From 012200749e9a82b1c379bdd87b674d0ce9dffbbb Mon Sep 17 00:00:00 2001 From: =?utf8?q?=EC=98=A4=ED=98=95=EC=84=9D/=EB=8F=99=EC=9E=91=EC=A0=9C?= =?utf8?q?=EC=96=B4Lab=28SR=29/Staff=20Engineer/=EC=82=BC=EC=84=B1?= =?utf8?q?=EC=A0=84=EC=9E=90?= Date: Thu, 19 Jul 2018 17:40:18 +0900 Subject: [PATCH] Fix benchmark failure (#2001) Fix benchmark test failure: migrate to armcl v18.05 Signed-off-by: Hyeongseok Oh --- benchmark/acl/benchmark_googlenet.cpp | 350 +++++++++++++++++-------------- benchmark/acl/benchmark_inception_v3.cpp | 340 +++++++++++++++++++----------- benchmark/acl/benchmark_mobilenet.cpp | 345 +++++++++++++++++++----------- 3 files changed, 636 insertions(+), 399 deletions(-) diff --git a/benchmark/acl/benchmark_googlenet.cpp b/benchmark/acl/benchmark_googlenet.cpp index aa949cf..de145dc 100644 --- a/benchmark/acl/benchmark_googlenet.cpp +++ b/benchmark/acl/benchmark_googlenet.cpp @@ -30,172 +30,212 @@ using namespace arm_compute::graph::frontend; -namespace +inline std::unique_ptr get_input_accessor(void) { - BranchLayer get_inception_node(const std::string &data_path, std::string &¶m_path, - unsigned int a_filt, - std::tuple b_filters, - std::tuple c_filters, - unsigned int d_filt) - { - Stream graph{0, "BENCHMARK_GOOGLENET"}; - SubStream i_a(graph); - i_a << ConvolutionLayer( - 1U, 1U, a_filt, - get_accessor(), - get_accessor(), - PadStrideInfo(1, 1, 0, 0)) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - - SubStream i_b(graph); - i_b << ConvolutionLayer( - 1U, 1U, std::get<0>(b_filters), - get_accessor(), - get_accessor(), - PadStrideInfo(1, 1, 0, 0)) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) - << ConvolutionLayer( - 3U, 3U, std::get<1>(b_filters), - get_accessor(), - get_accessor(), - PadStrideInfo(1, 1, 1, 1)) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - - SubStream i_c(graph); - i_c << ConvolutionLayer( - 1U, 1U, std::get<0>(c_filters), - get_accessor(), - get_accessor(), - PadStrideInfo(1, 1, 0, 0)) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) - << ConvolutionLayer( - 5U, 5U, std::get<1>(c_filters), - get_accessor(), - get_accessor(), - PadStrideInfo(1, 1, 2, 2)) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - - SubStream i_d(graph); - i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL))) - << ConvolutionLayer( - 1U, 1U, d_filt, - get_accessor(), - get_accessor(), - PadStrideInfo(1, 1, 0, 0)) - << 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)); - } -} // namespace + return get_accessor(); +} + +inline std::unique_ptr get_random_accessor(float lower, float upper) +{ + return get_accessor(); +} + +inline std::unique_ptr get_weights_accessor(const std::string &path, const std::string &data_file, DataLayout file_layout = DataLayout::NCHW) +{ + return get_accessor(); +} + +inline std::unique_ptr get_output_accessor(void) +{ + return get_accessor(); +} /** Example demonstrating how to implement Googlenet'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), [optional] Path to the weights folder, [optional] image, [optional] labels ) + * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) ) */ -void main_graph_googlenet(int argc, const char **argv) +class GraphGooglenetExample { - std::string data_path; /* Path to the trainable data */ - std::string image; /* Image data */ - std::string label; /* Label data */ - - constexpr float mean_r = 122.68f; /* Mean value to subtract from red channel */ - constexpr float mean_g = 116.67f; /* Mean value to subtract from green channel */ - constexpr float mean_b = 104.01f; /* Mean value to subtract from blue channel */ - - // Set target. 0 (NEON), 1 (OpenCL). By default it is NEON - Target target_hint = set_target_hint(argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0); - ConvolutionMethod convolution_hint = target_hint == Target::NEON ? ConvolutionMethod::GEMM : ConvolutionMethod::DIRECT; - - // Parse arguments - if(argc < 2) - { - // Print help - std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n"; - std::cout << "No data folder provided: using random values\n\n"; - } - else if(argc == 2) - { - std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n"; - std::cout << "No data folder provided: using random values\n\n"; - } - else if(argc == 3) - { - data_path = argv[2]; - std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n"; - std::cout << "No image provided: using random values\n\n"; - } - else if(argc == 4) - { - data_path = argv[2]; - image = argv[3]; - std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n"; - std::cout << "No text file with labels provided: skipping output accessor\n\n"; - } - else - { - data_path = argv[2]; - image = argv[3]; - label = argv[4]; - } - - Stream graph{0, "BENCHMARK_GOOGLENET"}; - - graph << target_hint - << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32), - get_accessor()) - << ConvolutionLayer( - 7U, 7U, 64U, - get_accessor(), - get_accessor(), - PadStrideInfo(2, 2, 3, 3)) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) - << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)) - << convolution_hint - << ConvolutionLayer( - 1U, 1U, 64U, - get_accessor(), - get_accessor(), - PadStrideInfo(1, 1, 0, 0)) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) - << ConvolutionLayer( - 3U, 3U, 192U, - get_accessor(), - get_accessor(), - PadStrideInfo(1, 1, 1, 1)) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) - << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)) - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) - << get_inception_node(data_path, "inception_3a", 64, std::make_tuple(96U, 128U), std::make_tuple(16U, 32U), 32U) - << get_inception_node(data_path, "inception_3b", 128, std::make_tuple(128U, 192U), std::make_tuple(32U, 96U), 64U) - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) - << get_inception_node(data_path, "inception_4a", 192, std::make_tuple(96U, 208U), std::make_tuple(16U, 48U), 64U) - << get_inception_node(data_path, "inception_4b", 160, std::make_tuple(112U, 224U), std::make_tuple(24U, 64U), 64U) - << get_inception_node(data_path, "inception_4c", 128, std::make_tuple(128U, 