Compute Library
 21.02
graph_inception_resnet_v2.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2018-2021 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"
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 
34 /** Example demonstrating how to implement InceptionV4's network using the Compute Library's graph API */
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  cmd_parser.validate();
47 
48  // Consume common parameters
49  common_params = consume_common_graph_parameters(common_opts);
50 
51  // Return when help menu is requested
52  if(common_params.help)
53  {
54  cmd_parser.print_help(argv[0]);
55  return false;
56  }
57 
58  // Set default layout if needed
59  if(!common_opts.data_layout->is_set() && common_params.target == Target::NEON)
60  {
61  common_params.data_layout = DataLayout::NCHW;
62  }
63 
64  // Checks
65  ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph");
66 
67  // Print parameter values
68  std::cout << common_params << std::endl;
69 
70  // Create model path
71  std::string data_path = common_params.data_path;
72  std::string model_path = "/cnn_data/inception_resnet_v2_model/";
73  if(!data_path.empty())
74  {
75  data_path += model_path;
76  }
77 
78  // Create a preprocessor object
79  std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<TFPreproccessor>(0.f, 1.f);
80 
81  // Create input descriptor
82  const auto operation_layout = common_params.data_layout;
83  const TensorShape tensor_shape = permute_shape(TensorShape(299U, 299U, 3U, 1U), DataLayout::NCHW, operation_layout);
84  TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout);
85 
86  // Set weights trained layout
87  const DataLayout weights_layout = DataLayout::NCHW;
88 
89  graph << common_params.target
90  << common_params.fast_math_hint
91  << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false))
92  // Conv2d_1a_3x3
93  << ConvolutionLayer(3U, 3U, 32U,
94  get_weights_accessor(data_path, "Conv2d_1a_3x3_weights.npy", weights_layout),
95  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
96  PadStrideInfo(2, 2, 0, 0))
97  .set_name("Conv2d_1a_3x3/convolution")
98  << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
99  get_weights_accessor(data_path, "Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
100  get_random_accessor(1.f, 1.f),
101  get_weights_accessor(data_path, "Conv2d_1a_3x3_BatchNorm_beta.npy"),
102  0.0010000000474974513f)
103  .set_name("Conv2d_1a_3x3/BatchNorm")
104  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_1a_3x3/Relu")
105  // Conv2d_2a_3x3
106  << ConvolutionLayer(3U, 3U, 32U,
107  get_weights_accessor(data_path, "Conv2d_2a_3x3_weights.npy", weights_layout),
108  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
109  PadStrideInfo(1, 1, 0, 0))
110  .set_name("Conv2d_2a_3x3/convolution")
111  << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_2a_3x3_BatchNorm_moving_mean.npy"),
112  get_weights_accessor(data_path, "Conv2d_2a_3x3_BatchNorm_moving_variance.npy"),
113  get_random_accessor(1.f, 1.f),
114  get_weights_accessor(data_path, "Conv2d_2a_3x3_BatchNorm_beta.npy"),
115  0.0010000000474974513f)
116  .set_name("Conv2d_2a_3x3/BatchNorm")
117  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2a_3x3/Relu")
118  // Conv2d_2b_3x3
119  << ConvolutionLayer(3U, 3U, 64U,
120  get_weights_accessor(data_path, "Conv2d_2b_3x3_weights.npy", weights_layout),
121  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
122  PadStrideInfo(1, 1, 1, 1))
123  .set_name("Conv2d_2b_3x3/convolution")
124  << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_2b_3x3_BatchNorm_moving_mean.npy"),
125  get_weights_accessor(data_path, "Conv2d_2b_3x3_BatchNorm_moving_variance.npy"),
126  get_random_accessor(1.f, 1.f),
127  get_weights_accessor(data_path, "Conv2d_2b_3x3_BatchNorm_beta.npy"),
128  0.0010000000474974513f)
129  .set_name("Conv2d_2b_3x3/BatchNorm")
130  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_2b_3x3/Relu")
131  // MaxPool_3a_3x3
132  << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)).set_name("MaxPool_3a_3x3/MaxPool")
133  // Conv2d_3b_1x1
134  << ConvolutionLayer(1U, 1U, 80U,
135  get_weights_accessor(data_path, "Conv2d_3b_1x1_weights.npy", weights_layout),
136  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
137  PadStrideInfo(1, 1, 0, 0))
138  .set_name("Conv2d_3b_1x1/convolution")
139  << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_3b_1x1_BatchNorm_moving_mean.npy"),
140  get_weights_accessor(data_path, "Conv2d_3b_1x1_BatchNorm_moving_variance.npy"),
141  get_random_accessor(1.f, 1.f),
142  get_weights_accessor(data_path, "Conv2d_3b_1x1_BatchNorm_beta.npy"),
143  0.0010000000474974513f)
144  .set_name("Conv2d_3b_1x1/BatchNorm")
145  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_3b_1x1/Relu")
146  // Conv2d_4a_3x3
147  << ConvolutionLayer(3U, 3U, 192U,
148  get_weights_accessor(data_path, "Conv2d_4a_3x3_weights.npy", weights_layout),
149  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
