examples/graph_inception_v4.cpp

   1 /*
   2  * Copyright (c) 2018 ARM Limited.
   3  *
   4  * SPDX-License-Identifier: MIT
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a copy
   7  * of this software and associated documentation files (the "Software"), to
   8  * deal in the Software without restriction, including without limitation the
   9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  10  * sell copies of the Software, and to permit persons to whom the Software is
  11  * furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in all
  14  * copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  */
  24 #include "arm_compute/graph.h"
  25 #include "support/ToolchainSupport.h"
  26 #include "utils/GraphUtils.h"
  27 #include "utils/Utils.h"
  28
  29 #include <cstdlib>
  30 #include <tuple>
  31
  32 using namespace arm_compute::utils;
  33 using namespace arm_compute::graph::frontend;
  34 using namespace arm_compute::graph_utils;
  35
  36 /** Example demonstrating how to implement InceptionV4's network using the Compute Library's graph API
  37  *
  38  * @param[in] argc Number of arguments
  39  * @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) )
  40  */
  41 class InceptionV4Example final : public Example
  42 {
  43 public:
  44     void do_setup(int argc, char **argv) override
  45     {
  46         // Disabled the test for now because the process gets killed on Linux Firefly 32 bit even when using ConvolutionMethodHint::DIRECT.
  47         // Needs to review/rework to run the code below.
  48 #if __aarch64__
  49         std::string data_path; /* Path to the trainable data */
  50         std::string image;     /* Image data */
  51         std::string label;     /* Label data */
  52
  53         // Create a preprocessor object
  54         std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<TFPreproccessor>();
  55
  56         // Set target. 0 (NEON), 1 (OpenCL). By default it is NEON
  57         const int    target         = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
  58         Target       target_hint    = set_target_hint(target);
  59         FastMathHint fast_math_hint = FastMathHint::DISABLED;
  60
  61         // Parse arguments
  62         if(argc < 2)
  63         {
  64             // Print help
  65             std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
  66             std::cout << "No data folder provided: using random values\n\n";
  67         }
  68         else if(argc == 2)
  69         {
  70             std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
  71             std::cout << "No data folder provided: using random values\n\n";
  72         }
  73         else if(argc == 3)
  74         {
  75             data_path = argv[2];
  76             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels] [fast_math_hint]\n\n";
  77             std::cout << "No image provided: using random values\n\n";
  78         }
  79         else if(argc == 4)
  80         {
  81             data_path = argv[2];
  82             image     = argv[3];
  83             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
  84             std::cout << "No text file with labels provided: skipping output accessor\n\n";
  85         }
  86         else if(argc == 5)
  87         {
  88             data_path = argv[2];
  89             image     = argv[3];
  90             label     = argv[4];
  91             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
  92             std::cout << "No fast math info provided: disabling fast math\n\n";
  93         }
  94         else
  95         {
  96             data_path      = argv[2];
  97             image          = argv[3];
  98             label          = argv[4];
  99             fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
 100         }
 101
 102         graph << target_hint
 103               << fast_math_hint
 104               << InputLayer(TensorDescriptor(TensorShape(299U, 299U, 3U, 1U), DataType::F32),
 105                             get_input_accessor(image, std::move(preprocessor), false))
 106               // Conv2d_1a_3x3
 107               << ConvolutionLayer(3U, 3U, 32U,
 108                                   get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_weights.npy"),
 109                                   std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
 110               << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
 111                                          get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
 112                                          get_random_accessor(1.f, 1.f),
 113                                          get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_beta.npy"),
 114                                          0.001f)
 115               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 116               // Conv2d_2a_3x3
 117               << ConvolutionLayer(3U, 3U, 32U,
 118                                   get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_weights.npy"),
 119                                   std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 120               << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_moving_mean.npy"),
 121                                          get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_moving_variance.npy"),
 122                                          get_random_accessor(1.f, 1.f),
 123                                          get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_beta.npy"),
 124                                          0.001f)
 125               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 126               // Conv2d_2b_3x3
 127               << ConvolutionLayer(3U, 3U, 64U,
 128                                   get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_weights.npy"),
 129                                   std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
 130               << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_moving_mean.npy"),
 131                                          get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_moving_variance.npy"),
 132                                          get_random_accessor(1.f, 1.f),
 133                                          get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_beta.npy"),
 134                                          0.001f)
 135               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 136
 137         graph << get_mixed_3a(data_path);
 138         graph << get_mixed_4a(data_path);
 139         graph << get_mixed_5a(data_path);
 140         // 4 inception A blocks
 141         graph << get_inceptionA_block(data_path, "Mixed_5b");
 142         graph << get_inceptionA_block(data_path, "Mixed_5c");
 143         graph << get_inceptionA_block(data_path, "Mixed_5d");
 144         graph << get_inceptionA_block(data_path, "Mixed_5e");
 145         // reduction A block
 146         graph << get_reductionA_block(data_path);
 147         // 7 inception B blocks
 148         graph << get_inceptionB_block(data_path, "Mixed_6b");