256U), std::make_tuple(24U, 64U), 64U) - << get_inception_node(data_path, "inception_4d", 112, std::make_tuple(144U, 288U), std::make_tuple(32U, 64U), 64U) - << get_inception_node(data_path, "inception_4e", 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U) - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) - << get_inception_node(data_path, "inception_5a", 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U) - << get_inception_node(data_path, "inception_5b", 384, std::make_tuple(192U, 384U), std::make_tuple(48U, 128U), 128U) - << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 7, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL))) - << FullyConnectedLayer( - 1000U, - get_accessor(), - get_accessor()) - << SoftmaxLayer() - << OutputLayer(get_accessor()); - - run_benchmark(graph); -} +public: + void do_setup(int argc, char **argv) + { + std::string data_path; /* Path to the trainable data */ + std::string image; /* Image data */ + std::string label; /* Label data */ + + const std::array mean_rgb{ { 122.68f, 116.67f, 104.01f } }; + // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON + const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; + Target target_hint = set_target_hint(target); + FastMathHint fast_math_hint = FastMathHint::DISABLED; + + // Parse arguments + if(argc < 2) + { + // Print help + std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\n\n"; + std::cout << "No data folder provided: using random values\n\n"; + } + else if(argc == 2) + { + std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n"; + std::cout << "No data folder provided: using random values\n\n"; + } + else if(argc == 3) + { + data_path = argv[2]; + std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels] [fast_math_hint]\n\n"; + std::cout << "No image provided: using random values\n\n"; + } + else if(argc == 4) + { + data_path = argv[2]; + image = argv[3]; + std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n"; + std::cout << "No text file with labels provided: skipping output accessor\n\n"; + } + else if(argc == 5) + { + data_path = argv[2]; + image = argv[3]; + label = argv[4]; + std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n"; + std::cout << "No fast math info provided: disabling fast math\n\n"; + } + else + { + data_path = argv[2]; + image = argv[3]; + label = argv[4]; + fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED; + } + + graph << target_hint + << fast_math_hint + << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32), + get_input_accessor()) + << ConvolutionLayer( + 7U, 7U, 64U, + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_w.npy"), + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_b.npy"), + PadStrideInfo(2, 2, 3, 3)) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)) + << ConvolutionLayer( + 1U, 1U, 64U, + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_w.npy"), + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + << ConvolutionLayer( + 3U, 3U, 192U, + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_w.npy"), + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_b.npy"), + PadStrideInfo(1, 1, 1, 1)) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)) + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))); + graph << get_inception_node(data_path, "inception_3a", 64, std::make_tuple(96U, 128U), std::make_tuple(16U, 32U), 32U); + graph << get_inception_node(data_path, "inception_3b", 128, std::make_tuple(128U, 192U), std::make_tuple(32U, 96U), 64U); + graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))); + graph << get_inception_node(data_path, "inception_4a", 192, std::make_tuple(96U, 208U), std::make_tuple(16U, 48U), 64U); + graph << get_inception_node(data_path, "inception_4b", 160, std::make_tuple(112U, 224U), std::make_tuple(24U, 64U), 64U); + graph << get_inception_node(data_path, "inception_4c", 128, std::make_tuple(128U, 256U), std::make_tuple(24U, 64U), 64U); + graph << get_inception_node(data_path, "inception_4d", 112, std::make_tuple(144U, 288U), std::make_tuple(32U, 64U), 64U); + graph << get_inception_node(data_path, "inception_4e", 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U); + graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))); + graph << get_inception_node(data_path, "inception_5a", 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U); + graph << get_inception_node(data_path, "inception_5b", 384, std::make_tuple(192U, 384U), std::make_tuple(48U, 128U), 128U); + graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 7, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL))) + << FullyConnectedLayer( + 1000U, + 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() + << OutputLayer(get_output_accessor()); + + // Finalize graph + GraphConfig config; + config.use_tuner = (target == 2); + graph.finalize(target_hint, config); + } + void do_run() + { + run_benchmark(graph); + } + +private: + Stream graph{ 0, "GoogleNet" }; + + BranchLayer get_inception_node(const std::string &data_path, std::string &¶m_path, + unsigned int a_filt, + std::tuple b_filters, + std::tuple c_filters, + unsigned int d_filt) + { + std::string total_path = "/cnn_data/googlenet_model/" + param_path + "/" + param_path + "_"; + SubStream i_a(graph); + i_a << ConvolutionLayer( + 1U, 1U, a_filt, + get_weights_accessor(data_path, total_path + "1x1_w.npy"), + get_weights_accessor(data_path, total_path + "1x1_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + << 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 + "3x3_reduce_w.npy"), + get_weights_accessor(data_path, total_path + "3x3_reduce_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + << ConvolutionLayer( + 3U, 3U, std::get<1>(b_filters), + get_weights_accessor(data_path, total_path + "3x3_w.npy"), + get_weights_accessor(data_path, total_path + "3x3_b.npy"), + PadStrideInfo(1, 1, 1, 1)) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + + 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"), + get_weights_accessor(data_path, total_path + "5x5_reduce_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + << ConvolutionLayer( + 5U, 5U, std::get<1>(c_filters), + get_weights_accessor(data_path, total_path + "5x5_w.npy"), + get_weights_accessor(data_path, total_path + "5x5_b.npy"), + PadStrideInfo(1, 1, 2, 2)) + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + + SubStream i_d(graph); + i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL))) + << ConvolutionLayer( + 1U, 1U, d_filt, + get_weights_accessor(data_path, total_path + "pool_proj_w.npy"), + get_weights_accessor(data_path, total_path + "pool_proj_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + << 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)); + } +}; /** Main program for Googlenet * * @param[in] argc Number of arguments - * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels ) + * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) ) */ -int main(int argc, const char **argv) +int main(int argc, char **argv) { - main_graph_googlenet(argc, argv); - return 0; + GraphGooglenetExample example; + + example.do_setup(argc, argv); + example.do_run(); + + return 0; } diff --git a/benchmark/acl/benchmark_inception_v3.cpp b/benchmark/acl/benchmark_inception_v3.cpp index 5a31d71..786ee6a 100644 --- a/benchmark/acl/benchmark_inception_v3.cpp +++ b/benchmark/acl/benchmark_inception_v3.cpp @@ -40,7 +40,7 @@ inline std::unique_ptr get_random_accessor( return get_accessor(); } -inline std::unique_ptr get_weights_accessor(const std::string &path, const std::string &data_file) +inline std::unique_ptr get_weights_accessor(const std::string &path, const std::string &data_file, DataLayout file_layout = DataLayout::NCHW) { return get_accessor(); } @@ -65,47 +65,59 @@ public: std::string label; /* Label data */ // 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; - Target 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_hint(target); + FastMathHint fast_math_hint = FastMathHint::DISABLED; // Parse arguments if(argc < 2) { // Print help - std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n"; + std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\n\n"; std::cout << "No data folder provided: using random values\n\n"; } else if(argc == 2) { - std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n"; + std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n"; std::cout << "No data folder provided: using random values\n\n"; } else if(argc == 3) { data_path = argv[2]; - std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n"; + std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels] [fast_math_hint]\n\n"; std::cout << "No image provided: using random values\n\n"; } else if(argc == 4) { data_path = argv[2]; image = argv[3]; - std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n"; + std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n"; std::cout << "No text file with labels provided: skipping output accessor\n\n"; } - else + else if(argc == 5) { data_path = argv[2]; image = argv[3]; label = argv[4]; + std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n"; + std::cout << "No fast math info provided: disabling fast math\n\n"; + } + else + { + data_path = argv[2]; + image = argv[3]; + label = argv[4]; + fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED; } - graph << target_hint << InputLayer(TensorDescriptor(TensorShape(299U, 299U, 3U, 1U), DataType::F32), + graph << target_hint + << fast_math_hint + << InputLayer(TensorDescriptor(TensorShape(299U, 299U, 3U, 1U), DataType::F32), get_input_accessor()) - << ConvolutionLayer(3U, 3U, 32U, get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_1a_3x3_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name("Conv2d_1a_3x3/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, @@ -113,11 +125,12 @@ public: get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_1a_3x3_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) - + .set_name("Conv2d_1a_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_1a_3x3/Relu") << ConvolutionLayer(3U, 3U, 32U, get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2a_3x3_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Conv2d_2a_3x3/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2a_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, @@ -125,11 +138,13 @@ public: get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2a_3x3_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_2a_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2a_3x3/Relu") << ConvolutionLayer(3U, 3U, 64U, get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2b_3x3_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name("Conv2d_2b_3x3/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2b_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, @@ -137,13 +152,15 @@ public: get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_2b_3x3_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_2b_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2b_3x3/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("MaxPool_3a_3x3/MaxPool") << ConvolutionLayer(1U, 1U, 80U, get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_3b_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Conv2d_3b_1x1/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_3b_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, @@ -151,11 +168,13 @@ public: get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_3b_1x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_3b_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_3b_1x1/Relu") << ConvolutionLayer(3U, 3U, 192U, get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_4a_3x3_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name("Conv2d_4a_3x3/convolution") << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_4a_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, @@ -163,48 +182,61 @@ public: get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Conv2d_4a_3x3_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("Conv2d_4a_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_4a_3x3/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("MaxPool_5a_3x3/MaxPool"); - << get_inception_node_A(data_path, "Mixed_5b", 