Georgios Pinitas | be2772a | 2018-08-17 15:33:39 +0100 | [diff] [blame] | 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/CommonGraphOptions.h" |
| 27 | #include "utils/GraphUtils.h" |
| 28 | #include "utils/Utils.h" |
| 29 | |
| 30 | using namespace arm_compute::utils; |
| 31 | using namespace arm_compute::graph::frontend; |
| 32 | using namespace arm_compute::graph_utils; |
| 33 | |
Georgios Pinitas | 108ab0b | 2018-09-14 18:35:11 +0100 | [diff] [blame] | 34 | /** Example demonstrating how to implement InceptionV4's network using the Compute Library's graph API */ |
Georgios Pinitas | be2772a | 2018-08-17 15:33:39 +0100 | [diff] [blame] | 35 | class InceptionResNetV2Example final : public Example |
| 36 | { |
| 37 | public: |
| 38 | InceptionResNetV2Example() |
| 39 | : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "InceptionResNetV2") |
| 40 | { |
| 41 | } |
| 42 | bool do_setup(int argc, char **argv) override |
| 43 | { |
| 44 | // Parse arguments |
| 45 | cmd_parser.parse(argc, argv); |
| 46 | |
| 47 | // Consume common parameters |
| 48 | common_params = consume_common_graph_parameters(common_opts); |
| 49 | |
| 50 | // Return when help menu is requested |
| 51 | if(common_params.help) |
| 52 | { |
| 53 | cmd_parser.print_help(argv[0]); |
| 54 | return false; |
| 55 | } |
| 56 | |
| 57 | // Set default layout if needed |
| 58 | if(!common_opts.data_layout->is_set() && common_params.target == Target::NEON) |
| 59 | { |
| 60 | common_params.data_layout = DataLayout::NCHW; |
| 61 | } |
| 62 | |
| 63 | // Checks |
Anthony Barbier | cdd68c0 | 2018-08-23 15:03:41 +0100 | [diff] [blame] | 64 | ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph"); |
Georgios Pinitas | be2772a | 2018-08-17 15:33:39 +0100 | [diff] [blame] | 65 | |
| 66 | // Print parameter values |
| 67 | std::cout << common_params << std::endl; |
| 68 | |
| 69 | // Create model path |
| 70 | std::string data_path = common_params.data_path; |
| 71 | std::string model_path = "/cnn_data/inception_resnet_v2_model/"; |
| 72 | if(!data_path.empty()) |
| 73 | { |
| 74 | data_path += model_path; |
| 75 | } |
| 76 | |
| 77 | // Create a preprocessor object |
| 78 | std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<TFPreproccessor>(0.f, 1.f); |
| 79 | |
| 80 | // Create input descriptor |
| 81 | const TensorShape tensor_shape = permute_shape(TensorShape(299U, 299U, 3U, 1U), DataLayout::NCHW, common_params.data_layout); |
| 82 | TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); |
| 83 | |
| 84 | // Set weights trained layout |
| 85 | const DataLayout weights_layout = DataLayout::NCHW; |
| 86 | |
| 87 | graph << common_params.target |
| 88 | << common_params.fast_math_hint |
| 89 | << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false)) |
| 90 | // Conv2d_1a_3x3 |
| 91 | << ConvolutionLayer(3U, 3U, 32U, |
| 92 | get_weights_accessor(data_path, "Conv2d_1a_3x3_weights.npy", weights_layout), |
| 93 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 94 | PadStrideInfo(2, 2, 0, 0)) |
| 95 | .set_name("Conv2d_1a_3x3/convolution") |
| 96 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), |
| 97 | get_weights_accessor(data_path, "Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), |
| 98 | get_random_accessor(1.f, 1.f), |
| 99 | get_weights_accessor(data_path, "Conv2d_1a_3x3_BatchNorm_beta.npy"), |
| 100 | 0.0010000000474974513f) |
| 101 | .set_name("Conv2d_1a_3x3/BatchNorm") |
| 102 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_1a_3x3/Relu") |
| 103 | // Conv2d_2a_3x3 |
| 104 | << ConvolutionLayer(3U, 3U, 32U, |
| 105 | get_weights_accessor(data_path, "Conv2d_2a_3x3_weights.npy", weights_layout), |
| 106 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 107 | PadStrideInfo(1, 1, 0, 0)) |
| 108 | .set_name("Conv2d_2a_3x3/convolution") |
| 109 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_2a_3x3_BatchNorm_moving_mean.npy"), |
| 110 | get_weights_accessor(data_path, "Conv2d_2a_3x3_BatchNorm_moving_variance.npy"), |
| 111 | get_random_accessor(1.f, 1.f), |
| 112 | get_weights_accessor(data_path, "Conv2d_2a_3x3_BatchNorm_beta.npy"), |
| 113 | 0.0010000000474974513f) |
| 114 | .set_name("Conv2d_2a_3x3/BatchNorm") |
| 115 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2a_3x3/Relu") |
| 116 | // Conv2d_2b_3x3 |
| 117 | << ConvolutionLayer(3U, 3U, 64U, |
| 118 | get_weights_accessor(data_path, "Conv2d_2b_3x3_weights.npy", weights_layout), |
| 119 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 120 | PadStrideInfo(1, 1, 1, 1)) |
| 121 | .set_name("Conv2d_2b_3x3/convolution") |
| 122 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_2b_3x3_BatchNorm_moving_mean.npy"), |
| 123 | get_weights_accessor(data_path, "Conv2d_2b_3x3_BatchNorm_moving_variance.npy"), |
| 124 | get_random_accessor(1.f, 1.f), |
| 125 | get_weights_accessor(data_path, "Conv2d_2b_3x3_BatchNorm_beta.npy"), |
| 126 | 0.0010000000474974513f) |
| 127 | .set_name("Conv2d_2b_3x3/BatchNorm") |
| 128 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2b_3x3/Relu") |
| 129 | // MaxPool_3a_3x3 |
| 130 | << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)).set_name("MaxPool_3a_3x3/MaxPool") |
