diff options
Diffstat (limited to 'examples/graph_googlenet.cpp')
| -rw-r--r-- | examples/graph_googlenet.cpp | 226 |
1 files changed, 131 insertions, 95 deletions
diff --git a/examples/graph_googlenet.cpp b/examples/graph_googlenet.cpp index 683205b3b5..f431fc412b 100644 --- a/examples/graph_googlenet.cpp +++ b/examples/graph_googlenet.cpp @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -35,8 +36,7 @@ using namespace arm_compute::graph_utils; class GraphGooglenetExample : public Example { public: - GraphGooglenetExample() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "GoogleNet") + GraphGooglenetExample() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "GoogleNet") { } bool do_setup(int argc, char **argv) override @@ -49,14 +49,15 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; } // Checks - ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph"); + ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), + "QASYMM8 not supported for this graph"); // Print parameter values std::cout << common_params << std::endl; @@ -65,64 +66,99 @@ public: std::string data_path = common_params.data_path; // Create a preprocessor object - const std::array<float, 3> mean_rgb{ { 122.68f, 116.67f, 104.01f } }; + const std::array<float, 3> mean_rgb{{122.68f, 116.67f, 104.01f}}; std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<CaffePreproccessor>(mean_rgb); // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; - graph << common_params.target - << common_params.fast_math_hint + graph << common_params.target << common_params.fast_math_hint << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor))) + << ConvolutionLayer(7U, 7U, 64U, + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_b.npy"), + PadStrideInfo(2, 2, 3, 3)) + .set_name("conv1/7x7_s2") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv1/relu_7x7") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool1/3x3_s2") + << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)) + .set_name("pool1/norm1") << ConvolutionLayer( - 7U, 7U, 64U, - get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_b.npy"), - PadStrideInfo(2, 2, 3, 3)) - .set_name("conv1/7x7_s2") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1/relu_7x7") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool1/3x3_s2") - << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)).set_name("pool1/norm1") - << ConvolutionLayer( - 1U, 1U, 64U, - get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2/3x3_reduce") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2/relu_3x3_reduce") + 1U, 1U, 64U, + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2/3x3_reduce") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv2/relu_3x3_reduce") << ConvolutionLayer( - 3U, 3U, 192U, - get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name("conv2/3x3") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2/relu_3x3") - << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)).set_name("conv2/norm2") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool2/3x3_s2"); - graph << get_inception_node(data_path, "inception_3a", weights_layout, 64, std::make_tuple(96U, 128U), std::make_tuple(16U, 32U), 32U).set_name("inception_3a/concat"); - graph << get_inception_node(data_path, "inception_3b", weights_layout, 128, std::make_tuple(128U, 192U), std::make_tuple(32U, 96U), 64U).set_name("inception_3b/concat"); - graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool3/3x3_s2"); - graph << get_inception_node(data_path, "inception_4a", weights_layout, 192, std::make_tuple(96U, 208U), std::make_tuple(16U, 48U), 64U).set_name("inception_4a/concat"); - graph << get_inception_node(data_path, "inception_4b", weights_layout, 160, std::make_tuple(112U, 224U), std::make_tuple(24U, 64U), 64U).set_name("inception_4b/concat"); - graph << get_inception_node(data_path, "inception_4c", weights_layout, 128, std::make_tuple(128U, 256U), std::make_tuple(24U, 64U), 64U).set_name("inception_4c/concat"); - graph << get_inception_node(data_path, "inception_4d", weights_layout, 112, std::make_tuple(144U, 288U), std::make_tuple(32U, 64U), 64U).set_name("inception_4d/concat"); - graph << get_inception_node(data_path, "inception_4e", weights_layout, 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U).set_name("inception_4e/concat"); - graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool4/3x3_s2"); - graph << get_inception_node(data_path, "inception_5a", weights_layout, 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U).set_name("inception_5a/concat"); - graph << get_inception_node(data_path, "inception_5b", weights_layout, 384, std::make_tuple(192U, 384U), std::make_tuple(48U, 128U), 128U).set_name("inception_5b/concat"); - graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 7, operation_layout, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL))).set_name("pool5/7x7_s1") + 3U, 3U, 192U, + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_b.npy"), + PadStrideInfo(1, 1, 1, 1)) + .set_name("conv2/3x3") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv2/relu_3x3") + << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)) + .set_name("conv2/norm2") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool2/3x3_s2"); + graph << get_inception_node(data_path, "inception_3a", weights_layout, 64, std::make_tuple(96U, 128U), + std::make_tuple(16U, 32U), 32U) + .set_name("inception_3a/concat"); + graph << get_inception_node(data_path, "inception_3b", weights_layout, 128, std::make_tuple(128U, 192U), + std::make_tuple(32U, 96U), 64U) + .set_name("inception_3b/concat"); + graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool3/3x3_s2"); + graph << get_inception_node(data_path, "inception_4a", weights_layout, 192, std::make_tuple(96U, 208U), + std::make_tuple(16U, 48U), 64U) + .set_name("inception_4a/concat"); + graph << get_inception_node(data_path, "inception_4b", weights_layout, 160, std::make_tuple(112U, 224U), + std::make_tuple(24U, 64U), 64U) + .set_name("inception_4b/concat"); + graph << get_inception_node(data_path, "inception_4c", weights_layout, 128, std::make_tuple(128U, 256U), + std::make_tuple(24U, 64U), 