 149         graph << get_inceptionB_block(data_path, "Mixed_6c");
 150         graph << get_inceptionB_block(data_path, "Mixed_6d");
 151         graph << get_inceptionB_block(data_path, "Mixed_6e");
 152         graph << get_inceptionB_block(data_path, "Mixed_6f");
 153         graph << get_inceptionB_block(data_path, "Mixed_6g");
 154         graph << get_inceptionB_block(data_path, "Mixed_6h");
 155         // reduction B block
 156         graph << get_reductionB_block(data_path);
 157         // 3 inception C blocks
 158         graph << get_inceptionC_block(data_path, "Mixed_7b");
 159         graph << get_inceptionC_block(data_path, "Mixed_7c");
 160         graph << get_inceptionC_block(data_path, "Mixed_7d");
 161         graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
 162               << FlattenLayer()
 163               << FullyConnectedLayer(
 164                   1001U,
 165                   get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_weights.npy"),
 166                   get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_biases.npy"))
 167               << SoftmaxLayer()
 168               << OutputLayer(get_output_accessor(label, 5));
 169
 170         // Finalize graph
 171         GraphConfig config;
 172         config.use_tuner = (target == 2);
 173         graph.finalize(target_hint, config);
 174 #else  /* __aarch64__ */
 175         using namespace arm_compute;
 176         ARM_COMPUTE_UNUSED(argc);
 177         ARM_COMPUTE_UNUSED(argv);
 178 #endif /* __aarch64__ */
 179     }
 180
 181     void do_run() override
 182     {
 183 #if __aarch64__
 184         graph.run();
 185 #endif /* __aarch64__ */
 186     }
 187
 188 private:
 189     Stream graph{ 0, "InceptionV4" };
 190
 191 private:
 192     BranchLayer get_mixed_3a(const std::string &data_path)
 193     {
 194         std::string total_path = "/cnn_data/inceptionv4_model/Mixed_3a_";
 195
 196         SubStream i_a(graph);
 197         i_a << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true));
 198
 199         SubStream i_b(graph);
 200         i_b << ConvolutionLayer(3U, 3U, 96U,
 201                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_weights.npy"),
 202                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
 203             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_moving_mean.npy"),
 204                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_moving_variance.npy"),
 205                                        get_random_accessor(1.f, 1.f),
 206                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_beta.npy"),
 207                                        0.001f)
 208             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 209
 210         return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
 211     }
 212
 213     BranchLayer get_mixed_4a(const std::string &data_path)
 214     {
 215         std::string total_path = "/cnn_data/inceptionv4_model/Mixed_4a_";
 216
 217         SubStream i_a(graph);
 218         i_a << ConvolutionLayer(1U, 1U, 64U,
 219                                 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
 220                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 221             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
 222                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
 223                                        get_random_accessor(1.f, 1.f),
 224                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
 225                                        0.001f)
 226             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 227             << ConvolutionLayer(3U, 3U, 96U,
 228                                 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"),
 229                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 230             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
 231                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
 232                                        get_random_accessor(1.f, 1.f),
 233                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"),
 234                                        0.001f)
 235             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 236
 237         SubStream i_b(graph);
 238         i_b << ConvolutionLayer(1U, 1U, 64U,
 239                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
 240                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 241             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
 242                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
 243                                        get_random_accessor(1.f, 1.f),
 244                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
 245                                        0.001f)
 246             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 247             << ConvolutionLayer(7U, 1U, 64U,
 248                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"),
 249                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0))
 250             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"),
 251                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"),
 252                                        get_random_accessor(1.f, 1.f),
 253                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"),
 254                                        0.001f)
 255             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 256             << ConvolutionLayer(1U, 7U, 64U,
 257                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"),
 258                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3))
 259             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"),
 260                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"),
 261                                        get_random_accessor(1.f, 1.f),
 262                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"),
 263                                        0.001f)
 264             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 265             << ConvolutionLayer(3U, 3U, 96U,
 266                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy"),
 267                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 268             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
 269                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
 270                                        get_random_accessor(1.f, 1.f),
 271                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"),
 272                                        0.001f)
 273             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 274
 275         return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
 276     }
 277
 278     BranchLayer get_mixed_5a(const std::string &data_path)
 279     {
 280         std::string total_path = "/cnn_data/inceptionv4_model/Mixed_5a_";
 281
 282         SubStream i_a(graph);
 283         i_a << ConvolutionLayer(3U, 3U, 192U,