150  PadStrideInfo(1, 1, 0, 0))
151  .set_name("Conv2d_4a_3x3/convolution")
152  << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_4a_3x3_BatchNorm_moving_mean.npy"),
153  get_weights_accessor(data_path, "Conv2d_4a_3x3_BatchNorm_moving_variance.npy"),
154  get_random_accessor(1.f, 1.f),
155  get_weights_accessor(data_path, "Conv2d_4a_3x3_BatchNorm_beta.npy"),
156  0.0010000000474974513f)
157  .set_name("Conv2d_4a_3x3/BatchNorm")
158  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_4a_3x3/Relu")
159  // MaxPool_5a_3x3
160  << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0), true)).set_name("MaxPool_5a_3x3/MaxPool");
161 
162  block_mixed_5b(data_path, weights_layout);
163  block35_repeat(data_path, weights_layout, 10);
164  block_mixed_6a(data_path, weights_layout);
165  block17_repeat(data_path, weights_layout, 20);
166  block_mixed_7a(data_path, weights_layout);
167  block8_repeat(data_path, weights_layout, 9, 0.2f, true);
168  block8_repeat(data_path, weights_layout, 1, 1.f, false);
169 
170  // Conv2d_7b_1x1
171  graph << ConvolutionLayer(1U, 1U, 1536U,
172  get_weights_accessor(data_path, "Conv2d_7b_1x1_weights.npy", weights_layout),
173  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
174  PadStrideInfo(1, 1, 0, 0))
175  .set_name("Conv2d_7b_1x1/convolution")
176  << BatchNormalizationLayer(get_weights_accessor(data_path, "Conv2d_7b_1x1_BatchNorm_moving_mean.npy"),
177  get_weights_accessor(data_path, "Conv2d_7b_1x1_BatchNorm_moving_variance.npy"),
178  get_random_accessor(1.f, 1.f),
179  get_weights_accessor(data_path, "Conv2d_7b_1x1_BatchNorm_beta.npy"),
180  0.0010000000474974513f)
181  .set_name("Conv2d_7b_1x1/BatchNorm")
182  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv2d_7b_1x1/Relu")
183  << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("Logits/AvgPool_1a_8x8")
184  << FlattenLayer().set_name("Logits/Flatten")
186  1001U,
187  get_weights_accessor(data_path, "Logits_Logits_weights.npy", weights_layout),
188  get_weights_accessor(data_path, "Logits_Logits_biases.npy"))
189  .set_name("Logits/Logits")
190  << SoftmaxLayer().set_name("Logits/Predictions")
191  << OutputLayer(get_output_accessor(common_params, 5));
192 
193  // Finalize graph
194  GraphConfig config;
195  config.num_threads = common_params.threads;
196  config.use_tuner = common_params.enable_tuner;
197  config.tuner_mode = common_params.tuner_mode;
198  config.tuner_file = common_params.tuner_file;
199  config.mlgo_file = common_params.mlgo_file;
200 
201  graph.finalize(common_params.target, config);
202 
203  return true;
204  }
205 
206  void do_run() override
207  {
208  graph.run();
209  }
210 
211 private:
212  CommandLineParser cmd_parser;
213  CommonGraphOptions common_opts;
214  CommonGraphParams common_params;
215  Stream graph;
216 
217 private:
218  void block_mixed_5b(const std::string &data_path, DataLayout weights_layout)
219  {
220  // Branch 0
221  SubStream i_a(graph);
222  i_a << ConvolutionLayer(1U, 1U, 96U,
223  get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_weights.npy", weights_layout),
224  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
225  PadStrideInfo(1, 1, 0, 0))
226  .set_name("Mixed_5b/Branch_0/Conv2d_1x1/convolution")
227  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"),
228  get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"),
229  get_random_accessor(1.f, 1.f),
230  get_weights_accessor(data_path, "Mixed_5b_Branch_0_Conv2d_1x1_BatchNorm_beta.npy"),
231  0.0010000000474974513f)
232  .set_name("Mixed_5b/Branch_0/Conv2d_1x1/BatchNorm")
233  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_0/Conv2d_1x1/Relu");
234 
235  // Branch 1
236  SubStream i_b(graph);
237  i_b << ConvolutionLayer(1U, 1U, 48U,
238  get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout),
239  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
240  PadStrideInfo(1, 1, 0, 0))
241  .set_name("Mixed_5b/Branch_1/Conv2d_0a_1x1/convolution")
242  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
243  get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
244  get_random_accessor(1.f, 1.f),
245  get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
246  0.0010000000474974513f)
247  .set_name("Mixed_5b/Branch_1/Conv2d_0a_1x1/BatchNorm")
248  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_1/Conv2d_0a_1x1/Relu")
249  << ConvolutionLayer(5U, 5U, 64U,
250  get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_weights.npy", weights_layout),
251  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
252  PadStrideInfo(1, 1, 2, 2))
253  .set_name("Mixed_5b/Branch_1/Conv2d_0b_5x5/convolution")
254  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_BatchNorm_moving_mean.npy"),
255  get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_BatchNorm_moving_variance.npy"),
256  get_random_accessor(1.f, 1.f),
257  get_weights_accessor(data_path, "Mixed_5b_Branch_1_Conv2d_0b_5x5_BatchNorm_beta.npy"),