64U) + .set_name("inception_4c/concat"); + graph << get_inception_node(data_path, "inception_4d", weights_layout, 112, std::make_tuple(144U, 288U), + std::make_tuple(32U, 64U), 64U) + .set_name("inception_4d/concat"); + graph << get_inception_node(data_path, "inception_4e", weights_layout, 256, std::make_tuple(160U, 320U), + std::make_tuple(32U, 128U), 128U) + .set_name("inception_4e/concat"); + graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool4/3x3_s2"); + graph << get_inception_node(data_path, "inception_5a", weights_layout, 256, std::make_tuple(160U, 320U), + std::make_tuple(32U, 128U), 128U) + .set_name("inception_5a/concat"); + graph << get_inception_node(data_path, "inception_5b", weights_layout, 384, std::make_tuple(192U, 384U), + std::make_tuple(48U, 128U), 128U) + .set_name("inception_5b/concat"); + graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 7, operation_layout, + PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL))) + .set_name("pool5/7x7_s1") << FullyConnectedLayer( - 1000U, - get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_b.npy")) - .set_name("loss3/classifier") - << SoftmaxLayer().set_name("prob") - << OutputLayer(get_output_accessor(common_params, 5)); + 1000U, + get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_w.npy", + weights_layout), + get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_b.npy")) + .set_name("loss3/classifier") + << SoftmaxLayer().set_name("prob") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph GraphConfig config; @@ -148,63 +184,63 @@ private: CommonGraphParams common_params; Stream graph; - ConcatLayer get_inception_node(const std::string &data_path, std::string &¶m_path, DataLayout weights_layout, - unsigned int a_filt, + ConcatLayer get_inception_node(const std::string &data_path, + std::string &¶m_path, + DataLayout weights_layout, + unsigned int a_filt, std::tuple<unsigned int, unsigned int> b_filters, std::tuple<unsigned int, unsigned int> c_filters, - unsigned int d_filt) + unsigned int d_filt) { std::string total_path = "/cnn_data/googlenet_model/" + param_path + "/" + param_path + "_"; SubStream i_a(graph); - i_a << ConvolutionLayer( - 1U, 1U, a_filt, - get_weights_accessor(data_path, total_path + "1x1_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "1x1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/1x1") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_1x1"); + i_a << ConvolutionLayer(1U, 1U, a_filt, + get_weights_accessor(data_path, total_path + "1x1_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "1x1_b.npy"), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/1x1") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_1x1"); SubStream i_b(graph); - i_b << ConvolutionLayer( - 1U, 1U, std::get<0>(b_filters), - get_weights_accessor(data_path, total_path + "3x3_reduce_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "3x3_reduce_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/3x3_reduce") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_3x3_reduce") - << ConvolutionLayer( - 3U, 3U, std::get<1>(b_filters), - get_weights_accessor(data_path, total_path + "3x3_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "3x3_b.npy"), - PadStrideInfo(1, 1, 1, 1)) - .set_name(param_path + "/3x3") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_3x3"); + i_b << ConvolutionLayer(1U, 1U, std::get<0>(b_filters), + get_weights_accessor(data_path, total_path + "3x3_reduce_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "3x3_reduce_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/3x3_reduce") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_3x3_reduce") + << ConvolutionLayer(3U, 3U, std::get<1>(b_filters), + get_weights_accessor(data_path, total_path + "3x3_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "3x3_b.npy"), PadStrideInfo(1, 1, 1, 1)) + .set_name(param_path + "/3x3") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_3x3"); SubStream i_c(graph); - i_c << ConvolutionLayer( - 1U, 1U, std::get<0>(c_filters), - get_weights_accessor(data_path, total_path + "5x5_reduce_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "5x5_reduce_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/5x5_reduce") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_5x5_reduce") - << ConvolutionLayer( - 5U, 5U, std::get<1>(c_filters), - get_weights_accessor(data_path, total_path + "5x5_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "5x5_b.npy"), - PadStrideInfo(1, 1, 2, 2)) - .set_name(param_path + "/5x5") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_5x5"); + i_c << ConvolutionLayer(1U, 1U, std::get<0>(c_filters), + get_weights_accessor(data_path, total_path + "5x5_reduce_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "5x5_reduce_b.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/5x5_reduce") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_5x5_reduce") + << ConvolutionLayer(5U, 5U, std::get<1>(c_filters), + get_weights_accessor(data_path, total_path + "5x5_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "5x5_b.npy"), PadStrideInfo(1, 1, 2, 2)) + .set_name(param_path + "/5x5") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_5x5"); SubStream i_d(graph); - i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL))).set_name(param_path + "/pool") + i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, + PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL))) + .set_name(param_path + "/pool") << ConvolutionLayer( - 1U, 1U, d_filt, - get_weights_accessor(data_path, total_path + "pool_proj_w.npy", weights_layout), - get_weights_accessor(data_path, total_path + "pool_proj_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name(param_path + "/pool_proj") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_pool_proj"); + 1U, 1U, d_filt, get_weights_accessor(data_path, total_path + "pool_proj_w.npy", weights_layout), + get_weights_accessor(data_path, total_path + "pool_proj_b.npy"), PadStrideInfo(1, 1, 0, 0)) + .set_name(param_path + "/pool_proj") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(param_path + "/relu_pool_proj"); return ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)); } |