From dacd3ded3daed84ba5cb10bddebebce8434c7d5a Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Tue, 4 Dec 2018 17:25:48 +0000 Subject: COMPMID-1804: Implement Inception-Resnet-V1 Change-Id: I002f99d7b8a816d780b7a941859babc264c92f65 Reviewed-on: https://review.mlplatform.org/366 Reviewed-by: Michele Di Giorgio Tested-by: Arm Jenkins --- examples/graph_inception_resnet_v1.cpp | 718 +++++++++++++++++++++++++++++++++ 1 file changed, 718 insertions(+) create mode 100644 examples/graph_inception_resnet_v1.cpp (limited to 'examples') diff --git a/examples/graph_inception_resnet_v1.cpp b/examples/graph_inception_resnet_v1.cpp new file mode 100644 index 0000000000..e99f688319 --- /dev/null +++ b/examples/graph_inception_resnet_v1.cpp @@ -0,0 +1,718 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/graph.h" +#include "support/ToolchainSupport.h" +#include "utils/CommonGraphOptions.h" +#include "utils/GraphUtils.h" +#include "utils/Utils.h" + +using namespace arm_compute::utils; +using namespace arm_compute::graph::frontend; +using namespace arm_compute::graph_utils; + +const float batch_norm_epsilon = 0.0010000000474974513f; + +/** Example demonstrating how to implement Inception ResNet V1 network using the Compute Library's graph API */ +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") + { + model_input_width = cmd_parser.add_option>("image-width", 512); + model_input_height = cmd_parser.add_option>("image-height", 512); + + // Add model id option + model_input_width->set_help("Input image width."); + model_input_height->set_help("Input image height."); + } + InceptionResNetV1Example(const InceptionResNetV1Example &) = delete; + InceptionResNetV1Example &operator=(const InceptionResNetV1Example &) = delete; + InceptionResNetV1Example(InceptionResNetV1Example &&) = default; // NOLINT + InceptionResNetV1Example &operator=(InceptionResNetV1Example &&) = default; // NOLINT + ~InceptionResNetV1Example() override = default; + bool do_setup(int argc, char **argv) override + { + // Parse arguments + cmd_parser.parse(argc, argv); + + // Consume common parameters + common_params = consume_common_graph_parameters(common_opts); + + // Return when help menu is requested + if(common_params.help) + { + cmd_parser.print_help(argv[0]); + return false; + } + // Get input image width and height + const unsigned int image_width = model_input_width->value(); + 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) + { + 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"); + + // Print parameter values + std::cout << common_params << std::endl; + std::cout << "Image width: " << image_width << std::endl; + std::cout << "Image height: " << image_height << std::endl; + + // 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()) + { + data_path += model_path; + } + + // Create a preprocessor object + std::unique_ptr preprocessor = arm_compute::support::cpp14::make_unique(0.f, 1.f); + + // Create input descriptor + const TensorShape tensor_shape = permute_shape(TensorShape(image_width, image_height, 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 + << 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(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") + // Conv2d_2a_3x3 + << ConvolutionLayer(3U, 3U, 32U, + get_weights_accessor(data_path, "Conv2d_2a_3x3_weights.npy", weights_layout), + std::unique_ptr(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") + // Conv2d_2b_3x3 + << ConvolutionLayer(3U, 3U, 64U, + get_weights_accessor(data_path, "Conv2d_2b_3x3_weights.npy", weights_layout), + std::unique_ptr(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") + // MaxPool_3a_3x3 + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, 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(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") + // Conv2d_4a_3x3 + << ConvolutionLayer(3U, 3U, 192U, + get_weights_accessor(data_path, "Conv2d_4a_3x3_weights.npy", weights_layout), + std::unique_ptr(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") + // Conv2d_4b_3x3 + << ConvolutionLayer(3U, 3U, 256U, + get_weights_accessor(data_path, "Conv2d_4b_3x3_weights.npy", weights_layout), + std::unique_ptr(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"); + + // 5 x Inception-resnet-A + block35_repeat(data_path, weights_layout, 5); + // Reduction-A + reduction_a(data_path, weights_layout); + // 10 x Inception-Resnet-B + block17_repeat(data_path, weights_layout, 10); + // Reduction-B + reduction_b(data_path, weights_layout); + // 5 x Inception-resnet-C + block8_repeat(data_path, weights_layout, 5, 0.2f, true); + + block8_repeat(data_path, weights_layout, 1, 1.f, false); + + // Logits tail + graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG)).