| 131 | // Conv2d_3b_1x1 |
| 132 | << ConvolutionLayer(1U, 1U, 80U, |
| 133 | get_weights_accessor(data_path, "Conv2d_3b_1x1_weights.npy", weights_layout), |
| 134 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 135 | PadStrideInfo(1, 1, 0, 0)) |
| 136 | .set_name("Conv2d_3b_1x1/convolution") |
| 137 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_3b_1x1_BatchNorm_moving_mean.npy"), |
| 138 | get_weights_accessor(data_path, "Conv2d_3b_1x1_BatchNorm_moving_variance.npy"), |
| 139 | get_random_accessor(1.f, 1.f), |
| 140 | get_weights_accessor(data_path, "Conv2d_3b_1x1_BatchNorm_beta.npy"), |
| 141 | 0.0010000000474974513f) |
| 142 | .set_name("Conv2d_3b_1x1/BatchNorm") |
| 143 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_3b_1x1/Relu") |
| 144 | // Conv2d_4a_3x3 |
| 145 | << ConvolutionLayer(3U, 3U, 192U, |
| 146 | get_weights_accessor(data_path, "Conv2d_4a_3x3_weights.npy", weights_layout), |
| 147 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 148 | PadStrideInfo(1, 1, 0, 0)) |
| 149 | .set_name("Conv2d_4a_3x3/convolution") |
| 150 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_4a_3x3_BatchNorm_moving_mean.npy"), |
| 151 | get_weights_accessor(data_path, "Conv2d_4a_3x3_BatchNorm_moving_variance.npy"), |
| 152 | get_random_accessor(1.f, 1.f), |
| 153 | get_weights_accessor(data_path, "Conv2d_4a_3x3_BatchNorm_beta.npy"), |
| 154 | 0.0010000000474974513f) |
| 155 | .set_name("Conv2d_4a_3x3/BatchNorm") |
| 156 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_4a_3x3/Relu") |
| 157 | // MaxPool_5a_3x3 |
| 158 | << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0), true)).set_name("MaxPool_5a_3x3/MaxPool"); |
| 159 | |
| 160 | block_mixed_5b(data_path, weights_layout); |
| 161 | block35_repeat(data_path, weights_layout, 10); |
| 162 | block_mixed_6a(data_path, weights_layout); |
| 163 | block17_repeat(data_path, weights_layout, 20); |
| 164 | block_mixed_7a(data_path, weights_layout); |
| 165 | block8_repeat(data_path, weights_layout, 9, 0.2f, true); |
| 166 | block8_repeat(data_path, weights_layout, 1, 1.f, false); |
| 167 | |
| 168 | // Conv2d_7b_1x1 |
| 169 | graph << ConvolutionLayer(1U, 1U, 1536U, |
| 170 | get_weights_accessor(data_path, "Conv2d_7b_1x1_weights.npy", weights_layout), |
| 171 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 172 | PadStrideInfo(1, 1, 0, 0)) |
| 173 | .set_name("Conv2d_7b_1x1/convolution") |
| 174 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_7b_1x1_BatchNorm_moving_mean.npy"), |
| 175 | get_weights_accessor(data_path, "Conv2d_7b_1x1_BatchNorm_moving_variance.npy"), |
| 176 | get_random_accessor(1.f, 1.f), |
| 177 | get_weights_accessor(data_path, "Conv2d_7b_1x1_BatchNorm_beta.npy"), |
| 178 | 0.0010000000474974513f) |
| 179 | .set_name("Conv2d_7b_1x1/BatchNorm") |
| 180 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_7b_1x1/Relu") |
| 181 | << PoolingLayer(PoolingLayerInfo(PoolingType::AVG)).set_name("Logits/AvgPool_1a_8x8") |
| 182 | << FlattenLayer().set_name("Logits/Flatten") |
| 183 | << FullyConnectedLayer( |
| 184 | 1001U, |
| 185 | get_weights_accessor(data_path, "Logits_Logits_weights.npy", weights_layout), |
| 186 | get_weights_accessor(data_path, "Logits_Logits_biases.npy")) |
| 187 | .set_name("Logits/Logits") |
| 188 | << SoftmaxLayer().set_name("Logits/Predictions") |
| 189 | << OutputLayer(get_output_accessor(common_params, 5)); |
| 190 | |
| 191 | // Finalize graph |
| 192 | GraphConfig config; |
| 193 | config.num_threads = common_params.threads; |
| 194 | config.use_tuner = common_params.enable_tuner; |
| 195 | config.tuner_file = common_params.tuner_file; |
| 196 | |
| 197 | graph.finalize(common_params.target, config); |
| 198 | |
| 199 | return true; |
| 200 | } |
| 201 | |
| 202 | void do_run() override |
| 203 | { |
| 204 | graph.run(); |
| 205 | } |
| 206 | |
| 207 | private: |
| 208 | CommandLineParser cmd_parser; |
| 209 | CommonGraphOptions common_opts; |
| 210 | CommonGraphParams common_params; |
| 211 | Stream graph; |
| 212 | |
| 213 | private: |
| 214 | void block_mixed_5b(const std::string &data_path, DataLayout weights_layout) |
| 215 | { |
| 216 | // Branch 0 |
| 217 | SubStream i_a(graph); |
| 218 | i_a << ConvolutionLayer(1U, 1U, 96U, |
| 219 | get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_weights.npy", weights_layout), |
| 220 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 221 | PadStrideInfo(1, 1, 0, 0)) |
| 222 | .set_name("Mixed_5b/Branch_0/Conv2d_1x1/convolution") |
| 223 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), |
| 224 | get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), |
| 225 | get_random_accessor(1.f, 1.f), |
| 226 | get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), |
| 227 | 0.0010000000474974513f) |
| 228 | .set_name("Mixed_5b/Branch_0/Conv2d_1x1/BatchNorm") |
| 229 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_0/Conv2d_1x1/Relu"); |
| 230 | |
| 231 | // Branch 1 |
| 232 | SubStream i_b(graph); |
| 233 | i_b << ConvolutionLayer(1U, 1U, 48U, |
| 234 | get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), |
| 235 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 236 | PadStrideInfo(1, 1, 0, 0)) |