 284                                 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"),
 285                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
 286             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
 287                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
 288                                        get_random_accessor(1.f, 1.f),
 289                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"),
 290                                        0.001f)
 291             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 292
 293         SubStream i_b(graph);
 294         i_b << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true));
 295
 296         return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
 297     }
 298
 299     BranchLayer get_inceptionA_block(const std::string &data_path, std::string &&param_path)
 300     {
 301         std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_";
 302
 303         SubStream i_a(graph);
 304         i_a << ConvolutionLayer(1U, 1U, 96U,
 305                                 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
 306                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 307             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
 308                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
 309                                        get_random_accessor(1.f, 1.f),
 310                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
 311                                        0.001f)
 312             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 313
 314         SubStream i_b(graph);
 315         i_b << ConvolutionLayer(1U, 1U, 64U,
 316                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
 317                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 318             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
 319                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
 320                                        get_random_accessor(1.f, 1.f),
 321                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
 322                                        0.001f)
 323             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 324             << ConvolutionLayer(3U, 3U, 96U,
 325                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy"),
 326                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
 327             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
 328                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
 329                                        get_random_accessor(1.f, 1.f),
 330                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"),
 331                                        0.001f)
 332             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 333
 334         SubStream i_c(graph);
 335         i_c << ConvolutionLayer(1U, 1U, 64U,
 336                                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"),
 337                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 338             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
 339                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
 340                                        get_random_accessor(1.f, 1.f),
 341                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"),
 342                                        0.001f)
 343             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 344             << ConvolutionLayer(3U, 3U, 96U,
 345                                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy"),
 346                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
 347             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
 348                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
 349                                        get_random_accessor(1.f, 1.f),
 350                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"),
 351                                        0.001f)
 352             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 353             << ConvolutionLayer(3U, 3U, 96U,
 354                                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_weights.npy"),
 355                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
 356             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"),
 357                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"),
 358                                        get_random_accessor(1.f, 1.f),
 359                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"),
 360                                        0.001f)
 361             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 362
 363         SubStream i_d(graph);
 364         i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true))
 365             << ConvolutionLayer(1U, 1U, 96U,
 366                                 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"),
 367                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 368             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"),
 369                                        get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"),
 370                                        get_random_accessor(1.f, 1.f),
 371                                        get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"),
 372                                        0.001f)
 373             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 374
 375         return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d));
 376     }
 377
 378     BranchLayer get_reductionA_block(const std::string &data_path)
 379     {
 380         std::string total_path = "/cnn_data/inceptionv4_model/Mixed_6a_";
 381
 382         SubStream i_a(graph);
 383         i_a << ConvolutionLayer(3U, 3U, 384U,
 384                                 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"),
 385                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
 386             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
 387                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
 388                                        get_random_accessor(1.f, 1.f),
 389                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"),
 390                                        0.001f)
 391             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 392
 393         SubStream i_b(graph);
 394         i_b << ConvolutionLayer(1U, 1U, 192U,
 395                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
 396                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 397             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
 398                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
 399                                        get_random_accessor(1.f, 1.f),
 400                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
 401                                        0.001f)