258  0.0010000000474974513f)
259  .set_name("Mixed_5b/Branch_1/Conv2d_0b_5x5/BatchNorm")
260  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_1/Conv2d_0b_5x5/Relu");
261 
262  // Branch 2
263  SubStream i_c(graph);
264  i_c << ConvolutionLayer(1U, 1U, 64U,
265  get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout),
266  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
267  PadStrideInfo(1, 1, 0, 0))
268  .set_name("Mixed_5b/Branch_2/Conv2d_0a_1x1/convolution")
269  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
270  get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
271  get_random_accessor(1.f, 1.f),
272  get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"),
273  0.0010000000474974513f)
274  .set_name("Mixed_5b/Branch_2/Conv2d_0a_1x1/BatchNorm")
275  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_2/Conv2d_0a_1x1/Relu")
276  << ConvolutionLayer(3U, 3U, 96U,
277  get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout),
278  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
279  PadStrideInfo(1, 1, 1, 1))
280  .set_name("Mixed_5b/Branch_2/Conv2d_0b_3x3/convolution")
281  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
282  get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
283  get_random_accessor(1.f, 1.f),
284  get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"),
285  0.0010000000474974513f)
286  .set_name("Mixed_5b/Branch_2/Conv2d_0b_3x3/BatchNorm")
287  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_2/Conv2d_0b_3x3/Relu")
288  << ConvolutionLayer(3U, 3U, 96U,
289  get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_weights.npy", weights_layout),
290  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
291  PadStrideInfo(1, 1, 1, 1))
292  .set_name("Mixed_5b/Branch_2/Conv2d_0c_3x3/convolution")
293  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"),
294  get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"),
295  get_random_accessor(1.f, 1.f),
296  get_weights_accessor(data_path, "Mixed_5b_Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"),
297  0.0010000000474974513f)
298  .set_name("Mixed_5b/Branch_2/Conv2d_0c_3x3/BatchNorm")
299  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_2/Conv2d_0c_3x3/Relu");
300 
301  // Branch 3
302  SubStream i_d(graph);
303  i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true)).set_name("Mixed_5b/Branch_3/AvgPool_0a_3x3")
304  << ConvolutionLayer(1U, 1U, 64U,
305  get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_weights.npy", weights_layout),
306  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
307  PadStrideInfo(1, 1, 0, 0))
308  .set_name("Mixed_5b/Branch_3/Conv2d_0b_1x1/convolution")
309  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"),
310  get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"),
311  get_random_accessor(1.f, 1.f),
312  get_weights_accessor(data_path, "Mixed_5b_Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"),
313  0.0010000000474974513f)
314  .set_name("Mixed_5b/Branch_3/Conv2d_0b_1x1/BatchNorm")
315  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_5b/Branch_3/Conv2d_0b_1x1/Relu");
316 
317  // Concatenate
318  graph << ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)).set_name("Mixed_5a/concat");
319  }
320 
321  void block_mixed_6a(const std::string &data_path, DataLayout weights_layout)
322  {
323  // Branch 0
324  SubStream i_a(graph);
325  i_a << ConvolutionLayer(3U, 3U, 384U,
326  get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout),
327  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
328  PadStrideInfo(2, 2, 0, 0))
329  .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/convolution")
330  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
331  get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
332  get_random_accessor(1.f, 1.f),
333  get_weights_accessor(data_path, "Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"),
334  0.0010000000474974513f)
335  .set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/BatchNorm")
336  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_0/Conv2d_1a_3x3/Relu");
337 
338  // Branch 1
339  SubStream i_b(graph);
340  i_b << ConvolutionLayer(1U, 1U, 256U,
341  get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout),
342  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
343  PadStrideInfo(1, 1, 0, 0))
344  .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/convolution")
345  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
346  get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
347  get_random_accessor(1.f, 1.f),
348  get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
349  0.0010000000474974513f)
350  .set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/BatchNorm")
351  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_1/Conv2d_0a_1x1/Relu")
352  << ConvolutionLayer(3U, 3U, 256U,
353  get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout),