| 237 | .set_name("Mixed_5b/Branch_1/Conv2d_0a_1x1/convolution") |
| 238 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), |
| 239 | get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), |
| 240 | get_random_accessor(1.f, 1.f), |
| 241 | get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), |
| 242 | 0.0010000000474974513f) |
| 243 | .set_name("Mixed_5b/Branch_1/Conv2d_0a_1x1/BatchNorm") |
| 244 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_1/Conv2d_0a_1x1/Relu") |
| 245 | << ConvolutionLayer(5U, 5U, 64U, |
| 246 | get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_weights.npy", weights_layout), |
| 247 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 248 | PadStrideInfo(1, 1, 2, 2)) |
| 249 | .set_name("Mixed_5b/Branch_1/Conv2d_0b_5x5/convolution") |
| 250 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_BatchNorm_moving_mean.npy"), |
| 251 | get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_BatchNorm_moving_variance.npy"), |
| 252 | get_random_accessor(1.f, 1.f), |
| 253 | get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_BatchNorm_beta.npy"), |
| 254 | 0.0010000000474974513f) |
| 255 | .set_name("Mixed_5b/Branch_1/Conv2d_0b_5x5/BatchNorm") |
| 256 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_1/Conv2d_0b_5x5/Relu"); |
| 257 | |
| 258 | // Branch 2 |
| 259 | SubStream i_c(graph); |
| 260 | i_c << ConvolutionLayer(1U, 1U, 64U, |
| 261 | get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), |
| 262 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 263 | PadStrideInfo(1, 1, 0, 0)) |
| 264 | .set_name("Mixed_5b/Branch_2/Conv2d_0a_1x1/convolution") |
| 265 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), |
| 266 | get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), |
| 267 | get_random_accessor(1.f, 1.f), |
| 268 | get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), |
| 269 | 0.0010000000474974513f) |
| 270 | .set_name("Mixed_5b/Branch_2/Conv2d_0a_1x1/BatchNorm") |
| 271 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_2/Conv2d_0a_1x1/Relu") |
| 272 | << ConvolutionLayer(3U, 3U, 96U, |
| 273 | get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), |
| 274 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 275 | PadStrideInfo(1, 1, 1, 1)) |
| 276 | .set_name("Mixed_5b/Branch_2/Conv2d_0b_3x3/convolution") |
| 277 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), |
| 278 | get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), |
| 279 | get_random_accessor(1.f, 1.f), |
| 280 | get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), |
| 281 | 0.0010000000474974513f) |
| 282 | .set_name("Mixed_5b/Branch_2/Conv2d_0b_3x3/BatchNorm") |
| 283 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_2/Conv2d_0b_3x3/Relu") |
| 284 | << ConvolutionLayer(3U, 3U, 96U, |
| 285 | get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_weights.npy", weights_layout), |
| 286 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 287 | PadStrideInfo(1, 1, 1, 1)) |
| 288 | .set_name("Mixed_5b/Branch_2/Conv2d_0c_3x3/convolution") |
| 289 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"), |
| 290 | get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"), |
| 291 | get_random_accessor(1.f, 1.f), |
| 292 | get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"), |
| 293 | 0.0010000000474974513f) |
| 294 | .set_name("Mixed_5b/Branch_2/Conv2d_0c_3x3/BatchNorm") |
| 295 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_2/Conv2d_0c_3x3/Relu"); |
| 296 | |
| 297 | // Branch 3 |
| 298 | SubStream i_d(graph); |
| 299 | i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)).set_name("Mixed_5b/Branch_3/AvgPool_0a_3x3") |
| 300 | << ConvolutionLayer(1U, 1U, 64U, |
| 301 | get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout), |
| 302 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 303 | PadStrideInfo(1, 1, 0, 0)) |
| 304 | .set_name("Mixed_5b/Branch_3/Conv2d_0b_1x1/convolution") |
| 305 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"), |
| 306 | get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"), |
| 307 | get_random_accessor(1.f, 1.f), |
| 308 | get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"), |
| 309 | 0.0010000000474974513f) |
| 310 | .set_name("Mixed_5b/Branch_3/Conv2d_0b_1x1/BatchNorm") |
| 311 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_3/Conv2d_0b_1x1/Relu"); |
| 312 | |
| 313 | // Concatenate |
Georgios Pinitas | 427bbbf | 2018-08-28 13:32:02 +0100 | [diff] [blame] | 314 | graph << ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)).set_name("Mixed_5a/concat"); |
Georgios Pinitas | be2772a | 2018-08-17 15:33:39 +0100 | [diff] [blame] | 315 | } |
| 316 | |
| 317 | void block_mixed_6a(const std::string &data_path, DataLayout weights_layout) |
| 318 | { |
| 319 | // Branch 0 |
| 320 | SubStream i_a(graph); |
| 321 | i_a << ConvolutionLayer(3U, 3U, 384U, |
| 322 | get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), |
| 323 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 324 | PadStrideInfo(2, 2, 0, 0)) |
| 325 | .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/convolution") |