64U, std::make_tuple(48U, 64U), std::make_tuple(64U, 96U, 96U), + graph << get_inception_node_A(data_path, "Mixed_5b", 64U, std::make_tuple(48U, 64U), std::make_tuple(64U, 96U, 96U), 32U) - << get_inception_node_A(data_path, "Mixed_5c", 64U, std::make_tuple(48U, 64U), std::make_tuple(64U, 96U, 96U), + .set_name("Mixed_5b/concat"); + graph << get_inception_node_A(data_path, "Mixed_5c", 64U, std::make_tuple(48U, 64U), std::make_tuple(64U, 96U, 96U), 64U, true) - << get_inception_node_A(data_path, "Mixed_5d", 64U, std::make_tuple(48U, 64U), std::make_tuple(64U, 96U, 96U), + .set_name("Mixed_5c/concat"); + graph << get_inception_node_A(data_path, "Mixed_5d", 64U, std::make_tuple(48U, 64U), std::make_tuple(64U, 96U, 96U), 64U) + .set_name("Mixed_5d/concat"); - << get_inception_node_B(data_path, "Mixed_6a", 384U, std::make_tuple(64U, 96U, 96U)) + graph << get_inception_node_B(data_path, "Mixed_6a", 384U, std::make_tuple(64U, 96U, 96U)).set_name("Mixed_6a/concat"); - << get_inception_node_C(data_path, "Mixed_6b", 192U, std::make_tuple(128U, 128U, 192U), + graph << get_inception_node_C(data_path, "Mixed_6b", 192U, std::make_tuple(128U, 128U, 192U), std::make_tuple(128U, 128U, 128U, 128U, 192U), 192U) - << get_inception_node_C(data_path, "Mixed_6c", 192U, std::make_tuple(160U, 160U, 192U), + .set_name("Mixed_6b/concat"); + graph << get_inception_node_C(data_path, "Mixed_6c", 192U, std::make_tuple(160U, 160U, 192U), std::make_tuple(160U, 160U, 160U, 160U, 192U), 192U) - << get_inception_node_C(data_path, "Mixed_6d", 192U, std::make_tuple(160U, 160U, 192U), + .set_name("Mixed_6c/concat"); + graph << get_inception_node_C(data_path, "Mixed_6d", 192U, std::make_tuple(160U, 160U, 192U), std::make_tuple(160U, 160U, 160U, 160U, 192U), 192U) - << get_inception_node_C(data_path, "Mixed_6e", 192U, std::make_tuple(192U, 192U, 192U), + .set_name("Mixed_6d/concat"); + graph << get_inception_node_C(data_path, "Mixed_6e", 192U, std::make_tuple(192U, 192U, 192U), std::make_tuple(192U, 192U, 192U, 192U, 192U), 192U) + .set_name("Mixed_6e/concat"); - << get_inception_node_D(data_path, "Mixed_7a", std::make_tuple(192U, 320U), + graph << get_inception_node_D(data_path, "Mixed_7a", std::make_tuple(192U, 320U), std::make_tuple(192U, 192U, 192U, 192U)) + .set_name("Mixed_7a/concat"); - << get_inception_node_E(data_path, "Mixed_7b", 320U, std::make_tuple(384U, 384U, 384U), + graph << get_inception_node_E(data_path, "Mixed_7b", 320U, std::make_tuple(384U, 384U, 384U), std::make_tuple(448U, 384U, 384U, 384U), 192U) - << get_inception_node_E(data_path, "Mixed_7c", 320U, std::make_tuple(384U, 384U, 384U), + .set_name("Mixed_7b/concat"); + graph << get_inception_node_E(data_path, "Mixed_7c", 320U, std::make_tuple(384U, 384U, 384U), std::make_tuple(448U, 384U, 384U, 384U), 192U, true) + .set_name("Mixed_7c/concat"); - << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 8, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL))) + graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 8, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL))).set_name("Logits/AvgPool_1a_8x8/AvgPool") << ConvolutionLayer(1U, 1U, 1001U, get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Logits_Conv2d_1c_1x1_weights.npy"), get_weights_accessor(data_path, "/cnn_data/inceptionv3_model/Logits_Conv2d_1c_1x1_biases.npy"), PadStrideInfo(1, 1, 0, 0)) - << ReshapeLayer(TensorShape(1001U)) << SoftmaxLayer() + .set_name("Logits/Conv2d_1c_1x1/convolution") + << ReshapeLayer(TensorShape(1001U)).set_name("Predictions/Reshape") + << SoftmaxLayer().set_name("Predictions/Softmax") << OutputLayer(get_output_accessor()); // Finalize graph GraphConfig config; - config.use_tuner = (int_target_hint == 2); + config.use_tuner = (target == 2); graph.finalize(target_hint, config); } @@ -214,7 +246,7 @@ public: } private: - Stream graph{0, "BENCHMARK_INCEPTION_V3"}; + Stream graph{ 0, "InceptionV3" }; private: BranchLayer get_inception_node_A(const std::string &data_path, std::string &¶m_path, @@ -241,13 +273,15 @@ private: get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu"); SubStream i_b(graph); i_b << ConvolutionLayer( @@ -255,25 +289,29 @@ private: get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + "1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d" + conv_id0 + "1x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + "1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id0 + "1x1_BatchNorm_moving_variance.npy"), 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) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d" + conv_id0 + "1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d" + conv_id0 + "1x1/Relu") << ConvolutionLayer( 5U, 5U, std::get<1>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 2, 2)) + .set_name(param_path + "/Branch_1/Conv2d" + conv_id1 + "5x5/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_1_Conv" + conv_id1 + "5x5_BatchNorm_moving_variance.npy"), 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) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_1/Conv2d" + conv_id1 + "5x5/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d" + conv_id1 + "5x5/Relu"); SubStream i_c(graph); i_c << ConvolutionLayer( @@ -281,52 +319,60 @@ private: get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu") << ConvolutionLayer( 3U, 3U, std::get<1>(c_filters), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), 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)) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0b_3x3/Relu") << ConvolutionLayer( 3U, 3U, std::get<2>(c_filters), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name(param_path + "/Branch_2/Conv2d_0c_3x3/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_2/Conv2d_0c_3x3/BatchNorm/batcnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0c_3x3/Relu"); SubStream i_d(graph); - i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) + i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)).set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool") << ConvolutionLayer( 1U, 