354  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
355  PadStrideInfo(1, 1, 1, 1))
356  .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/convolution")
357  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
358  get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
359  get_random_accessor(1.f, 1.f),
360  get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"),
361  0.0010000000474974513f)
362  .set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/BatchNorm")
363  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_1/Conv2d_0b_3x3/Relu")
364  << ConvolutionLayer(3U, 3U, 384U,
365  get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout),
366  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
367  PadStrideInfo(2, 2, 0, 0))
368  .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/convolution")
369  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
370  get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
371  get_random_accessor(1.f, 1.f),
372  get_weights_accessor(data_path, "Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"),
373  0.0010000000474974513f)
374  .set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/BatchNorm")
375  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_6a/Branch_1/Conv2d_1a_3x3/Relu");
376 
377  // Branch 2
378  SubStream i_c(graph);
379  i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(2, 2, 0, 0), true)).set_name("Mixed_6a/Branch_2/MaxPool_1a_3x3");
380 
381  // Concatenate
382  graph << ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c)).set_name("Mixed_6a/concat");
383  }
384 
385  void block_mixed_7a(const std::string &data_path, DataLayout weights_layout)
386  {
387  // Branch 0
388  SubStream i_a(graph);
389  i_a << ConvolutionLayer(1U, 1U, 256U,
390  get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_weights.npy", weights_layout),
391  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
392  PadStrideInfo(1, 1, 0, 0))
393  .set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/convolution")
394  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
395  get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
396  get_random_accessor(1.f, 1.f),
397  get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
398  0.0010000000474974513f)
399  .set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/BatchNorm")
400  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_0/Conv2d_0a_1x1/Relu")
401  << ConvolutionLayer(3U, 3U, 384U,
402  get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_weights.npy", weights_layout),
403  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
404  PadStrideInfo(2, 2, 0, 0))
405  .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/convolution")
406  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
407  get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
408  get_random_accessor(1.f, 1.f),
409  get_weights_accessor(data_path, "Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"),
410  0.0010000000474974513f)
411  .set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/BatchNorm")
412  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_0/Conv2d_1a_3x3/Relu");
413 
414  // Branch 1
415  SubStream i_b(graph);
416  i_b << ConvolutionLayer(1U, 1U, 256U,
417  get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout),
418  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
419  PadStrideInfo(1, 1, 0, 0))
420  .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/convolution")
421  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
422  get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
423  get_random_accessor(1.f, 1.f),
424  get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
425  0.0010000000474974513f)
426  .set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/BatchNorm")
427  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_1/Conv2d_0a_1x1/Relu")
428  << ConvolutionLayer(3U, 3U, 288U,
429  get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_weights.npy", weights_layout),
430  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
431  PadStrideInfo(2, 2, 0, 0))
432  .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/convolution")
433  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
434  get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
435  get_random_accessor(1.f, 1.f),
436  get_weights_accessor(data_path, "Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"),
437  0.0010000000474974513f)
438  .set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/BatchNorm")
439  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_1/Conv2d_1a_3x3/Relu");
440 
441  // Branch 2
442  SubStream i_c(graph);
443  i_c << ConvolutionLayer(1U, 1U, 256U,
444  get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout),
445  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
446  PadStrideInfo(1, 1, 0, 0))
447  .set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/convolution")
448  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
