Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 240cfa65322da0c9b2ec0e78960ff4bd87c7573e / . / examples / graph_inception_v4.cpp
blob: f004b41fb09d2f212c32a3d114a87ba820520fa8 [file] [log] [blame]
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[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/Graph.h"
25#include "arm_compute/graph/Nodes.h"
26#include "arm_compute/graph/SubGraph.h"
27#include "support/ToolchainSupport.h"
28#include "utils/GraphUtils.h"
29#include "utils/Utils.h"
30
31#include <cstdlib>
32#include <tuple>
33
34using namespace arm_compute::utils;
35using namespace arm_compute::graph;
36using namespace arm_compute::graph_utils;
37
38/** Example demonstrating how to implement InceptionV4's network using the Compute Library's graph API
39 *
40 * @param[in] argc Number of arguments
41 * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels )
42 */
43class InceptionV4Example final : public Example
44{
45public:
46 void do_setup(int argc, char **argv) override
47 {
48 std::string data_path; /* Path to the trainable data */
49 std::string image; /* Image data */
50 std::string label; /* Label data */
51
52 // Create a preprocessor object
53 std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<TFPreproccessor>();
54
55 // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
56 const int int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
57 TargetHint target_hint = set_target_hint(int_target_hint);
58
59 // Parse arguments
60 if(argc < 2)
61 {
62 // Print help
63 std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
64 std::cout << "No data folder provided: using random values\n\n";
65 }
66 else if(argc == 2)
67 {
68 std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
69 std::cout << "No data folder provided: using random values\n\n";
70 }
71 else if(argc == 3)
72 {
73 data_path = argv[2];
74 std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
75 std::cout << "No image provided: using random values\n\n";
76 }
77 else if(argc == 4)
78 {
79 data_path = argv[2];
80 image = argv[3];
81 std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
82 std::cout << "No text file with labels provided: skipping output accessor\n\n";
83 }
84 else
85 {
86 data_path = argv[2];
87 image = argv[3];
88 label = argv[4];
89 }
90
91 graph << target_hint << Tensor(TensorInfo(TensorShape(299U, 299U, 3U, 1U), 1, DataType::F32),
92 get_input_accessor(image, std::move(preprocessor), false))
93
94 // Conv2d_1a_3x3
95 << ConvolutionLayer(3U, 3U, 32U,
96 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_weights.npy"),
97 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
98 << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
99 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
100 get_random_accessor(1.f, 1.f),
101 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_beta.npy"),
102 0.001f)
103 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
104 // Conv2d_2a_3x3
105 << ConvolutionLayer(3U, 3U, 32U,
106 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_weights.npy"),
107 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
108 << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_moving_mean.npy"),
109 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_moving_variance.npy"),
110 get_random_accessor(1.f, 1.f),
111 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_beta.npy"),
112 0.001f)
113 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
114 // Conv2d_2b_3x3
115 << ConvolutionLayer(3U, 3U, 64U,
116 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_weights.npy"),
117 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
118 << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_moving_mean.npy"),
119 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_moving_variance.npy"),
120 get_random_accessor(1.f, 1.f),
121 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_beta.npy"),
122 0.001f)
123 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
124
125 << get_mixed_3a(data_path)
126 << get_mixed_4a(data_path)
127 << get_mixed_5a(data_path)
128 // 4 inception A blocks
129 << get_inceptionA_block(data_path, "Mixed_5b")
130 << get_inceptionA_block(data_path, "Mixed_5c")
131 << get_inceptionA_block(data_path, "Mixed_5d")
132 << get_inceptionA_block(data_path, "Mixed_5e")
133 // reduction A block
134 << get_reductionA_block(data_path)
135 // 7 inception B blocks
136 << get_inceptionB_block(data_path, "Mixed_6b")
137 << get_inceptionB_block(data_path, "Mixed_6c")
138 << get_inceptionB_block(data_path, "Mixed_6d")
139 << get_inceptionB_block(data_path, "Mixed_6e")
140 << get_inceptionB_block(data_path, "Mixed_6f")
141 << get_inceptionB_block(data_path, "Mixed_6g")
142 << get_inceptionB_block(data_path, "Mixed_6h")
143 // reduction B block
144 << get_reductionB_block(data_path)
145 // 3 inception C blocks
146 << get_inceptionC_block(data_path, "Mixed_7b")
147 << get_inceptionC_block(data_path, "Mixed_7c")
148 << get_inceptionC_block(data_path, "Mixed_7d")
149 << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
150 << FlattenLayer()
151 << FullyConnectedLayer(
152 1001U,
153 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_weights.npy"),
154 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_biases.npy"))
155 << SoftmaxLayer()
156 << Tensor(get_output_accessor(label, 5));
157
158 // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
159 graph.graph_init(int_target_hint == 2);
160 }
161
162 void do_run() override
163 {
164 graph.run();
165 }
166
167private:
168 Graph graph{};
169
170private:
171 BranchLayer get_mixed_3a(const std::string &data_path)
172 {
173 std::string total_path = "/cnn_data/inceptionv4_model/Mixed_3a_";
174
175 SubGraph i_a;
176 i_a << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true))
177 // TODO (geopin01) : Remove once we understand why a single node graph does not run in CL
178 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f));
179
180 SubGraph i_b;
181 i_b << ConvolutionLayer(3U, 3U, 96U,
182 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_weights.npy"),
183 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
184 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_moving_mean.npy"),
185 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_moving_variance.npy"),
186 get_random_accessor(1.f, 1.f),
187 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_beta.npy"),
188 0.001f)
189 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
190
191 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
192 }
193
194 BranchLayer get_mixed_4a(const std::string &data_path)
195 {
196 std::string total_path = "/cnn_data/inceptionv4_model/Mixed_4a_";
197
198 SubGraph i_a;
199 i_a << ConvolutionLayer(1U, 1U, 64U,
200 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
201 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
202 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
203 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
204 get_random_accessor(1.f, 1.f),
205 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
206 0.001f)
207 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
208 << ConvolutionLayer(3U, 3U, 96U,
209 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"),
