diff options
Diffstat (limited to 'examples/graph_resnet50.cpp')
| -rw-r--r-- | examples/graph_resnet50.cpp | 190 |
1 files changed, 99 insertions, 91 deletions
diff --git a/examples/graph_resnet50.cpp b/examples/graph_resnet50.cpp index 0d3322c886..ba0f0d5fb6 100644 --- a/examples/graph_resnet50.cpp +++ b/examples/graph_resnet50.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 GraphResNetV1_50Example : public Example { public: - GraphResNetV1_50Example() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ResNetV1_50") + GraphResNetV1_50Example() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ResNetV1_50") { } bool do_setup(int argc, char **argv) override @@ -49,7 +49,7 @@ 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; @@ -62,36 +62,40 @@ 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 } }; - std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<CaffePreproccessor>(mean_rgb, - false /* Do not convert to BGR */); + const std::array<float, 3> mean_rgb{{122.68f, 116.67f, 104.01f}}; + std::unique_ptr<IPreprocessor> preprocessor = + std::make_unique<CaffePreproccessor>(mean_rgb, false /* Do not convert to BGR */); // 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 - << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false /* Do not convert to BGR */)) + graph << common_params.target << common_params.fast_math_hint + << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), + false /* Do not convert to BGR */)) << ConvolutionLayer( - 7U, 7U, 64U, - get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(2, 2, 3, 3)) - .set_name("conv1/convolution") + 7U, 7U, 64U, + get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 3, 3)) + .set_name("conv1/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_beta.npy"), - 0.0000100099996416f) - .set_name("conv1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1/Relu") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))).set_name("pool1/MaxPool"); + get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_beta.npy"), + 0.0000100099996416f) + .set_name("conv1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name("conv1/Relu") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, + PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))) + .set_name("pool1/MaxPool"); add_residual_block(data_path, "block1", weights_layout, 64, 3, 2); add_residual_block(data_path, "block2", weights_layout, 128, 4, 2); @@ -100,13 +104,12 @@ public: graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("pool5") << ConvolutionLayer( - 1U, 1U, 1000U, - get_weights_accessor(data_path, "/cnn_data/resnet50_model/logits_weights.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/resnet50_model/logits_biases.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("logits/convolution") - << FlattenLayer().set_name("predictions/Reshape") - << SoftmaxLayer().set_name("predictions/Softmax") + 1U, 1U, 1000U, + get_weights_accessor(data_path, "/cnn_data/resnet50_model/logits_weights.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/resnet50_model/logits_biases.npy"), + PadStrideInfo(1, 1, 0, 0)) + .set_name("logits/convolution") + << FlattenLayer().set_name("predictions/Reshape") << SoftmaxLayer().set_name("predictions/Softmax") << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph @@ -136,10 +139,14 @@ private: CommonGraphParams common_params; Stream graph; - void add_residual_block(const std::string &data_path, const std::string &name, DataLayout weights_layout, - unsigned int base_depth, unsigned int num_units, unsigned int stride) + void add_residual_block(const std::string &data_path, + const std::string &name, + DataLayout weights_layout, + unsigned int base_depth, + unsigned int num_units, + unsigned int stride) { - for(unsigned int i = 0; i < num_units; ++i) + for (unsigned int i = 0; i < num_units; ++i) { std::stringstream unit_path_ss; unit_path_ss << "/cnn_data/resnet50_model/" << name << "_unit_" << (i + 1) << "_bottleneck_v1_"; @@ -151,89 +158,90 @@ private: unsigned int middle_stride = 1; - if(i == (num_units - 1)) + if (i == (num_units - 1)) { middle_stride = stride; } SubStream right(graph); - right << ConvolutionLayer( - 1U, 1U, base_depth, - get_weights_accessor(data_path, unit_path + "conv1_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "conv1/convolution") + right << ConvolutionLayer(1U, 1U, base_depth, + get_weights_accessor(data_path, unit_path + "conv1_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "conv1/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_beta.npy"), - 0.0000100099996416f) - .set_name(unit_name + "conv1/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu") + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_beta.npy"), 0.0000100099996416f) + .set_name(unit_name + "conv1/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "conv1/Relu") - << ConvolutionLayer( - 3U, 3U, base_depth, - get_weights_accessor(data_path, unit_path + "conv2_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(middle_stride, middle_stride, 1, 1)) - .set_name(unit_name + "conv2/convolution") + << ConvolutionLayer(3U, 3U, base_depth, + get_weights_accessor(data_path, unit_path + "conv2_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(middle_stride, middle_stride, 1, 1)) + .set_name(unit_name + "conv2/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_beta.npy"), - 0.0000100099996416f) - .set_name(unit_name + "conv2/BatchNorm") - << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu") + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_beta.npy"), 0.0000100099996416f) + .set_name(unit_name + "conv2/BatchNorm") + << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "conv1/Relu") - << ConvolutionLayer( - 1U, 1U, base_depth * 4, - get_weights_accessor(data_path, unit_path + "conv3_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "conv3/convolution") + << ConvolutionLayer(1U, 1U, base_depth * 4, + get_weights_accessor(data_path, unit_path + "conv3_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), + PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "conv3/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_beta.npy"), - 0.0000100099996416f) - .set_name(unit_name + "conv2/BatchNorm"); + get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_beta.npy"), 0.0000100099996416f) + .set_name(unit_name + "conv2/BatchNorm"); - if(i == 0) + if (i == 0) { SubStream left(graph); left << ConvolutionLayer( - 1U, 1U, base_depth * 4, - get_weights_accessor(data_path, unit_path + "shortcut_weights.npy", weights_layout), - std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), - PadStrideInfo(1, 1, 0, 0)) - .set_name(unit_name + "shortcut/convolution") + 1U, 1U, base_depth * 4, + get_weights_accessor(data_path, unit_path + "shortcut_weights.npy", weights_layout), + std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) + .set_name(unit_name + "shortcut/convolution") << BatchNormalizationLayer( - get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_moving_mean.npy"), - get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_moving_variance.npy"), - get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_gamma.npy"), - get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_beta.npy"), - 0.0000100099996416f) - .set_name(unit_name + "shortcut/BatchNorm"); + get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_moving_mean.npy"), + get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_moving_variance.npy"), + get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_gamma.npy"), + get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_beta.npy"), + 0.0000100099996416f) + .set_name(unit_name + "shortcut/BatchNorm"); - graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(unit_name + "add"); + graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add) + .set_name(unit_name + "add"); } - else if(middle_stride > 1) + else if (middle_stride > 1) { SubStream left(graph); - left << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 1, common_params.data_layout, PadStrideInfo(middle_stride, middle_stride, 0, 0), true)).set_name(unit_name + "shortcut/MaxPool"); + left << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 1, common_params.data_layout, + PadStrideInfo(middle_stride, middle_stride, 0, 0), true)) + .set_name(unit_name + "shortcut/MaxPool"); - graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(unit_name + "add"); + graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add) + .set_name(unit_name + "add"); } else { SubStream left(graph); - graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(unit_name + "add"); + graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add) + .set_name(unit_name + "add"); } - graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu"); + graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) + .set_name(unit_name + "Relu"); } } }; |