449  get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
450  get_random_accessor(1.f, 1.f),
451  get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"),
452  0.0010000000474974513f)
453  .set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/BatchNorm")
454  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_2/Conv2d_0a_1x1/Relu")
455  << ConvolutionLayer(3U, 3U, 288U,
456  get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout),
457  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
458  PadStrideInfo(1, 1, 1, 1))
459  .set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/convolution")
460  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
461  get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
462  get_random_accessor(1.f, 1.f),
463  get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"),
464  0.0010000000474974513f)
465  .set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/BatchNorm")
466  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_2/Conv2d_0b_3x3/Relu")
467  << ConvolutionLayer(3U, 3U, 320U,
468  get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_weights.npy", weights_layout),
469  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
470  PadStrideInfo(2, 2, 0, 0))
471  .set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/convolution")
472  << BatchNormalizationLayer(get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
473  get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
474  get_random_accessor(1.f, 1.f),
475  get_weights_accessor(data_path, "Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm_beta.npy"),
476  0.0010000000474974513f)
477  .set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/BatchNorm")
478  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Mixed_7a/Branch_2/Conv2d_1a_3x3/Relu");
479 
480  // Branch 3
481  SubStream i_d(graph);
482  i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true)).set_name("Mixed_7a/Branch_3/MaxPool_1a_3x3");
483 
484  // Concatenate
485  graph << ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d)).set_name("Mixed_7a/concat");
486  }
487 
488  void block35_repeat(const std::string &data_path, DataLayout weights_layout, unsigned int num_blocks)
489  {
490  for(unsigned int i = 0; i < num_blocks; ++i)
491  {
492  std::stringstream unit_path_ss;
493  unit_path_ss << "Repeat_block35_" << (i + 1) << "_";
494  std::stringstream unit_name_ss;
495  unit_name_ss << "Repeat/block35_" << (i + 1) << "/";
496 
497  std::string unit_path = unit_path_ss.str();
498  std::string unit_name = unit_name_ss.str();
499 
500  // Create left and write substreams
501  SubStream i_l(graph);
502  SubStream i_r(graph);
503 
504  // Branch 0
505  SubStream i_la(i_l);
506  i_la << ConvolutionLayer(1U, 1U, 32U,
507  get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout),
508  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
509  PadStrideInfo(1, 1, 0, 0))
510  .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution")
511  << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"),
512  get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"),
513  get_random_accessor(1.f, 1.f),
514  get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"),
515  0.0010000000474974513f)
516  .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm")
517  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_0/Conv2d_1x1/Relu");
518 
519  // Branch 1
520  SubStream i_lb(i_l);
521  i_lb << ConvolutionLayer(1U, 1U, 32U,
522  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout),
523  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
524  PadStrideInfo(1, 1, 0, 0))
525  .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution")
526  << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
527  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
528  get_random_accessor(1.f, 1.f),
529  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
530  0.0010000000474974513f)
531  .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm")
532  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu")
533  << ConvolutionLayer(3U, 3U, 32U,
534  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_weights.npy", weights_layout),
535  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
536  PadStrideInfo(1, 1, 1, 1))
537  .set_name(unit_name + "Branch_1/Conv2d_0b_3x3/convolution")
538  << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
539  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
540  get_random_accessor(1.f, 1.f),
541  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"),
542  0.0010000000474974513f)
543  .set_name(unit_name + "Branch_1/Conv2d_0b_3x3/BatchNorm")
544  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0b_3x3/Relu");
545 
546  // Branch 2
547  SubStream i_lc(i_l);
548  i_lc << ConvolutionLayer(1U, 1U, 32U,
549  get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_weights.npy", weights_layout),