| 326 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), |
| 327 | get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), |
| 328 | get_random_accessor(1.f, 1.f), |
| 329 | get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), |
| 330 | 0.0010000000474974513f) |
| 331 | .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/BatchNorm") |
| 332 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/Relu"); |
| 333 | |
| 334 | // Branch 1 |
| 335 | SubStream i_b(graph); |
| 336 | i_b << ConvolutionLayer(1U, 1U, 256U, |
| 337 | get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), |
| 338 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 339 | PadStrideInfo(1, 1, 0, 0)) |
| 340 | .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/convolution") |
| 341 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), |
| 342 | get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), |
| 343 | get_random_accessor(1.f, 1.f), |
| 344 | get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), |
| 345 | 0.0010000000474974513f) |
| 346 | .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/BatchNorm") |
| 347 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/Relu") |
| 348 | << ConvolutionLayer(3U, 3U, 256U, |
| 349 | get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), |
| 350 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 351 | PadStrideInfo(1, 1, 1, 1)) |
| 352 | .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/convolution") |
| 353 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), |
| 354 | get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), |
| 355 | get_random_accessor(1.f, 1.f), |
| 356 | get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), |
| 357 | 0.0010000000474974513f) |
| 358 | .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/BatchNorm") |
| 359 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/Relu") |
| 360 | << ConvolutionLayer(3U, 3U, 384U, |
| 361 | get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), |
| 362 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 363 | PadStrideInfo(2, 2, 0, 0)) |
| 364 | .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/convolution") |
| 365 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), |
| 366 | get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), |
| 367 | get_random_accessor(1.f, 1.f), |
| 368 | get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), |
| 369 | 0.0010000000474974513f) |
| 370 | .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/BatchNorm") |
| 371 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/Relu"); |
| 372 | |
| 373 | // Branch 2 |
| 374 | SubStream i_c(graph); |
| 375 | i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0), true)).set_name("Mixed_6a/Branch_2/MaxPool_1a_3x3"); |
| 376 | |
| 377 | // Concatenate |
Georgios Pinitas | 427bbbf | 2018-08-28 13:32:02 +0100 | [diff] [blame] | 378 | graph << ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c)).set_name("Mixed_6a/concat"); |
Georgios Pinitas | be2772a | 2018-08-17 15:33:39 +0100 | [diff] [blame] | 379 | } |
| 380 | |
| 381 | void block_mixed_7a(const std::string &data_path, DataLayout weights_layout) |
| 382 | { |
| 383 | // Branch 0 |
| 384 | SubStream i_a(graph); |
| 385 | i_a << ConvolutionLayer(1U, 1U, 256U, |
| 386 | get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout), |
| 387 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 388 | PadStrideInfo(1, 1, 0, 0)) |
| 389 | .set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/convolution") |
| 390 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), |
| 391 | get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), |
| 392 | get_random_accessor(1.f, 1.f), |
| 393 | get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"), |
| 394 | 0.0010000000474974513f) |
| 395 | .set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/BatchNorm") |
| 396 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/Relu") |
| 397 | << ConvolutionLayer(3U, 3U, 384U, |
| 398 | get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout), |
| 399 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 400 | PadStrideInfo(2, 2, 0, 0)) |
| 401 | .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/convolution") |
| 402 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), |
| 403 | get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), |
| 404 | get_random_accessor(1.f, 1.f), |
| 405 | get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"), |
| 406 | 0.0010000000474974513f) |
| 407 | .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/BatchNorm") |
| 408 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/Relu"); |
| 409 | |
| 410 | // Branch 1 |
| 411 | SubStream i_b(graph); |
| 412 | i_b << ConvolutionLayer(1U, 1U, 256U, |
| 413 | get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), |
| 414 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 415 | PadStrideInfo(1, 1, 0, 0)) |
| 416 | .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/convolution") |
| 417 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), |