 402             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 403             << ConvolutionLayer(3U, 3U, 224U,
 404                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy"),
 405                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
 406             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
 407                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
 408                                        get_random_accessor(1.f, 1.f),
 409                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"),
 410                                        0.001f)
 411             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 412             << ConvolutionLayer(3U, 3U, 256U,
 413                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy"),
 414                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
 415             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
 416                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
 417                                        get_random_accessor(1.f, 1.f),
 418                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"),
 419                                        0.001f)
 420             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 421
 422         SubStream i_c(graph);
 423         i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true));
 424
 425         return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c));
 426     }
 427
 428     BranchLayer get_inceptionB_block(const std::string &data_path, std::string &&param_path)
 429     {
 430         std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_";
 431
 432         SubStream i_a(graph);
 433         i_a << ConvolutionLayer(1U, 1U, 384U,
 434                                 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
 435                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 436             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
 437                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
 438                                        get_random_accessor(1.f, 1.f),
 439                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
 440                                        0.001f)
 441             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 442
 443         SubStream i_b(graph);
 444         i_b << ConvolutionLayer(1U, 1U, 192U,
 445                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
 446                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 447             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
 448                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
 449                                        get_random_accessor(1.f, 1.f),
 450                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
 451                                        0.001f)
 452             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 453             << ConvolutionLayer(7U, 1U, 224U,
 454                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"),
 455                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0))
 456             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"),
 457                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"),
 458                                        get_random_accessor(1.f, 1.f),
 459                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"),
 460                                        0.001f)
 461             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 462             << ConvolutionLayer(1U, 7U, 256U,
 463                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"),
 464                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3))
 465             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"),
 466                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"),
 467                                        get_random_accessor(1.f, 1.f),
 468                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"),
 469                                        0.001f)
 470             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 471
 472         SubStream i_c(graph);
 473         i_c << ConvolutionLayer(1U, 1U, 192U,
 474                                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"),
 475                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 476             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
 477                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
 478                                        get_random_accessor(1.f, 1.f),
 479                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"),
 480                                        0.001f)
 481             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 482             << ConvolutionLayer(1U, 7U, 192U,
 483                                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_weights.npy"),
 484                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3))
 485             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_mean.npy"),
 486                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_variance.npy"),
 487                                        get_random_accessor(1.f, 1.f),
 488                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_beta.npy"),
 489                                        0.001f)
 490             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 491             << ConvolutionLayer(7U, 1U, 224U,
 492                                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_weights.npy"),
 493                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0))
 494             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_mean.npy"),
 495                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_variance.npy"),
 496                                        get_random_accessor(1.f, 1.f),
 497                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_beta.npy"),
 498                                        0.001f)
 499             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 500             << ConvolutionLayer(1U, 7U, 224U,
 501                                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_weights.npy"),
 502                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3))
 503             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_mean.npy"),
 504                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_variance.npy"),
 505                                        get_random_accessor(1.f, 1.f),
 506                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_beta.npy"),
 507                                        0.001f)
 508             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 509             << ConvolutionLayer(7U, 1U, 256U,
 510                                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_weights.npy"),
