From 97988a4b3ef0f840432daf95b6e4b2ad7e5feefd Mon Sep 17 00:00:00 2001 From: Isabella Gottardi Date: Fri, 3 Nov 2017 14:39:44 +0000 Subject: COMPMID-657 - Add PPMAccessor and TopNPredictionsAccessor to SqueezeNet Change-Id: I8baaee68b82e200a4829c3fc6c60dd211c06e14f Reviewed-on: http://mpd-gerrit.cambridge.arm.com/94534 Tested-by: Kaizen Reviewed-by: Gian Marco Iodice Reviewed-by: Anthony Barbier --- examples/graph_squeezenet.cpp | 107 ++++++++++++++++++++---------------------- 1 file changed, 50 insertions(+), 57 deletions(-) (limited to 'examples/graph_squeezenet.cpp') diff --git a/examples/graph_squeezenet.cpp b/examples/graph_squeezenet.cpp index a1a7b8619e..195fe0addb 100644 --- a/examples/graph_squeezenet.cpp +++ b/examples/graph_squeezenet.cpp @@ -43,43 +43,22 @@ using namespace arm_compute::graph; using namespace arm_compute::graph_utils; using namespace arm_compute::logging; -/** Generates appropriate accessor according to the specified path - * - * @note If path is empty will generate a DummyAccessor else will generate a NumPyBinLoader - * - * @param path Path to the data files - * @param data_file Relative path to the data files from path - * - * @return An appropriate tensor accessor - */ -std::unique_ptr get_accessor(const std::string &path, const std::string &data_file) -{ - if(path.empty()) - { - return arm_compute::support::cpp14::make_unique(); - } - else - { - return arm_compute::support::cpp14::make_unique(path + data_file); - } -} - BranchLayer get_expand_fire_node(const std::string &data_path, std::string &¶m_path, unsigned int expand1_filt, unsigned int expand3_filt) { std::string total_path = "/cnn_data/squeezenet_v1.0_model/" + param_path + "_"; SubGraph i_a; i_a << ConvolutionLayer( 1U, 1U, expand1_filt, - get_accessor(data_path, total_path + "expand1x1_w.npy"), - get_accessor(data_path, total_path + "expand1x1_b.npy"), + get_weights_accessor(data_path, total_path + "expand1x1_w.npy"), + get_weights_accessor(data_path, total_path + "expand1x1_b.npy"), PadStrideInfo(1, 1, 0, 0)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); SubGraph i_b; i_b << ConvolutionLayer( 3U, 3U, expand3_filt, - get_accessor(data_path, total_path + "expand3x3_w.npy"), - get_accessor(data_path, total_path + "expand3x3_b.npy"), + get_weights_accessor(data_path, total_path + "expand3x3_w.npy"), + get_weights_accessor(data_path, total_path + "expand3x3_b.npy"), PadStrideInfo(1, 1, 1, 1)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); @@ -89,32 +68,44 @@ BranchLayer get_expand_fire_node(const std::string &data_path, std::string &&par /** Example demonstrating how to implement Squeezenet's network using the Compute Library's graph API * * @param[in] argc Number of arguments - * @param[in] argv Arguments ( [optional] Path to the weights folder, [optional] batches ) + * @param[in] argv Arguments ( [optional] Path to the weights folder, [optional] image, [optional] labels ) */ void main_graph_squeezenet(int argc, const char **argv) { - std::string data_path; /** Path to the trainable data */ - unsigned int batches = 4; /** Number of batches */ + 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 */ // Parse arguments if(argc < 2) { // Print help - std::cout << "Usage: " << argv[0] << " [path_to_data] [batches]\n\n"; + std::cout << "Usage: " << argv[0] << " [path_to_data] [image] [labels]\n\n"; std::cout << "No data folder provided: using random values\n\n"; } else if(argc == 2) { //Do something with argv[1] data_path = argv[1]; - std::cout << "Usage: " << argv[0] << " [path_to_data] [batches]\n\n"; - std::cout << "No number of batches where specified, thus will use the default : " << batches << "\n\n"; + std::cout << "Usage: " << argv[0] << " " << argv[1] << " [image] [labels]\n\n"; + std::cout << "No image provided: using random values\n"; + } + else if(argc == 3) + { + data_path = argv[1]; + image = argv[2]; + std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [labels]\n\n"; + std::cout << "No text file with labels provided: skipping output accessor\n"; } else { - //Do something with argv[1] and argv[2] data_path = argv[1]; - batches = std::strtol(argv[2], nullptr, 0); + image = argv[2]; + label = argv[3]; } // Check if OpenCL is available and initialize the scheduler @@ -126,81 +117,83 @@ void main_graph_squeezenet(int argc, const char **argv) Graph graph; graph << TargetHint::OPENCL - << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, batches), 1, DataType::F32), DummyAccessor()) + << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32), + get_input_accessor(image, mean_r, mean_g, mean_b)) << ConvolutionLayer( 7U, 7U, 96U, - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_w.npy"), - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_b.