| 418 | get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), |
| 419 | get_random_accessor(1.f, 1.f), |
| 420 | get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), |
| 421 | 0.0010000000474974513f) |
| 422 | .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/BatchNorm") |
| 423 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/Relu") |
| 424 | << ConvolutionLayer(3U, 3U, 288U, |
| 425 | get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout), |
| 426 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 427 | PadStrideInfo(2, 2, 0, 0)) |
| 428 | .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/convolution") |
| 429 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), |
| 430 | get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), |
| 431 | get_random_accessor(1.f, 1.f), |
| 432 | get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"), |
| 433 | 0.0010000000474974513f) |
| 434 | .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/BatchNorm") |
| 435 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/Relu"); |
| 436 | |
| 437 | // Branch 2 |
| 438 | SubStream i_c(graph); |
| 439 | i_c << ConvolutionLayer(1U, 1U, 256U, |
| 440 | get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), |
| 441 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 442 | PadStrideInfo(1, 1, 0, 0)) |
| 443 | .set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/convolution") |
| 444 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), |
| 445 | get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), |
| 446 | get_random_accessor(1.f, 1.f), |
| 447 | get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), |
| 448 | 0.0010000000474974513f) |
| 449 | .set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/BatchNorm") |
| 450 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/Relu") |
| 451 | << ConvolutionLayer(3U, 3U, 288U, |
| 452 | get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), |
| 453 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 454 | PadStrideInfo(1, 1, 1, 1)) |
| 455 | .set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/convolution") |
| 456 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), |
| 457 | get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), |
| 458 | get_random_accessor(1.f, 1.f), |
| 459 | get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), |
| 460 | 0.0010000000474974513f) |
| 461 | .set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/BatchNorm") |
| 462 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/Relu") |
| 463 | << ConvolutionLayer(3U, 3U, 320U, |
| 464 | get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_weights.npy", weights_layout), |
| 465 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 466 | PadStrideInfo(2, 2, 0, 0)) |
| 467 | .set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/convolution") |
| 468 | << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"), |
| 469 | get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"), |
| 470 | get_random_accessor(1.f, 1.f), |
| 471 | get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_beta.npy"), |
| 472 | 0.0010000000474974513f) |
| 473 | .set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/BatchNorm") |
| 474 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/Relu"); |
| 475 | |
| 476 | // Branch 3 |
| 477 | SubStream i_d(graph); |
| 478 | i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)).set_name("Mixed_7a/Branch_3/MaxPool_1a_3x3"); |
| 479 | |
| 480 | // Concatenate |
Georgios Pinitas | 427bbbf | 2018-08-28 13:32:02 +0100 | [diff] [blame] | 481 | graph << ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)).set_name("Mixed_7a/concat"); |
Georgios Pinitas | be2772a | 2018-08-17 15:33:39 +0100 | [diff] [blame] | 482 | } |
| 483 | |
| 484 | void block35_repeat(const std::string &data_path, DataLayout weights_layout, unsigned int num_blocks) |
| 485 | { |
| 486 | for(unsigned int i = 0; i < num_blocks; ++i) |
| 487 | { |
| 488 | std::stringstream unit_path_ss; |
| 489 | unit_path_ss << "Repeat_block35_" << (i + 1) << "_"; |
| 490 | std::stringstream unit_name_ss; |
| 491 | unit_name_ss << "Repeat/block35_" << (i + 1) << "/"; |
| 492 | |
| 493 | std::string unit_path = unit_path_ss.str(); |
| 494 | std::string unit_name = unit_name_ss.str(); |
| 495 | |
| 496 | // Create left and write substreams |
| 497 | SubStream i_l(graph); |
| 498 | SubStream i_r(graph); |
| 499 | |
| 500 | // Branch 0 |
| 501 | SubStream i_la(i_l); |
| 502 | i_la << ConvolutionLayer(1U, 1U, 32U, |
| 503 | get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout), |
| 504 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 505 | PadStrideInfo(1, 1, 0, 0)) |
| 506 | .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution") |
| 507 | << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), |
| 508 | get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), |
| 509 | get_random_accessor(1.f, 1.f), |
| 510 | get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), |
| 511 | 0.0010000000474974513f) |
| 512 | .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm") |