550  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
551  PadStrideInfo(1, 1, 0, 0))
552  .set_name(unit_name + "Branch_2/Conv2d_0a_1x1/convolution")
553  << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
554  get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
555  get_random_accessor(1.f, 1.f),
556  get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"),
557  0.0010000000474974513f)
558  .set_name(unit_name + "Branch_2/Conv2d_0a_1x1/BatchNorm")
559  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_2/Conv2d_0a_1x1/Relu")
560  << ConvolutionLayer(3U, 3U, 48U,
561  get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_weights.npy", weights_layout),
562  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
563  PadStrideInfo(1, 1, 1, 1))
564  .set_name(unit_name + "Branch_2/Conv2d_0b_3x3/convolution")
565  << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
566  get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
567  get_random_accessor(1.f, 1.f),
568  get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"),
569  0.0010000000474974513f)
570  .set_name(unit_name + "Branch_2/Conv2d_0b_3x3/BatchNorm")
571  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_2/Conv2d_0b_3x3/Relu")
572  << ConvolutionLayer(3U, 3U, 64U,
573  get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_weights.npy", weights_layout),
574  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
575  PadStrideInfo(1, 1, 1, 1))
576  .set_name(unit_name + "Branch_2/Conv2d_0c_3x3/convolution")
577  << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"),
578  get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"),
579  get_random_accessor(1.f, 1.f),
580  get_weights_accessor(data_path, unit_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"),
581  0.0010000000474974513f)
582  .set_name(unit_name + "Branch_2/Conv2d_0c_3x3/BatchNorm")
583  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_2/Conv2d_0c_3x3/Relu");
584 
585  // Concatenate
586  i_l << ConcatLayer(std::move(i_la), std::move(i_lb), std::move(i_lc)).set_name(unit_name + "concat")
587  << ConvolutionLayer(1U, 1U, 320U,
588  get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout),
589  get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout),
590  PadStrideInfo(1, 1, 0, 0))
591  .set_name(unit_name + "Conv2d_1x1/convolution")
592  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 0.17f, 0.f)).set_name(unit_name + "mul");
593 
594  graph << EltwiseLayer(std::move(i_l), std::move(i_r), EltwiseOperation::Add).set_name(unit_name + "add")
595  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu");
596  }
597  }
598 
599  void block17_repeat(const std::string &data_path, DataLayout weights_layout, unsigned int num_blocks)
600  {
601  for(unsigned int i = 0; i < num_blocks; ++i)
602  {
603  std::stringstream unit_path_ss;
604  unit_path_ss << "Repeat_1_block17_" << (i + 1) << "_";
605  std::stringstream unit_name_ss;
606  unit_name_ss << "Repeat_1/block17_" << (i + 1) << "/";
607 
608  std::string unit_path = unit_path_ss.str();
609  std::string unit_name = unit_name_ss.str();
610 
611  // Create left and write substreams
612  SubStream i_l(graph);
613  SubStream i_r(graph);
614 
615  // Branch 0
616  SubStream i_la(i_l);
617  i_la << ConvolutionLayer(1U, 1U, 192U,
618  get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout),
619  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
620  PadStrideInfo(1, 1, 0, 0))
621  .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution")
622  << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"),
623  get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"),
624  get_random_accessor(1.f, 1.f),
625  get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"),
626  0.0010000000474974513f)
627  .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm")
628  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_0/Conv2d_1x1/Relu");
629 
630  // Branch 1
631  SubStream i_lb(i_l);
632  i_lb << ConvolutionLayer(1U, 1U, 128U,
633  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout),
634  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
635  PadStrideInfo(1, 1, 0, 0))
636  .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution")
637  << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
638  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
639  get_random_accessor(1.f, 1.f),
640  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
641  0.0010000000474974513f)
642  .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm")
643  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu")
644  << ConvolutionLayer(7U, 1U, 160U,
645  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_weights.npy", weights_layout),
646  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
647  PadStrideInfo(1, 1, 3, 0))
648  .set_name(unit_name + "Branch_1/Conv2d_0b_1x7/convolution")
649  << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"),
650  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"),