 511                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0))
 512             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_mean.npy"),
 513                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_variance.npy"),
 514                                        get_random_accessor(1.f, 1.f),
 515                                        get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_beta.npy"),
 516                                        0.001f)
 517             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 518
 519         SubStream i_d(graph);
 520         i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true))
 521             << ConvolutionLayer(1U, 1U, 128U,
 522                                 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"),
 523                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 524             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"),
 525                                        get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"),
 526                                        get_random_accessor(1.f, 1.f),
 527                                        get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"),
 528                                        0.001f)
 529             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 530
 531         return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d));
 532     }
 533
 534     BranchLayer get_reductionB_block(const std::string &data_path)
 535     {
 536         std::string total_path = "/cnn_data/inceptionv4_model/Mixed_7a_";
 537
 538         SubStream i_a(graph);
 539         i_a << ConvolutionLayer(1U, 1U, 192U,
 540                                 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
 541                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 542             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
 543                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
 544                                        get_random_accessor(1.f, 1.f),
 545                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
 546                                        0.001f)
 547             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 548             << ConvolutionLayer(3U, 3U, 192U,
 549                                 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"),
 550                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
 551             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
 552                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
 553                                        get_random_accessor(1.f, 1.f),
 554                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"),
 555                                        0.001f)
 556             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 557
 558         SubStream i_b(graph);
 559         i_b << ConvolutionLayer(1U, 1U, 256U,
 560                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
 561                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 562             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
 563                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
 564                                        get_random_accessor(1.f, 1.f),
 565                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
 566                                        0.001f)
 567             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 568             << ConvolutionLayer(7U, 1U, 256U,
 569                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"),
 570                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0))
 571             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"),
 572                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"),
 573                                        get_random_accessor(1.f, 1.f),
 574                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"),
 575                                        0.001f)
 576             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 577             << ConvolutionLayer(1U, 7U, 320U,
 578                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"),
 579                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3))
 580             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"),
 581                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"),
 582                                        get_random_accessor(1.f, 1.f),
 583                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"),
 584                                        0.001f)
 585             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 586             << ConvolutionLayer(3U, 3U, 320U,
 587                                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy"),
 588                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
 589             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
 590                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
 591                                        get_random_accessor(1.f, 1.f),
 592                                        get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"),
 593                                        0.001f)
 594             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 595
 596         SubStream i_c(graph);
 597         i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true));
 598
 599         return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c));
 600     }
 601
 602     BranchLayer get_inceptionC_block(const std::string &data_path, std::string &&param_path)
 603     {
 604         std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_";
 605
 606         SubStream i_a(graph);
 607         i_a << ConvolutionLayer(1U, 1U, 256U,
 608                                 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
 609                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 610             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
 611                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
 612                                        get_random_accessor(1.f, 1.f),
 613                                        get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
 614                                        0.001f)
 615             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 616
 617         SubStream i_b(graph);
 618         i_b << ConvolutionLayer(
 619                 1U, 1U, 384U,
 620                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
 621                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
 622                 PadStrideInfo(1, 1, 0, 0))
 623             << BatchNormalizationLayer(
 624                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
 625                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
 626                 get_random_accessor(1.f, 1.f),
 627                 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