| 513 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_0/Conv2d_1x1/Relu"); |
| 514 | |
| 515 | // Branch 1 |
| 516 | SubStream i_lb(i_l); |
| 517 | i_lb << ConvolutionLayer(1U, 1U, 32U, |
| 518 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), |
| 519 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 520 | PadStrideInfo(1, 1, 0, 0)) |
| 521 | .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution") |
| 522 | << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), |
| 523 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), |
| 524 | get_random_accessor(1.f, 1.f), |
| 525 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), |
| 526 | 0.0010000000474974513f) |
| 527 | .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm") |
| 528 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu") |
| 529 | << ConvolutionLayer(3U, 3U, 32U, |
| 530 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout), |
| 531 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 532 | PadStrideInfo(1, 1, 1, 1)) |
| 533 | .set_name(unit_name + "Branch_1/Conv2d_0b_3x3/convolution") |
| 534 | << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), |
| 535 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), |
| 536 | get_random_accessor(1.f, 1.f), |
| 537 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"), |
| 538 | 0.0010000000474974513f) |
| 539 | .set_name(unit_name + "Branch_1/Conv2d_0b_3x3/BatchNorm") |
| 540 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0b_3x3/Relu"); |
| 541 | |
| 542 | // Branch 2 |
| 543 | SubStream i_lc(i_l); |
| 544 | i_lc << ConvolutionLayer(1U, 1U, 32U, |
| 545 | get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout), |
| 546 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 547 | PadStrideInfo(1, 1, 0, 0)) |
| 548 | .set_name(unit_name + "Branch_2/Conv2d_0a_1x1/convolution") |
| 549 | << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), |
| 550 | get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), |
| 551 | get_random_accessor(1.f, 1.f), |
| 552 | get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"), |
| 553 | 0.0010000000474974513f) |
| 554 | .set_name(unit_name + "Branch_2/Conv2d_0a_1x1/BatchNorm") |
| 555 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_2/Conv2d_0a_1x1/Relu") |
| 556 | << ConvolutionLayer(3U, 3U, 48U, |
| 557 | get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout), |
| 558 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 559 | PadStrideInfo(1, 1, 1, 1)) |
| 560 | .set_name(unit_name + "Branch_2/Conv2d_0b_3x3/convolution") |
| 561 | << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"), |
| 562 | get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"), |
| 563 | get_random_accessor(1.f, 1.f), |
| 564 | get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"), |
| 565 | 0.0010000000474974513f) |
| 566 | .set_name(unit_name + "Branch_2/Conv2d_0b_3x3/BatchNorm") |
| 567 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_2/Conv2d_0b_3x3/Relu") |
| 568 | << ConvolutionLayer(3U, 3U, 64U, |
| 569 | get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_weights.npy", weights_layout), |
| 570 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 571 | PadStrideInfo(1, 1, 1, 1)) |
| 572 | .set_name(unit_name + "Branch_2/Conv2d_0c_3x3/convolution") |
| 573 | << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"), |
| 574 | get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"), |
| 575 | get_random_accessor(1.f, 1.f), |
| 576 | get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"), |
| 577 | 0.0010000000474974513f) |
| 578 | .set_name(unit_name + "Branch_2/Conv2d_0c_3x3/BatchNorm") |
| 579 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_2/Conv2d_0c_3x3/Relu"); |
| 580 | |
| 581 | // Concatenate |
Georgios Pinitas | 427bbbf | 2018-08-28 13:32:02 +0100 | [diff] [blame] | 582 | i_l << ConcatLayer(std::move(i_la), std::move(i_lb), std::move(i_lc)).set_name(unit_name + "concat") |
Georgios Pinitas | be2772a | 2018-08-17 15:33:39 +0100 | [diff] [blame] | 583 | << ConvolutionLayer(1U, 1U, 320U, |
| 584 | get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout), |
| 585 | get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout), |
| 586 | PadStrideInfo(1, 1, 0, 0)) |
| 587 | .set_name(unit_name + "Conv2d_1x1/convolution") |
| 588 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 0.17f, 0.f)).set_name(unit_name + "mul"); |
| 589 | |
Georgios Pinitas | 427bbbf | 2018-08-28 13:32:02 +0100 | [diff] [blame] | 590 | graph << EltwiseLayer(std::move(i_l), std::move(i_r), EltwiseOperation::Add).set_name(unit_name + "add") |
Georgios Pinitas | be2772a | 2018-08-17 15:33:39 +0100 | [diff] [blame] | 591 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu"); |
| 592 | } |
| 593 | } |
| 594 | |
| 595 | void block17_repeat(const std::string &data_path, DataLayout weights_layout, unsigned int num_blocks) |
| 596 | { |
| 597 | for(unsigned int i = 0; i < num_blocks; ++i) |
| 598 | { |
| 599 | std::stringstream unit_path_ss; |
| 600 | unit_path_ss << "Repeat_1_block17_" << (i + 1) << "_"; |
| 601 | std::stringstream unit_name_ss; |
| 602 | unit_name_ss << "Repeat_1/block17_" << (i + 1) << "/"; |