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_w.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_b.npy"), PadStrideInfo(2, 2, 0, 0)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) << ConvolutionLayer( 1U, 1U, 16U, - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_w.npy"), - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_b.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_w.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_b.npy"), PadStrideInfo(1, 1, 0, 0)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << get_expand_fire_node(data_path, "fire2", 64U, 64U) << ConvolutionLayer( 1U, 1U, 16U, - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_w.npy"), - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_b.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_w.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_b.npy"), PadStrideInfo(1, 1, 0, 0)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << get_expand_fire_node(data_path, "fire3", 64U, 64U) << ConvolutionLayer( 1U, 1U, 32U, - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_w.npy"), - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_b.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_w.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_b.npy"), PadStrideInfo(1, 1, 0, 0)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << get_expand_fire_node(data_path, "fire4", 128U, 128U) << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) << ConvolutionLayer( 1U, 1U, 32U, - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_w.npy"), - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_b.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_w.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_b.npy"), PadStrideInfo(1, 1, 0, 0)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << get_expand_fire_node(data_path, "fire5", 128U, 128U) << ConvolutionLayer( 1U, 1U, 48U, - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_w.npy"), - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_b.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_w.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_b.npy"), PadStrideInfo(1, 1, 0, 0)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << get_expand_fire_node(data_path, "fire6", 192U, 192U) << ConvolutionLayer( 1U, 1U, 48U, - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_w.npy"), - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_b.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_w.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_b.npy"), PadStrideInfo(1, 1, 0, 0)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << get_expand_fire_node(data_path, "fire7", 192U, 192U) << ConvolutionLayer( 1U, 1U, 64U, - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_w.npy"), - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_b.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_w.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_b.npy"), PadStrideInfo(1, 1, 0, 0)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << get_expand_fire_node(data_path, "fire8", 256U, 256U) << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) << ConvolutionLayer( 1U, 1U, 64U, - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_w.npy"), - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_b.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_w.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_b.npy"), PadStrideInfo(1, 1, 0, 0)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << get_expand_fire_node(data_path, "fire9", 256U, 256U) << ConvolutionLayer( 1U, 1U, 1000U, - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_w.npy"), - get_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_b.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_w.npy"), + get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_b.npy"), PadStrideInfo(1, 1, 0, 0)) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 13, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL))) + << FlattenLayer() << SoftmaxLayer() - << Tensor(DummyAccessor()); + << Tensor(get_output_accessor(label, 5)); graph.run(); } @@ -208,9 +201,9 @@ void main_graph_squeezenet(int argc, const char **argv) /** Main program for Squeezenet v1.0 * * @param[in] argc Number of arguments - * @param[in] argv Arguments ( [optional] Path to the weights folder, [optional] batches ) + * @param[in] argv Arguments ( [optional] Path to the weights folder, [optional] image, [optional] labels ) */ int main(int argc, const char **argv) { return arm_compute::utils::run_example(argc, argv, main_graph_squeezenet); -} +} \ No newline at end of file -- cgit v1.2.1