651  get_random_accessor(1.f, 1.f),
652  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"),
653  0.0010000000474974513f)
654  .set_name(unit_name + "Branch_1/Conv2d_0b_1x7/BatchNorm")
655  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0b_1x7/Relu")
656  << ConvolutionLayer(1U, 7U, 192U,
657  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_weights.npy", weights_layout),
658  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
659  PadStrideInfo(1, 1, 0, 3))
660  .set_name(unit_name + "Branch_1/Conv2d_0c_7x1/convolution")
661  << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"),
662  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"),
663  get_random_accessor(1.f, 1.f),
664  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"),
665  0.0010000000474974513f)
666  .set_name(unit_name + "Branch_1/Conv2d_0c_7x1/BatchNorm")
667  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0c_7x1/Relu");
668 
669  // Concatenate
670  i_l << ConcatLayer(std::move(i_la), std::move(i_lb)).set_name(unit_name + "concat")
671  << ConvolutionLayer(1U, 1U, 1088U,
672  get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout),
673  get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout),
674  PadStrideInfo(1, 1, 0, 0))
675  .set_name(unit_name + "Conv2d_1x1/convolution")
676  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 0.10f, 0.f)).set_name(unit_name + "mul");
677 
678  graph << EltwiseLayer(std::move(i_l), std::move(i_r), EltwiseOperation::Add).set_name(unit_name + "add")
679  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu");
680  }
681  }
682 
683  void block8_repeat(const std::string &data_path, DataLayout weights_layout, unsigned int num_blocks, float scale, bool has_activation)
684  {
685  for(unsigned int i = 0; i < num_blocks; ++i)
686  {
687  std::stringstream unit_path_ss;
688  std::stringstream unit_name_ss;
689  if(num_blocks != 1)
690  {
691  unit_path_ss << "Repeat_2_block8_" << (i + 1) << "_";
692  unit_name_ss << "Repeat_2/block8_" << (i + 1) << "/";
693  }
694  else
695  {
696  unit_path_ss << "Block8_";
697  unit_name_ss << "Block8/";
698  }
699 
700  std::string unit_path = unit_path_ss.str();
701  std::string unit_name = unit_name_ss.str();
702 
703  // Create left and write substreams
704  SubStream i_l(graph);
705  SubStream i_r(graph);
706 
707  // Branch 0
708  SubStream i_la(i_l);
709  i_la << ConvolutionLayer(1U, 1U, 192U,
710  get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_weights.npy", weights_layout),
711  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
712  PadStrideInfo(1, 1, 0, 0))
713  .set_name(unit_name + "Branch_0/Conv2d_1x1/convolution")
714  << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_mean.npy"),
715  get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_moving_variance.npy"),
716  get_random_accessor(1.f, 1.f),
717  get_weights_accessor(data_path, unit_path + "Branch_0_Conv2d_1x1_BatchNorm_beta.npy"),
718  0.0010000000474974513f)
719  .set_name(unit_name + "Branch_0/Conv2d_1x1/BatchNorm")
720  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_0/Conv2d_1x1/Relu");
721 
722  // Branch 1
723  SubStream i_lb(i_l);
724  i_lb << ConvolutionLayer(1U, 1U, 192U,
725  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_weights.npy", weights_layout),
726  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
727  PadStrideInfo(1, 1, 0, 0))
728  .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/convolution")
729  << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
730  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
731  get_random_accessor(1.f, 1.f),
732  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
733  0.0010000000474974513f)
734  .set_name(unit_name + "Branch_1/Conv2d_0a_1x1/BatchNorm")
735  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0a_1x1/Relu")
736  << ConvolutionLayer(3U, 1U, 224U,
737  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_weights.npy", weights_layout),
738  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
739  PadStrideInfo(1, 1, 1, 0))
740  .set_name(unit_name + "Branch_1/Conv2d_0b_1x3/convolution")
741  << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"),
742  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"),
743  get_random_accessor(1.f, 1.f),
744  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"),
745  0.0010000000474974513f)
746  .set_name(unit_name + "Branch_1/Conv2d_0b_1x3/BatchNorm")
747  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0b_1x3/Relu")
748  << ConvolutionLayer(1U, 3U, 256U,
749  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_weights.npy", weights_layout),
750  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
751  PadStrideInfo(1, 1, 0, 1))
752  .set_name(unit_name + "Branch_1/Conv2d_0c_3x1/convolution")
753  << BatchNormalizationLayer(get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_mean.npy"),
754  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_variance.npy"),