1U, d_filt, get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu"); return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)); } @@ -336,19 +382,21 @@ private: std::tuple b_filters) { std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_"; - SubStream i_a(graph); + 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"), std::unique_ptr(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name(param_path + "/Branch_0/Conv2d_1a_1x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_1x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_0/Conv2d_1a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_1a_1x1/Relu"); SubStream i_b(graph); i_b << ConvolutionLayer( @@ -356,41 +404,46 @@ private: get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), 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)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu") << ConvolutionLayer( 3U, 3U, std::get<1>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 1, 1)) + .set_name(param_path + "/Branch_1/Conv2d_0b_3x3/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0b_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0b_3x3/Relu") << ConvolutionLayer( 3U, 3U, std::get<2>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d_1a_1x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_1x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_1/Conv2d_1a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_1a_1x1/Relu"); SubStream i_c(graph); - i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f)); + i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name(param_path + "/Branch_2/MaxPool_1a_3x3/MaxPool"); return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c)); } @@ -402,19 +455,21 @@ private: unsigned int d_filt) { std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_"; - SubStream i_a(graph); + 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"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu"); SubStream i_b(graph); i_b << ConvolutionLayer( @@ -422,37 +477,43 @@ private: get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), 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)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu") << ConvolutionLayer( 7U, 1U, std::get<1>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 3, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0b_1x7/Relu") << ConvolutionLayer( 1U, 7U, std::get<2>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 3)) + .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0c_7x1/Relu"); SubStream i_c(graph); i_c << ConvolutionLayer( @@ -460,76 +521,88 @@ private: get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu") << ConvolutionLayer( 1U, 7U, std::get<1>(c_filters), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 3)) + .set_name(param_path + "/Branch_2/Conv2d_0b_7x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0b_7x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0b_7x1/Relu") << ConvolutionLayer( 7U, 1U, std::get<2>(c_filters), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 3, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0c_1x7/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0c_1x7/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0c_1x7/Relu") << ConvolutionLayer( 1U, 7U, std::get<3>(c_filters), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 3)) + .set_name(param_path + "/Branch_2/Conv2d_0d_7x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_2/Conv2d_0d_7x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0d_7x1/Relu") << ConvolutionLayer( 7U, 1U, std::get<4>(c_filters), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 3, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0e_1x7/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_2/Conv2d_0e_1x7/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0e_1x7/Relu"); SubStream i_d(graph); - i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) + i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)).set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool") << ConvolutionLayer( 1U, 1U, d_filt, get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu"); return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)); } @@ -539,31 +612,35 @@ private: std::tuple b_filters) { std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_"; - SubStream i_a(graph); + 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"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu") << ConvolutionLayer( 3U, 3U, std::get<1>(a_filters), get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name(param_path + "/Branch_0/Conv2d_1a_3x3/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_0/Conv2d_1a_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_1a_3x3/Relu"); SubStream i_b(graph); i_b << ConvolutionLayer( @@ -571,53 +648,60 @@ private: get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), 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)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu") << ConvolutionLayer( 7U, 1U, std::get<1>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 3, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x7/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0b_1x7/Relu") << ConvolutionLayer( 1U, 7U, std::get<2>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 3)) + .