diff options
Diffstat (limited to 'examples/graph_inception_v4.cpp')
-rw-r--r-- | examples/graph_inception_v4.cpp | 163 |
1 files changed, 84 insertions, 79 deletions
diff --git a/examples/graph_inception_v4.cpp b/examples/graph_inception_v4.cpp index 4e405923fc..b61acfcb3f 100644 --- a/examples/graph_inception_v4.cpp +++ b/examples/graph_inception_v4.cpp @@ -60,7 +60,6 @@ public: // Checks ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "Unsupported data type!"); - ARM_COMPUTE_EXIT_ON_MSG(common_params.data_layout == DataLayout::NHWC, "Unsupported data layout!"); // Print parameter values std::cout << common_params << std::endl; @@ -71,13 +70,19 @@ public: // Create a preprocessor object std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<TFPreproccessor>(); + // Create input descriptor + const TensorShape tensor_shape = permute_shape(TensorShape(299U, 299U, 3U, 1U), DataLayout::NCHW, common_params.data_layout); + TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); + + // Set weights trained layout + const DataLayout weights_layout = DataLayout::NCHW; + graph << common_params.target << common_params.fast_math_hint - << InputLayer(TensorDescriptor(TensorShape(299U, 299U, 3U, 1U), common_params.data_type), - get_input_accessor(common_params, std::move(preprocessor), false)) + << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false)) // Conv2d_1a_3x3 << ConvolutionLayer(3U, 3U, 32U, - get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_weights.npy"), + get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), @@ -87,7 +92,7 @@ public: << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) // Conv2d_2a_3x3 << ConvolutionLayer(3U, 3U, 32U, - get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_weights.npy"), + get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_moving_variance.npy"), @@ -97,7 +102,7 @@ public: << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) // Conv2d_2b_3x3 << ConvolutionLayer(3U, 3U, 64U, - get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_weights.npy"), + get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_moving_variance.npy"), @@ -106,35 +111,35 @@ public: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)); - graph << get_mixed_3a(data_path); - graph << get_mixed_4a(data_path); - graph << get_mixed_5a(data_path); + graph << get_mixed_3a(data_path, weights_layout); + graph << get_mixed_4a(data_path, weights_layout); + graph << get_mixed_5a(data_path, weights_layout); // 4 inception A blocks - graph << get_inceptionA_block(data_path, "Mixed_5b"); - graph << get_inceptionA_block(data_path, "Mixed_5c"); - graph << get_inceptionA_block(data_path, "Mixed_5d"); - graph << get_inceptionA_block(data_path, "Mixed_5e"); + graph << get_inceptionA_block(data_path, weights_layout, "Mixed_5b"); + graph << get_inceptionA_block(data_path, weights_layout, "Mixed_5c"); + graph << get_inceptionA_block(data_path, weights_layout, "Mixed_5d"); + graph << get_inceptionA_block(data_path, weights_layout, "Mixed_5e"); // reduction A block - graph << get_reductionA_block(data_path); + graph << get_reductionA_block(data_path, weights_layout); // 7 inception B blocks - graph << get_inceptionB_block(data_path, "Mixed_6b"); - graph << get_inceptionB_block(data_path, "Mixed_6c"); - graph << get_inceptionB_block(data_path, "Mixed_6d"); - graph << get_inceptionB_block(data_path, "Mixed_6e"); - graph << get_inceptionB_block(data_path, "Mixed_6f"); - graph << get_inceptionB_block(data_path, "Mixed_6g"); - graph << get_inceptionB_block(data_path, "Mixed_6h"); + graph << get_inceptionB_block(data_path, weights_layout, "Mixed_6b"); + graph << get_inceptionB_block(data_path, weights_layout, "Mixed_6c"); + graph << get_inceptionB_block(data_path, weights_layout, "Mixed_6d"); + graph << get_inceptionB_block(data_path, weights_layout, "Mixed_6e"); + graph << get_inceptionB_block(data_path, weights_layout, "Mixed_6f"); + graph << get_inceptionB_block(data_path, weights_layout, "Mixed_6g"); + graph << get_inceptionB_block(data_path, weights_layout, "Mixed_6h"); // reduction B block - graph << get_reductionB_block(data_path); + graph << get_reductionB_block(data_path, weights_layout); // 3 inception C blocks - graph << get_inceptionC_block(data_path, "Mixed_7b"); - graph << get_inceptionC_block(data_path, "Mixed_7c"); - graph << get_inceptionC_block(data_path, "Mixed_7d"); + graph << get_inceptionC_block(data_path, weights_layout, "Mixed_7b"); + graph << get_inceptionC_block(data_path, weights_layout, "Mixed_7c"); + graph << get_inceptionC_block(data_path, weights_layout, "Mixed_7d"); graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG)) << FlattenLayer() << FullyConnectedLayer( 1001U, - get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_weights.npy"), + get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_weights.npy", weights_layout), get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_biases.npy")) << SoftmaxLayer() << OutputLayer(get_output_accessor(common_params, 5)); @@ -162,7 +167,7 @@ private: Stream graph; private: - BranchLayer get_mixed_3a(const std::string &data_path) + BranchLayer get_mixed_3a(const std::string &data_path, DataLayout weights_layout) { std::string total_path = "/cnn_data/inceptionv4_model/Mixed_3a_"; @@ -171,7 +176,7 @@ private: SubStream i_b(graph); i_b << ConvolutionLayer(3U, 3U, 96U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_moving_mean.npy"), get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_moving_variance.npy"), @@ -183,13 +188,13 @@ private: return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b)); } - BranchLayer get_mixed_4a(const std::string &data_path) + BranchLayer get_mixed_4a(const std::string &data_path, DataLayout weights_layout) { std::string total_path = "/cnn_data/inceptionv4_model/Mixed_4a_"; SubStream i_a(graph); i_a << ConvolutionLayer(1U, 1U, 64U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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"), @@ -198,7 +203,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(3U, 3U, 96U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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"), @@ -209,7 +214,7 @@ private: SubStream i_b(graph); i_b << ConvolutionLayer(1U, 1U, 64U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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"), @@ -218,7 +223,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(7U, 1U, 64U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) << 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"), @@ -227,7 +232,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(1U, 7U, 64U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) << 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"), @@ -236,7 +241,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(3U, 3U, 96U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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"), @@ -248,13 +253,13 @@ private: return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b)); } - BranchLayer get_mixed_5a(const std::string &data_path) + BranchLayer get_mixed_5a(const std::string &data_path, DataLayout weights_layout) { std::string total_path = "/cnn_data/inceptionv4_model/Mixed_5a_"; SubStream i_a(graph); i_a << ConvolutionLayer(3U, 3U, 192U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) << 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"), @@ -269,13 +274,13 @@ private: return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b)); } - BranchLayer get_inceptionA_block(const std::string &data_path, std::string &¶m_path) + BranchLayer get_inceptionA_block(const std::string &data_path, DataLayout weights_layout, std::string &¶m_path) { std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_"; SubStream i_a(graph); i_a << ConvolutionLayer(1U, 1U, 96U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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"), @@ -286,7 +291,7 @@ private: SubStream i_b(graph); i_b << ConvolutionLayer(1U, 1U, 64U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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"), @@ -295,7 +300,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(3U, 3U, 96U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) << 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"), @@ -306,7 +311,7 @@ private: SubStream i_c(graph); i_c << ConvolutionLayer(1U, 1U, 64U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(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"), @@ -315,7 +320,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(3U, 3U, 96U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(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"), @@ -324,7 +329,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(3U, 3U, 96U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) << 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"), @@ -336,7 +341,7 @@ private: SubStream i_d(graph); i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) << ConvolutionLayer(1U, 1U, 96U, - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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"), @@ -348,13 +353,13 @@ private: return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)); } - BranchLayer get_reductionA_block(const std::string &data_path) + BranchLayer get_reductionA_block(const std::string &data_path, DataLayout weights_layout) { std::string total_path = "/cnn_data/inceptionv4_model/Mixed_6a_"; SubStream i_a(graph); i_a << ConvolutionLayer(3U, 3U, 384U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) << 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"), @@ -365,7 +370,7 @@ private: SubStream i_b(graph); i_b << ConvolutionLayer(1U, 1U, 192U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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"), @@ -374,7 +379,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(3U, 3U, 224U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1)) << 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"), @@ -383,7 +388,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(3U, 3U, 256U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) << 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"), @@ -398,13 +403,13 @@ private: return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c)); } - BranchLayer get_inceptionB_block(const std::string &data_path, std::string &¶m_path) + BranchLayer get_inceptionB_block(const std::string &data_path, DataLayout weights_layout, std::string &¶m_path) { std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_"; SubStream i_a(graph); i_a << ConvolutionLayer(1U, 1U, 384U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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"), @@ -415,7 +420,7 @@ private: SubStream i_b(graph); i_b << ConvolutionLayer(1U, 1U, 