| 603 | |
| 604 | std::string unit_path = unit_path_ss.str(); |
| 605 | std::string unit_name = unit_name_ss.str(); |
| 606 | |
| 607 | // Create left and write substreams |
| 608 | SubStream i_l(graph); |
| 609 | SubStream i_r(graph); |
| 610 | |
| 611 | // Branch 0 |
| 612 | SubStream i_la(i_l); |
| 613 | i_la << ConvolutionLayer(1U, 1U, 192U, |
| 614 | get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout), |
| 615 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 616 | PadStrideInfo(1, 1, 0, 0)) |
| 617 | .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution") |
| 618 | << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), |
| 619 | get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), |
| 620 | get_random_accessor(1.f, 1.f), |
| 621 | get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), |
| 622 | 0.0010000000474974513f) |
| 623 | .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm") |
| 624 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_0/Conv2d_1x1/Relu"); |
| 625 | |
| 626 | // Branch 1 |
| 627 | SubStream i_lb(i_l); |
| 628 | i_lb << ConvolutionLayer(1U, 1U, 128U, |
| 629 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), |
| 630 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 631 | PadStrideInfo(1, 1, 0, 0)) |
| 632 | .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution") |
| 633 | << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), |
| 634 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), |
| 635 | get_random_accessor(1.f, 1.f), |
| 636 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), |
| 637 | 0.0010000000474974513f) |
| 638 | .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm") |
| 639 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu") |
| 640 | << ConvolutionLayer(7U, 1U, 160U, |
| 641 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout), |
| 642 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 643 | PadStrideInfo(1, 1, 3, 0)) |
| 644 | .set_name(unit_name + "Branch_1/Conv2d_0b_1x7/convolution") |
| 645 | << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"), |
| 646 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"), |
| 647 | get_random_accessor(1.f, 1.f), |
| 648 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"), |
| 649 | 0.0010000000474974513f) |
| 650 | .set_name(unit_name + "Branch_1/Conv2d_0b_1x7/BatchNorm") |
| 651 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0b_1x7/Relu") |
| 652 | << ConvolutionLayer(1U, 7U, 192U, |
| 653 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout), |
| 654 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 655 | PadStrideInfo(1, 1, 0, 3)) |
| 656 | .set_name(unit_name + "Branch_1/Conv2d_0c_7x1/convolution") |
| 657 | << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"), |
| 658 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"), |
| 659 | get_random_accessor(1.f, 1.f), |
| 660 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"), |
| 661 | 0.0010000000474974513f) |
| 662 | .set_name(unit_name + "Branch_1/Conv2d_0c_7x1/BatchNorm") |
| 663 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0c_7x1/Relu"); |
| 664 | |
| 665 | // Concatenate |
Georgios Pinitas | 427bbbf | 2018-08-28 13:32:02 +0100 | [diff] [blame] | 666 | i_l << ConcatLayer(std::move(i_la), std::move(i_lb)).set_name(unit_name + "concat") |
Georgios Pinitas | be2772a | 2018-08-17 15:33:39 +0100 | [diff] [blame] | 667 | << ConvolutionLayer(1U, 1U, 1088U, |
| 668 | get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout), |
| 669 | get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout), |
| 670 | PadStrideInfo(1, 1, 0, 0)) |
| 671 | .set_name(unit_name + "Conv2d_1x1/convolution") |
| 672 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 0.10f, 0.f)).set_name(unit_name + "mul"); |
| 673 | |
Georgios Pinitas | 427bbbf | 2018-08-28 13:32:02 +0100 | [diff] [blame] | 674 | graph << EltwiseLayer(std::move(i_l), std::move(i_r), EltwiseOperation::Add).set_name(unit_name + "add") |
Georgios Pinitas | be2772a | 2018-08-17 15:33:39 +0100 | [diff] [blame] | 675 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu"); |
| 676 | } |
| 677 | } |
| 678 | |
| 679 | void block8_repeat(const std::string &data_path, DataLayout weights_layout, unsigned int num_blocks, float scale, bool has_activation) |
| 680 | { |
| 681 | for(unsigned int i = 0; i < num_blocks; ++i) |
| 682 | { |
| 683 | std::stringstream unit_path_ss; |
| 684 | std::stringstream unit_name_ss; |
| 685 | if(num_blocks != 1) |
| 686 | { |
| 687 | unit_path_ss << "Repeat_2_block8_" << (i + 1) << "_"; |
| 688 | unit_name_ss << "Repeat_2/block8_" << (i + 1) << "/"; |
| 689 | } |
| 690 | else |
| 691 | { |
| 692 | unit_path_ss << "Block8_"; |
| 693 | unit_name_ss << "Block8/"; |
| 694 | } |
| 695 | |
| 696 | std::string unit_path = unit_path_ss.str(); |
| 697 | std::string unit_name = unit_name_ss.str(); |
| 698 | |
| 699 | // Create left and write substreams |
| 700 | SubStream i_l(graph); |
| 701 | SubStream i_r(graph); |
| 702 | |
| 703 | // Branch 0 |
| 704 | SubStream i_la(i_l); |
| 705 | i_la << ConvolutionLayer(1U, 1U, 192U, |