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/Branch_1/Conv2d_0c_7x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0c_7x1/Relu") << ConvolutionLayer( 3U, 3U, std::get<3>(b_filters), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(2, 2, 0, 0)) + .set_name(param_path + "/Branch_1/Conv2d_1a_3x3/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_1/Conv2d_1a_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_1a_3x3/Relu"); SubStream i_c(graph); - i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f)); + i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name(param_path + "/Branch_2/MaxPool_1a_3x3/MaxPool"); return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c)); } @@ -637,132 +721,154 @@ private: } std::string total_path = "/cnn_data/inceptionv3_model/" + param_path + "_"; - SubStream i_a(graph); + 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"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_0/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_0/Conv2d_0a_1x1/Relu"); - SubStream i_b1(graph); + 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)) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_1/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0a_1x1/Relu"); + + 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"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 1, 0)) + .set_name(param_path + "/Branch_1/Conv2d_0b_1x3/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"), 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)); + .set_name(param_path + "/Branch_1/Conv2d_0b_1x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d_0b_1x3/Relu"); - SubStream i_b2(graph); + 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"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 1)) + .set_name(param_path + "/Branch_1/Conv2d" + conv_id + "3x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d" + conv_id + "3x1_BatchNorm_moving_variance.npy"), 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) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_1/Conv2d" + conv_id + "3x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_1/Conv2d" + conv_id + "3x1/Relu"); - 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"), + // Merge b1 and b2 + i_b << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_b1), std::move(i_b2)).set_name(param_path + "/Branch_1/concat"); + + SubStream i_c(graph); + i_c << ConvolutionLayer( + 1U, 1U, std::get<0>(c_filters), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, total_path + "Branch_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"), + get_weights_accessor(data_path, total_path + "Branch_2_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)); + .set_name(param_path + "/Branch_2/Conv2d_0a_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0a_1x1/Relu") + << ConvolutionLayer( + 3U, 3U, std::get<1>(c_filters), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy"), + std::unique_ptr(nullptr), + PadStrideInfo(1, 1, 1, 1)) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/convolution") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), + get_random_accessor(1.f, 1.f), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), + 0.001f) + .set_name(param_path + "/Branch_2/Conv2d_0b_3x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0b_3x3/Relu"); - SubStream i_c1(graph); + 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"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 1, 0)) + .set_name(param_path + "/Branch_2/Conv2d_0c_1x3/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_variance.npy"), 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)); + .set_name(param_path + "/Branch_2/Conv2d_0c_1x3/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0c_1x3/Relu"); - SubStream i_c2(graph); + 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"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 1)) + .set_name(param_path + "/Branch_2/Conv2d_0d_3x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_3x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_2/Conv2d_0d_3x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_2/Conv2d_0d_3x1/Relu"); - 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_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) - << 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)) - << 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)).set_name(param_path + "/Branch_2/concat"); SubStream i_d(graph); - i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) + i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)).set_name(param_path + "/Branch_3/AvgPool_0a_3x3/AvgPool") << ConvolutionLayer( 1U, 1U, d_filt, get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"), std::unique_ptr(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/convolution") << BatchNormalizationLayer( get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), get_random_accessor(1.f, 1.f), get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); + .set_name(param_path + "/Branch_3/Conv2d_0b_1x1/BatchNorm/batchnorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/Branch_3/Conv2d_0b_1x1/Relu"); return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)); } @@ -771,7 +877,7 @@ private: /** Main program for Inception V3 * * @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] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) ) */ int main(int argc, char **argv) { diff --git a/benchmark/acl/benchmark_mobilenet.cpp b/benchmark/acl/benchmark_mobilenet.cpp index aecb62a..2ed6965 100644 --- a/benchmark/acl/benchmark_mobilenet.cpp +++ b/benchmark/acl/benchmark_mobilenet.cpp @@ -29,145 +29,236 @@ using namespace arm_compute::graph::frontend; -namespace +inline std::unique_ptr get_input_accessor(void) { - 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 = "/cnn_data/mobilenet_v1_model/" + param_path + "_"; - Stream graph{0, "BENCHMARK_MOBILENET"}; - SubStream sg(graph); - sg << DepthwiseConvolutionLayer( - 3U, 3U, - get_accessor(), - std::unique_ptr(nullptr), - dwc_pad_stride_info) - << BatchNormalizationLayer( - get_accessor(), - get_accessor(), - get_accessor(), - get_accessor(), - 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) - << ConvolutionLayer( - 1U, 1U, conv_filt, - get_accessor(), - std::unique_ptr(nullptr), - conv_pad_stride_info) - << BatchNormalizationLayer( - get_accessor(), - get_accessor(), - get_accessor(), - get_accessor(), - 0.001f) - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)); - - return BranchLayer(std::move(sg)); - } -} // namespace + return get_accessor(); +} + +inline std::unique_ptr get_random_accessor(float lower, float upper) +{ + return get_accessor(); +} + +inline std::unique_ptr get_weights_accessor(const std::string &path, const std::string &data_file, DataLayout file_layout = DataLayout::NCHW) +{ + return get_accessor(); +} + +inline std::unique_ptr get_output_accessor(void) +{ + return get_accessor(); +} /** Example demonstrating how to implement MobileNet'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), [optional] Path to the weights folder, [optional] image, [optional] labels ) */ -void main_graph_mobilenet(int argc, const char **argv) +class GraphMobilenetExample { - std::string data_path; /* Path to the trainable data */ - std::string image; /* Image data */ - std::string label; /* Label data */ - - constexpr float mean_r = 122.68f; /* Mean value to subtract from red channel */ - constexpr float mean_g = 116.67f; /* Mean value to subtract from green channel */ - constexpr float mean_b = 104.01f; /* Mean value to subtract from blue channel */ - - // Set target. 