210 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
211 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
212 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
213 get_random_accessor(1.f, 1.f),
214 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"),
215 0.001f)
216 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
217
218 SubGraph i_b;
219 i_b << ConvolutionLayer(1U, 1U, 64U,
220 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
221 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
222 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
223 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
224 get_random_accessor(1.f, 1.f),
225 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
226 0.001f)
227 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
228 << ConvolutionLayer(7U, 1U, 64U,
229 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"),
230 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0))
231 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"),
232 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"),
233 get_random_accessor(1.f, 1.f),
234 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"),
235 0.001f)
236 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
237 << ConvolutionLayer(1U, 7U, 64U,
238 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"),
239 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3))
240 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"),
241 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"),
242 get_random_accessor(1.f, 1.f),
243 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"),
244 0.001f)
245 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
246 << ConvolutionLayer(3U, 3U, 96U,
247 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy"),
248 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
249 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
250 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
251 get_random_accessor(1.f, 1.f),
252 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"),
253 0.001f)
254 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
255
256 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
257 }
258
259 BranchLayer get_mixed_5a(const std::string &data_path)
260 {
261 std::string total_path = "/cnn_data/inceptionv4_model/Mixed_5a_";
262
263 SubGraph i_a;
264 i_a << ConvolutionLayer(3U, 3U, 192U,
265 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"),
266 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
267 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
268 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
269 get_random_accessor(1.f, 1.f),
270 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"),
271 0.001f)
272 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
273
274 SubGraph i_b;
275 i_b << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true))
276 // TODO (geopin01) : Remove once we understand why a single node graph does not run in CL
277 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f));
278
279 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
280 }
281
282 BranchLayer get_inceptionA_block(const std::string &data_path, std::string &&param_path)
283 {
284 std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_";
285
286 SubGraph i_a;
287 i_a << ConvolutionLayer(1U, 1U, 96U,
288 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
289 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
290 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
291 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
292 get_random_accessor(1.f, 1.f),
293 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
294 0.001f)
295 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
296
297 SubGraph i_b;
298 i_b << ConvolutionLayer(1U, 1U, 64U,
299 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
300 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
301 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
302 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
303 get_random_accessor(1.f, 1.f),
304 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
305 0.001f)
306 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
307 << ConvolutionLayer(3U, 3U, 96U,
308 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy"),
309 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
310 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
311 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
312 get_random_accessor(1.f, 1.f),
313 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"),
314 0.001f)
315 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
316
317 SubGraph i_c;
318 i_c << ConvolutionLayer(1U, 1U, 64U,
319 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"),
320 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
321 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
322 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
323 get_random_accessor(1.f, 1.f),
324 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"),
325 0.001f)
326 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
327 << ConvolutionLayer(3U, 3U, 96U,
328 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy"),
329 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
330 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
331 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
332 get_random_accessor(1.f, 1.f),
333 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"),
334 0.001f)
335 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
336 << ConvolutionLayer(3U, 3U, 96U,
337 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_weights.npy"),
338 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
339 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"),
340 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"),
341 get_random_accessor(1.f, 1.f),
342 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"),
343 0.001f)
344 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
345
346 SubGraph i_d;
347 i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true))
348 << ConvolutionLayer(1U, 1U, 96U,
349 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"),
350 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
351 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"),
352 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"),
353 get_random_accessor(1.f, 1.f),
354 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"),
355 0.001f)
356 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
357
358 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d));
359 }
360
361 BranchLayer get_reductionA_block(const std::string &data_path)
362 {
363 std::string total_path = "/cnn_data/inceptionv4_model/Mixed_6a_";
364
365 SubGraph i_a;
366 i_a << ConvolutionLayer(3U, 3U, 384U,
367 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"),
368 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
369 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
370 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
371 get_random_accessor(1.f, 1.f),
372 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"),
373 0.001f)
374 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
375
376 SubGraph i_b;