 628                 0.001f)
 629             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 630
 631         SubStream i_b1(static_cast<IStream &>(i_b));
 632         i_b1 << ConvolutionLayer(
 633                  3U, 1U, 256U,
 634                  get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_weights.npy"),
 635                  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
 636                  PadStrideInfo(1, 1, 1, 0))
 637              << BatchNormalizationLayer(
 638                  get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"),
 639                  get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"),
 640                  get_random_accessor(1.f, 1.f),
 641                  get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"),
 642                  0.001f)
 643              << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 644
 645         SubStream i_b2(static_cast<IStream &>(i_b));
 646         i_b2 << ConvolutionLayer(
 647                  1U, 3U, 256U,
 648                  get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_weights.npy"),
 649                  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
 650                  PadStrideInfo(1, 1, 0, 1))
 651              << BatchNormalizationLayer(
 652                  get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_mean.npy"),
 653                  get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_variance.npy"),
 654                  get_random_accessor(1.f, 1.f),
 655                  get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_beta.npy"),
 656                  0.001f)
 657              << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 658
 659         // Merge b1 and b2
 660         i_b << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_b1), std::move(i_b2));
 661
 662         SubStream i_c(graph);
 663         i_c << ConvolutionLayer(
 664                 1U, 1U, 384U,
 665                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"),
 666                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
 667                 PadStrideInfo(1, 1, 0, 0))
 668             << BatchNormalizationLayer(
 669                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
 670                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
 671                 get_random_accessor(1.f, 1.f),
 672                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"),
 673                 0.001f)
 674             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 675             << ConvolutionLayer(
 676                 1U, 3U, 448U,
 677                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_weights.npy"),
 678                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
 679                 PadStrideInfo(1, 1, 0, 1))
 680             << BatchNormalizationLayer(
 681                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_BatchNorm_moving_mean.npy"),
 682                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_BatchNorm_moving_variance.npy"),
 683                 get_random_accessor(1.f, 1.f),
 684                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_BatchNorm_beta.npy"),
 685                 0.001f)
 686             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 687             << ConvolutionLayer(
 688                 3U, 1U, 512U,
 689                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_weights.npy"),
 690                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
 691                 PadStrideInfo(1, 1, 1, 0))
 692             << BatchNormalizationLayer(
 693                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_mean.npy"),
 694                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_variance.npy"),
 695                 get_random_accessor(1.f, 1.f),
 696                 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_beta.npy"),
 697                 0.001f)
 698             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 699
 700         SubStream i_c1(static_cast<IStream &>(i_c));
 701         i_c1 << ConvolutionLayer(
 702                  3U, 1U, 256U,
 703                  get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_weights.npy"),
 704                  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
 705                  PadStrideInfo(1, 1, 1, 0))
 706              << BatchNormalizationLayer(
 707                  get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_moving_mean.npy"),
 708                  get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_moving_variance.npy"),
 709                  get_random_accessor(1.f, 1.f),
 710                  get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_beta.npy"),
 711                  0.001f)
 712              << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 713
 714         SubStream i_c2(static_cast<IStream &>(i_c));
 715         i_c2 << ConvolutionLayer(
 716                  1U, 3U, 256U,
 717                  get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_weights.npy"),
 718                  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
 719                  PadStrideInfo(1, 1, 0, 1))
 720              << BatchNormalizationLayer(
 721                  get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_moving_mean.npy"),
 722                  get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_moving_variance.npy"),
 723                  get_random_accessor(1.f, 1.f),
 724                  get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_beta.npy"),
 725                  0.001f)
 726              << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 727
 728         // Merge i_c1 and i_c2
 729         i_c << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_c1), std::move(i_c2));
 730
 731         SubStream i_d(graph);
 732         i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true))
 733             << ConvolutionLayer(1U, 1U, 256U,
 734                                 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"),
 735                                 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
 736             << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"),
 737                                        get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"),
 738                                        get_random_accessor(1.f, 1.f),
 739                                        get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"),
 740                                        0.001f)
 741             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 742
 743         return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d));
 744     }
 745 };
 746
 747 /** Main program for Inception V4
 748  *
 749  * @param[in] argc Number of arguments
 750  * @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) )
 751  */
 752 int main(int argc, char **argv)
 753 {
 754     return arm_compute::utils::run_example<InceptionV4Example>(argc, argv);
 755 }