| 706 | get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout), |
| 707 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 708 | PadStrideInfo(1, 1, 0, 0)) |
| 709 | .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution") |
| 710 | << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"), |
| 711 | get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"), |
| 712 | get_random_accessor(1.f, 1.f), |
| 713 | get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"), |
| 714 | 0.0010000000474974513f) |
| 715 | .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm") |
| 716 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_0/Conv2d_1x1/Relu"); |
| 717 | |
| 718 | // Branch 1 |
| 719 | SubStream i_lb(i_l); |
| 720 | i_lb << ConvolutionLayer(1U, 1U, 192U, |
| 721 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout), |
| 722 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 723 | PadStrideInfo(1, 1, 0, 0)) |
| 724 | .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution") |
| 725 | << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"), |
| 726 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"), |
| 727 | get_random_accessor(1.f, 1.f), |
| 728 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"), |
| 729 | 0.0010000000474974513f) |
| 730 | .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm") |
| 731 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu") |
| 732 | << ConvolutionLayer(3U, 1U, 224U, |
| 733 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_weights.npy", weights_layout), |
| 734 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 735 | PadStrideInfo(1, 1, 1, 0)) |
| 736 | .set_name(unit_name + "Branch_1/Conv2d_0b_1x3/convolution") |
| 737 | << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"), |
| 738 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"), |
| 739 | get_random_accessor(1.f, 1.f), |
| 740 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"), |
| 741 | 0.0010000000474974513f) |
| 742 | .set_name(unit_name + "Branch_1/Conv2d_0b_1x3/BatchNorm") |
| 743 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0b_1x3/Relu") |
| 744 | << ConvolutionLayer(1U, 3U, 256U, |
| 745 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_weights.npy", weights_layout), |
| 746 | std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), |
| 747 | PadStrideInfo(1, 1, 0, 1)) |
| 748 | .set_name(unit_name + "Branch_1/Conv2d_0c_3x1/convolution") |
| 749 | << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_mean.npy"), |
| 750 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_variance.npy"), |
| 751 | get_random_accessor(1.f, 1.f), |
| 752 | get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_beta.npy"), |
| 753 | 0.0010000000474974513f) |
| 754 | .set_name(unit_name + "Branch_1/Conv2d_0c_3x1/BatchNorm") |
| 755 | << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0c_3x1/Relu"); |
| 756 | |
| 757 | // Concatenate |
Georgios Pinitas | 427bbbf | 2018-08-28 13:32:02 +0100 | [diff] [blame] | 758 | i_l << ConcatLayer(std::move(i_la), std::move(i_lb)).set_name(unit_name + "concat") |
Georgios Pinitas | be2772a | 2018-08-17 15:33:39 +0100 | [diff] [blame] | 759 | << ConvolutionLayer(1U, 1U, 2080U, |
| 760 | get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout), |
| 761 | get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout), |
| 762 | PadStrideInfo(1, 1, 0, 0)) |
| 763 | .set_name(unit_name + "Conv2d_1x1/convolution"); |
| 764 | |
| 765 | // Scale result |
| 766 | if(scale != 1.f) |
| 767 | { |
| 768 | i_l << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, scale, 0.f)).set_name(unit_name + "mul"); |
| 769 | } |
| 770 | |
| 771 | // Residual add |
Georgios Pinitas | 427bbbf | 2018-08-28 13:32:02 +0100 | [diff] [blame] | 772 | graph << EltwiseLayer(std::move(i_l), std::move(i_r), EltwiseOperation::Add).set_name(unit_name + "add"); |
Georgios Pinitas | be2772a | 2018-08-17 15:33:39 +0100 | [diff] [blame] | 773 | |
| 774 | // Apply activation if needed |
| 775 | if(has_activation) |
| 776 | { |
| 777 | graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu"); |
| 778 | } |
| 779 | } |
| 780 | } |
| 781 | }; |
| 782 | |
| 783 | /** Main program for Inception ResNet V2 |
| 784 | * |
Georgios Pinitas | bdbbbe8 | 2018-11-07 16:06:47 +0000 | [diff] [blame] | 785 | * Model is based on: |
| 786 | * https://arxiv.org/abs/1602.07261 |
| 787 | * "Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning" |
| 788 | * Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, Alex Alemi |
| 789 | * |
Georgios Pinitas | 588ebc5 | 2018-12-21 13:39:07 +0000 | [diff] [blame] | 790 | * Provenance: download.tensorflow.org/models/inception_resnet_v2_2016_08_30.tar.gz |
| 791 | * |
Georgios Pinitas | be2772a | 2018-08-17 15:33:39 +0100 | [diff] [blame] | 792 | * @note To list all the possible arguments execute the binary appended with the --help option |
| 793 | * |
| 794 | * @param[in] argc Number of arguments |
| 795 | * @param[in] argv Arguments |
| 796 | */ |
| 797 | int main(int argc, char **argv) |
| 798 | { |
| 799 | return arm_compute::utils::run_example<InceptionResNetV2Example>(argc, argv); |
| 800 | } |