192U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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"), @@ -424,7 +429,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(7U, 1U, 224U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) << 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"), @@ -433,7 +438,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(1U, 7U, 256U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) << 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"), @@ -444,7 +449,7 @@ private: SubStream i_c(graph); i_c << ConvolutionLayer(1U, 1U, 192U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(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"), @@ -453,7 +458,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(1U, 7U, 192U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) << 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"), @@ -462,7 +467,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(7U, 1U, 224U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) << 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"), @@ -471,7 +476,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(1U, 7U, 224U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) << 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"), @@ -480,7 +485,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(7U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) << 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"), @@ -492,7 +497,7 @@ private: SubStream i_d(graph); i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) << ConvolutionLayer(1U, 1U, 128U, - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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"), @@ -504,13 +509,13 @@ private: return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)); } - BranchLayer get_reductionB_block(const std::string &data_path) + BranchLayer get_reductionB_block(const std::string &data_path, DataLayout weights_layout) { std::string total_path = "/cnn_data/inceptionv4_model/Mixed_7a_"; SubStream i_a(graph); i_a << ConvolutionLayer(1U, 1U, 192U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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"), @@ -519,7 +524,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(3U, 3U, 192U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) << 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"), @@ -530,7 +535,7 @@ private: SubStream i_b(graph); i_b << ConvolutionLayer(1U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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"), @@ -539,7 +544,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(7U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0)) << 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"), @@ -548,7 +553,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(1U, 7U, 320U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3)) << 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"), @@ -557,7 +562,7 @@ private: 0.001f) << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer(3U, 3U, 320U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0)) << 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"), @@ -572,13 +577,13 @@ private: return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c)); } - BranchLayer get_inceptionC_block(const std::string &data_path, std::string &¶m_path) + BranchLayer get_inceptionC_block(const std::string &data_path, DataLayout weights_layout, std::string &¶m_path) { std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_"; SubStream i_a(graph); i_a << ConvolutionLayer(1U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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"), @@ -590,7 +595,7 @@ private: SubStream i_b(graph); i_b << ConvolutionLayer( 1U, 1U, 384U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << BatchNormalizationLayer( @@ -604,7 +609,7 @@ private: SubStream i_b1(i_b); i_b1 << ConvolutionLayer( 3U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 0)) << BatchNormalizationLayer( @@ -618,7 +623,7 @@ private: SubStream i_b2(i_b); i_b2 << ConvolutionLayer( 1U, 3U, 256U, - get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 1)) << BatchNormalizationLayer( @@ -635,7 +640,7 @@ private: SubStream i_c(graph); i_c << ConvolutionLayer( 1U, 1U, 384U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << BatchNormalizationLayer( @@ -647,7 +652,7 @@ private: << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 1U, 3U, 448U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 1)) << BatchNormalizationLayer( @@ -659,7 +664,7 @@ private: << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)) << ConvolutionLayer( 3U, 1U, 512U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 0)) << BatchNormalizationLayer( @@ -673,7 +678,7 @@ private: SubStream i_c1(i_c); i_c1 << ConvolutionLayer( 3U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 0)) << BatchNormalizationLayer( @@ -687,7 +692,7 @@ private: SubStream i_c2(i_c); i_c2 << ConvolutionLayer( 1U, 3U, 256U, - get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 1)) << BatchNormalizationLayer( @@ -704,7 +709,7 @@ private: SubStream i_d(graph); i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)) << ConvolutionLayer(1U, 1U, 256U, - get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"), + get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0)) << 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"), |