U, 64U, @@ -224,10 +225,10 @@ public: get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc8_b.npy")) // Softmax << SoftmaxLayer() - << Tensor(get_output_accessor(label, 5)); + << OutputLayer(get_output_accessor(label, 5)); - // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated - graph.graph_init(int_target_hint == 2); + // Finalize graph + graph.finalize(target_hint, enable_tuning, enable_memory_management); } void do_run() override { @@ -236,7 +237,7 @@ public: } private: - Graph graph{}; + Stream graph{ 0, "VGG16" }; }; /** Main program for VGG16 diff --git a/examples/graph_vgg19.cpp b/examples/graph_vgg19.cpp index 69ae23d87c..55502e0e00 100644 --- a/examples/graph_vgg19.cpp +++ b/examples/graph_vgg19.cpp @@ -21,8 +21,7 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/graph/Graph.h" -#include "arm_compute/graph/Nodes.h" +#include "arm_compute/graph2.h" #include "support/ToolchainSupport.h" #include "utils/GraphUtils.h" #include "utils/Utils.h" @@ -30,7 +29,7 @@ #include using namespace arm_compute::utils; -using namespace arm_compute::graph; +using namespace arm_compute::graph2::frontend; using namespace arm_compute::graph_utils; /** Example demonstrating how to implement VGG19's network using the Compute Library's graph API @@ -52,9 +51,11 @@ public: std::unique_ptr preprocessor = arm_compute::support::cpp14::make_unique(mean_rgb); // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON - const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; - TargetHint target_hint = set_target_hint(int_target_hint); - ConvolutionMethodHint convolution_hint = ConvolutionMethodHint::DIRECT; + const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; + Target target_hint = set_target_hint2(target); + ConvolutionMethod convolution_hint = ConvolutionMethod::DIRECT; + bool enable_tuning = (target == 2); + bool enable_memory_management = true; // Parse arguments if(argc < 2) @@ -90,8 +91,8 @@ public: graph << target_hint << convolution_hint - << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32), - get_input_accessor(image, std::move(preprocessor))) + << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32), + get_input_accessor(image, std::move(preprocessor))) // Layer 1 << ConvolutionLayer( 3U, 3U, 64U, @@ -217,10 +218,10 @@ public: get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc8_b.npy")) // Softmax << SoftmaxLayer() - << Tensor(get_output_accessor(label, 5)); + << OutputLayer(get_output_accessor(label, 5)); - // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated - graph.graph_init(int_target_hint == 2); + // Finalize graph + graph.finalize(target_hint, enable_tuning, enable_memory_management); } void do_run() override { @@ -229,7 +230,7 @@ public: } private: - Graph graph{}; + Stream graph{ 0, "VGG19" }; }; /** Main program for VGG19 -- cgit v1.2.1