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") + << OutputLayer(arm_compute::support::cpp14::make_unique(0)); + + // Finalize graph + GraphConfig config; + config.num_threads = common_params.threads; + config.use_tuner = common_params.enable_tuner; + config.tuner_file = common_params.tuner_file; + + graph.finalize(common_params.target, config); + + return true; + } + + void do_run() override + { + graph.run(); + } + +private: + CommandLineParser cmd_parser; + CommonGraphOptions common_opts; + CommonGraphParams common_params; + SimpleOption *model_input_width{ nullptr }; + SimpleOption *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) + { + std::stringstream unit_path_ss; + unit_path_ss << "Repeat_block35_" << (i + 1) << "_"; + std::stringstream unit_name_ss; + unit_name_ss << "Repeat/block35_" << (i + 1) << "/"; + + std::string unit_path = unit_path_ss.str(); + std::string unit_name = unit_name_ss.str(); + + // Create left and write substreams + SubStream i_l(graph); + SubStream i_r(graph); + + // 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(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), + std::unique_ptr(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") + << ConvolutionLayer(3U, 3U, 32U, + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), + std::unique_ptr(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"); + + // 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), + std::unique_ptr(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") + << ConvolutionLayer(3U, 3U, 32U, + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), + std::unique_ptr(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") + << ConvolutionLayer(3U, 3U, 32U, + get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_weights.npy", weights_layout), + std::unique_ptr(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"); + + // 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"); + + 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"); + } + } + + void block17_repeat(const std::string &data_path, DataLayout weights_layout, unsigned int num_blocks) + { + for(unsigned int i = 0; i < num_blocks; ++i) + { + std::stringstream unit_path_ss; + unit_path_ss << "Repeat_1_block17_" << (i + 1) << "_"; + std::stringstream unit_name_ss; + unit_name_ss << "Repeat_1/block17_" << (i + 1) << "/"; + + std::string unit_path = unit_path_ss.str(); + std::string unit_name = unit_name_ss.str(); + + // Create left and write substreams + SubStream i_l(graph); + SubStream i_r(graph); + + // 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(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), + std::unique_ptr(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") + << ConvolutionLayer(7U, 1U, 128U, + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), + std::unique_ptr(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") + << ConvolutionLayer(1U, 7U, 128U, + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), + std::unique_ptr(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"); + + // 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"); + + 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"); + } + } + + 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) + { + std::stringstream unit_path_ss; + std::stringstream unit_name_ss; + if(num_blocks != 1) + { + unit_path_ss << "Repeat_2_block8_" << (i + 1) << "_"; + unit_name_ss << "Repeat_2/block8_" << (i + 1) << "/"; + } + else + { + unit_path_ss << "Block8_"; + unit_name_ss << "Block8/"; + } + + std::string unit_path = unit_path_ss.str(); + std::string unit_name = unit_name_ss.str(); + + // Create left and write substreams + SubStream i_l(graph); + SubStream i_r(graph); + + // 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(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), + std::unique_ptr(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") + << ConvolutionLayer(3U, 1U, 192U, + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_weights.npy", weights_layout), + std::unique_ptr(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") + << ConvolutionLayer(1U, 3U, 192U, + get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_weights.npy", weights_layout), + std::unique_ptr(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"); + + // 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"); + + // Scale result + if(scale != 1.f) + { + 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) + { + graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu"); + } + } + } + + void reduction_a(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(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(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(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(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, 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"); + } + + void reduction_b(const std::string &data_path, DataLayout weights_layout) + { + // 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(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(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(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(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(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(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(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, 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"); + } +}; + +/** Main program for Inception ResNet V1 + * + * Model is based on: + * https://arxiv.org/abs/1602.07261 + * "Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning" + * Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, Alex Alemi + * + * @note To list all the possible arguments execute the binary appended with the --help option + * + * @param[in] argc Number of arguments + * @param[in] argv Arguments + */ +int main(int argc, char **argv) +{ + return arm_compute::utils::run_example(argc, argv); +} -- cgit v1.2.1