755  get_random_accessor(1.f, 1.f),
756  get_weights_accessor(data_path, unit_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_beta.npy"),
757  0.0010000000474974513f)
758  .set_name(unit_name + "Branch_1/Conv2d_0c_3x1/BatchNorm")
759  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Branch_1/Conv2d_0c_3x1/Relu");
760 
761  // Concatenate
762  i_l << ConcatLayer(std::move(i_la), std::move(i_lb)).set_name(unit_name + "concat")
763  << ConvolutionLayer(1U, 1U, 2080U,
764  get_weights_accessor(data_path, unit_path + "Conv2d_1x1_weights.npy", weights_layout),
765  get_weights_accessor(data_path, unit_path + "Conv2d_1x1_biases.npy", weights_layout),
766  PadStrideInfo(1, 1, 0, 0))
767  .set_name(unit_name + "Conv2d_1x1/convolution");
768 
769  // Scale result
770  if(scale != 1.f)
771  {
772  i_l << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, scale, 0.f)).set_name(unit_name + "mul");
773  }
774 
775  // Residual add
776  graph << EltwiseLayer(std::move(i_l), std::move(i_r), EltwiseOperation::Add).set_name(unit_name + "add");
777 
778  // Apply activation if needed
779  if(has_activation)
780  {
781  graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu");
782  }
783  }
784  }
785 };
786 
787 /** Main program for Inception ResNet V2
788  *
789  * Model is based on:
790  * https://arxiv.org/abs/1602.07261
791  * "Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning"
792  * Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, Alex Alemi
793  *
794  * Provenance: download.tensorflow.org/models/inception_resnet_v2_2016_08_30.tar.gz
795  *
796  * @note To list all the possible arguments execute the binary appended with the --help option
797  *
798  * @param[in] argc Number of arguments
799  * @param[in] argv Arguments
800  */
801 int main(int argc, char **argv)
802 {
803  return arm_compute::utils::run_example<InceptionResNetV2Example>(argc, argv);
804 }
Graph configuration structure Device target types.
Definition: Types.h:80
CLTunerMode tuner_mode
Tuner mode to be used by the CL tuner.
Definition: Types.h:87
std::unique_ptr< graph::ITensorAccessor > get_input_accessor(const arm_compute::utils::CommonGraphParams &graph_parameters, std::unique_ptr< IPreprocessor > preprocessor=nullptr, bool bgr=true)
Generates appropriate input accessor according to the specified graph parameters. ...
Definition: GraphUtils.h:497
void consume_common_graph_parameters(CommonGraphValidateOptions &options, CommonParams &common_params)
Consumes the consume_common_graph_parameters graph options and creates a structure containing any inf...
Includes all the Graph headers at once.
Common command line options used to configure the graph examples.
int main(int argc, char **argv)
Main program for Inception ResNet V2.
Class to parse command line arguments.
std::string mlgo_file
Filename to load MLGO heuristics from.
Definition: Types.h:90
std::unique_ptr< graph::ITensorAccessor > get_random_accessor(PixelValue lower, PixelValue upper, const std::random_device::result_type seed=0)
Generates appropriate random accessor.
Definition: GraphUtils.h:460
std::string tuner_file
File to load/store tuning values from.
Definition: Types.h:89
#define ARM_COMPUTE_EXIT_ON_MSG(cond, msg)
If the condition is true, the given message is printed and program exits.
Definition: Error.h:379
Abstract Example class.
Definition: Utils.h:78
Num samples, channels, height, width.
TensorShape permute_shape(TensorShape tensor_shape, DataLayout in_data_layout, DataLayout out_data_layout)
Permutes a given tensor shape given the input and output data layout.
Definition: GraphUtils.h:664
bool is_data_type_quantized_asymmetric(DataType dt)
Check if a given data type is of asymmetric quantized type.
Definition: Utils.h:1190
TensorDescriptor & set_layout(DataLayout data_layout)
Sets tensor descriptor data layout.
Structure holding all the common graph parameters.
std::unique_ptr< graph::ITensorAccessor > get_output_accessor(const arm_compute::utils::CommonGraphParams &graph_parameters, size_t top_n=5, bool is_validation=false, std::ostream &output_stream=std::cout)
Generates appropriate output accessor according to the specified graph parameters.
Definition: GraphUtils.h:543
bool use_tuner
Use a tuner in tunable backends.
Definition: Types.h:85
std::unique_ptr< graph::ITensorAccessor > get_weights_accessor(const std::string &path, const std::string &data_file, DataLayout file_layout=DataLayout::NCHW)
Generates appropriate weights accessor according to the specified path.
Definition: GraphUtils.h:475
int num_threads
Number of threads to use (thread capable backends), if 0 the backend will auto-initialize, if -1 the backend will stay as it is.
Definition: Types.h:88
Stream frontend class to construct simple graphs in a stream fashion.
Definition: Stream.h:45
DataLayout
[DataLayout enum definition]
Definition: Types.h:120
ILayer & set_name(std::string name)
Sets the name of the layer.
Definition: ILayer.h:55