0 (NEON), 1 (OpenCL). By default it is NEON - Target target_hint = set_target_hint(argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0); - - // Parse arguments - if(argc < 2) - { - // Print help - std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n"; - std::cout << "No data folder provided: using random values\n\n"; - } - else if(argc == 2) - { - std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n"; - std::cout << "No data folder provided: using random values\n\n"; - } - else if(argc == 3) - { - data_path = argv[2]; - std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n"; - std::cout << "No image provided: using random values\n\n"; - } - else if(argc == 4) - { - data_path = argv[2]; - image = argv[3]; - std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n"; - std::cout << "No text file with labels provided: skipping output accessor\n\n"; - } - else - { - data_path = argv[2]; - image = argv[3]; - label = argv[4]; - } - - Stream graph{0, "BENCHMARK_MOBILENET"}; - - graph << target_hint - << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32), - get_accessor()) - << ConvolutionLayer( - 3U, 3U, 32U, - get_accessor(), - std::unique_ptr(nullptr), - PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR)) - << BatchNormalizationLayer( - get_accessor(), - get_accessor(), - get_accessor(), - get_accessor(), - 0.001f) - - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)) - << get_dwsc_node(data_path, "Conv2d_1", 64, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0)) - << get_dwsc_node(data_path, "Conv2d_2", 128, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR), PadStrideInfo(1, 1, 0, 0)) - << get_dwsc_node(data_path, "Conv2d_3", 128, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR), PadStrideInfo(1, 1, 0, 0)) - << get_dwsc_node(data_path, "Conv2d_4", 256, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR), PadStrideInfo(1, 1, 0, 0)) - << get_dwsc_node(data_path, "Conv2d_5", 256, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR), PadStrideInfo(1, 1, 0, 0)) - << get_dwsc_node(data_path, "Conv2d_6", 512, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR), PadStrideInfo(1, 1, 0, 0)) - << get_dwsc_node(data_path, "Conv2d_7", 512, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR), PadStrideInfo(1, 1, 0, 0)) - << get_dwsc_node(data_path, "Conv2d_8", 512, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR), PadStrideInfo(1, 1, 0, 0)) - << get_dwsc_node(data_path, "Conv2d_9", 512, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR), PadStrideInfo(1, 1, 0, 0)) - << get_dwsc_node(data_path, "Conv2d_10", 512, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR), PadStrideInfo(1, 1, 0, 0)) - << get_dwsc_node(data_path, "Conv2d_11", 512, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR), PadStrideInfo(1, 1, 0, 0)) - << get_dwsc_node(data_path, "Conv2d_12", 1024, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR), PadStrideInfo(1, 1, 0, 0)) - << get_dwsc_node(data_path, "Conv2d_13", 1024, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR), PadStrideInfo(1, 1, 0, 0)) - << PoolingLayer(PoolingLayerInfo(PoolingType::AVG)) - << ConvolutionLayer( - 1U, 1U, 1001U, - get_accessor(), - get_accessor(), - PadStrideInfo(1, 1, 0, 0)) - << ReshapeLayer(TensorShape(1001U)) - << SoftmaxLayer() - << OutputLayer(get_accessor()); - - run_benchmark(graph); -} +public: + void do_setup(int argc, char **argv) + { + std::string data_path; /* Path to the trainable data */ + std::string image; /* Image data */ + std::string label; /* Label data */ + + // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON + const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0; + Target target_hint = set_target_hint(target); + ConvolutionMethod convolution_hint = ConvolutionMethod::GEMM; + DepthwiseConvolutionMethod depthwise_convolution_hint = DepthwiseConvolutionMethod::OPTIMIZED_3x3; + FastMathHint fast_math_hint = FastMathHint::DISABLED; + + // 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; + ARM_COMPUTE_ERROR_ON_MSG(model_id > 1, "Invalid model ID. Model must be 0 (MobileNetV1_1.0_224) or 1 (MobileNetV1_0.75_160)"); + int layout_id = (argc > 3) ? std::strtol(argv[3], nullptr, 10) : 0; + ARM_COMPUTE_ERROR_ON_MSG(layout_id > 1, "Invalid layout ID. Layout must be 0 (NCHW) or 1 (NHWC)"); + + float depth_scale = (model_id == 0) ? 1.f : 0.75; + unsigned int spatial_size = (model_id == 0) ? 224 : 160; + std::string model_path = (model_id == 0) ? "/cnn_data/mobilenet_v1_1_224_model/" : "/cnn_data/mobilenet_v1_075_160_model/"; + TensorDescriptor input_descriptor_nchw = TensorDescriptor(TensorShape(spatial_size, spatial_size, 3U, 1U), DataType::F32); + TensorDescriptor input_descriptor_nhwc = TensorDescriptor(TensorShape(3U, spatial_size, spatial_size, 1U), DataType::F32).set_layout(DataLayout::NHWC); + TensorDescriptor input_descriptor = (layout_id == 0) ? input_descriptor_nchw : input_descriptor_nhwc; + + // Parse arguments + if(argc < 2) + { + // Print help + std::cout << "Usage: " << argv[0] << " [target] [model] [layout] [path_to_data] [image] [labels] [fast_math_hint]\n\n"; + std::cout << "No model ID provided: using MobileNetV1_1.0_224\n\n"; + std::cout << "No data layout provided: using NCHW\n\n"; + std::cout << "No data folder provided: using random values\n\n"; + } + else if(argc == 2) + { + std::cout << "Usage: " << argv[0] << " " << argv[1] << " [model] [layout] [path_to_data] [image] [labels] [fast_math_hint]\n\n"; + std::cout << "No model ID provided: using MobileNetV1_1.0_224\n\n"; + std::cout << "No data layout provided: using NCHW\n\n"; + std::cout << "No data folder provided: using random values\n\n"; + } + else if(argc == 3) + { + std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [layout] [path_to_data] [image] [labels] [fast_math_hint]\n\n"; + std::cout << "No data layout provided: using NCHW\n\n"; + std::cout << "No data folder provided: using random values\n\n"; + } + else if(argc == 4) + { + std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n"; + std::cout << "No data folder provided: using random values\n\n"; + } + else if(argc == 5) + { + data_path = argv[4]; + std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [image] [labels] [fast_math_hint]\n\n"; + std::cout << "No image provided: using random values\n\n"; + std::cout << "No text file with labels provided: skipping output accessor\n\n"; + } + else if(argc == 6) + { + data_path = argv[4]; + image = argv[5]; + std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n"; + std::cout << "No text file with labels provided: skipping output accessor\n\n"; + } + else if(argc == 7) + { + data_path = argv[4]; + image = argv[5]; + label = argv[6]; + std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n"; + std::cout << "No fast math info provided: disabling fast math\n\n"; + } + else + { + data_path = argv[4]; + image = argv[5]; + label = argv[6]; + fast_math_hint = (std::strtol(argv[7], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED; + } + + // Add model path to data path + if(!data_path.empty()) + { + data_path += model_path; + } + + graph << target_hint + << convolution_hint + << depthwise_convolution_hint + << fast_math_hint + << InputLayer(input_descriptor, + get_input_accessor()) + << ConvolutionLayer( + 3U, 3U, 32U * depth_scale, + get_weights_accessor(data_path, "Conv2d_0_weights.npy", DataLayout::NCHW), + std::unique_ptr(nullptr), + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR)) + .set_name("Conv2d_0") + << BatchNormalizationLayer( + get_weights_accessor(data_path, "Conv2d_0_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "Conv2d_0_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, "Conv2d_0_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, "Conv2d_0_BatchNorm_beta.npy"), + 0.001f) + .set_name("Conv2d_0/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name("Conv2d_0/Relu6"); + graph << get_dwsc_node(data_path, "Conv2d_1", 64 * depth_scale, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node(data_path, "Conv2d_2", 128 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node(data_path, "Conv2d_3", 128 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node(data_path, "Conv2d_4", 256 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node(data_path, "Conv2d_5", 256 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node(data_path, "Conv2d_6", 512 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node(data_path, "Conv2d_7", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node(data_path, "Conv2d_8", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node(data_path, "Conv2d_9", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node(data_path, "Conv2d_10", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node(data_path, "Conv2d_11", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node(data_path, "Conv2d_12", 1024 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); + graph << get_dwsc_node(data_path, "Conv2d_13", 1024 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0)); + graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG)).set_name("Logits/AvgPool_1a") + << ConvolutionLayer( + 1U, 1U, 1001U, + get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy", DataLayout::NCHW), + get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_biases.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("Logits/Conv2d_1c_1x1") + << ReshapeLayer(TensorShape(1001U)).set_name("Reshape") + << SoftmaxLayer().set_name("Softmax") + << OutputLayer(get_output_accessor()); + + // Finalize graph + GraphConfig config; + config.use_tuner = (target == 2); + graph.finalize(target_hint, config); + } + void do_run() + { + run_benchmark(graph); + } + +private: + 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 + "_"; + SubStream sg(graph); + sg << DepthwiseConvolutionLayer( + 3U, 3U, + get_weights_accessor(data_path, total_path + "depthwise_depthwise_weights.npy", DataLayout::NCHW), + std::unique_ptr(nullptr), + dwc_pad_stride_info) + .set_name(total_path + "depthwise/depthwise") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_beta.npy"), + 0.001f) + .set_name(total_path + "depthwise/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name(total_path + "depthwise/Relu6") + << ConvolutionLayer( + 1U, 1U, conv_filt, + get_weights_accessor(data_path, total_path + "pointwise_weights.npy", DataLayout::NCHW), + std::unique_ptr(nullptr), + conv_pad_stride_info) + .set_name(total_path + "pointwise/Conv2D") + << BatchNormalizationLayer( + get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_beta.npy"), + 0.001f) + .set_name(total_path + "pointwise/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name(total_path + "pointwise/Relu6"); + + return BranchLayer(std::move(sg)); + } +}; /** Main program for MobileNetV1 * * @param[in] argc Number of arguments - * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels ) + * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), + * [optional] Model ID (0 = MobileNetV1_1.0_224, 1 = MobileNetV1_0.75_160), + * [optional] Path to the weights folder, + * [optional] image, + * [optional] labels, + * [optional] data layout, + * [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) ) */ -int main(int argc, const char **argv) +int main(int argc, char **argv) { - main_graph_mobilenet(argc, argv); - return 0; + GraphMobilenetExample example; + + example.do_setup(argc, argv); + example.do_run(); + + return 0; } -- 2.7.4