377 i_b << ConvolutionLayer(1U, 1U, 192U,
378 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
379 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
380 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
381 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
382 get_random_accessor(1.f, 1.f),
383 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
384 0.001f)
385 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
386 << ConvolutionLayer(3U, 3U, 224U,
387 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy"),
388 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
389 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
390 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
391 get_random_accessor(1.f, 1.f),
392 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"),
393 0.001f)
394 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
395 << ConvolutionLayer(3U, 3U, 256U,
396 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy"),
397 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
398 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
399 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
400 get_random_accessor(1.f, 1.f),
401 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"),
402 0.001f)
403 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
404
405 SubGraph i_c;
406 i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true))
407 // TODO (geopin01) : Remove once we understand why a single node graph does not run in CL
408 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f));
409 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c));
410 }
411
412 BranchLayer get_inceptionB_block(const std::string &data_path, std::string &&param_path)
413 {
414 std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_";
415
416 SubGraph i_a;
417 i_a << ConvolutionLayer(1U, 1U, 384U,
418 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
419 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
420 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
421 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
422 get_random_accessor(1.f, 1.f),
423 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
424 0.001f)
425 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
426
427 SubGraph i_b;
428 i_b << ConvolutionLayer(1U, 1U, 192U,
429 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
430 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
431 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
432 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
433 get_random_accessor(1.f, 1.f),
434 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
435 0.001f)
436 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
437 << ConvolutionLayer(7U, 1U, 224U,
438 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"),
439 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0))
440 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"),
441 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"),
442 get_random_accessor(1.f, 1.f),
443 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"),
444 0.001f)
445 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
446 << ConvolutionLayer(1U, 7U, 256U,
447 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"),
448 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3))
449 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"),
450 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"),
451 get_random_accessor(1.f, 1.f),
452 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"),
453 0.001f)
454 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
455
456 SubGraph i_c;
457 i_c << ConvolutionLayer(1U, 1U, 192U,
458 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"),
459 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
460 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
461 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
462 get_random_accessor(1.f, 1.f),
463 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"),
464 0.001f)
465 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
466 << ConvolutionLayer(1U, 7U, 192U,
467 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_weights.npy"),
468 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3))
469 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_mean.npy"),
470 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_variance.npy"),
471 get_random_accessor(1.f, 1.f),
472 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_beta.npy"),
473 0.001f)
474 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
475 << ConvolutionLayer(7U, 1U, 224U,
476 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_weights.npy"),
477 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0))
478 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_mean.npy"),
479 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_variance.npy"),
480 get_random_accessor(1.f, 1.f),
481 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_beta.npy"),
482 0.001f)
483 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
484 << ConvolutionLayer(1U, 7U, 224U,
485 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_weights.npy"),
486 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3))
487 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_mean.npy"),
488 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_variance.npy"),
489 get_random_accessor(1.f, 1.f),
490 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_beta.npy"),
491 0.001f)
492 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
493 << ConvolutionLayer(7U, 1U, 256U,
494 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_weights.npy"),
495 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0))
496 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_mean.npy"),
497 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_variance.npy"),
498 get_random_accessor(1.f, 1.f),
499 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_beta.npy"),
500 0.001f)
501 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
502
503 SubGraph i_d;
504 i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true))
505 << ConvolutionLayer(1U, 1U, 128U,
506 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"),
507 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
508 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"),
509 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"),
510 get_random_accessor(1.f, 1.f),
511 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"),
512 0.001f)
513 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
514
515 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d));
516 }
517
518 BranchLayer get_reductionB_block(const std::string &data_path)
519 {
520 std::string total_path = "/cnn_data/inceptionv4_model/Mixed_7a_";
521
522 SubGraph i_a;
523 i_a << ConvolutionLayer(1U, 1U, 192U,
524 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
525 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
526 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
527 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
528 get_random_accessor(1.f, 1.f),
529 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
530 0.001f)
531 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
532 << ConvolutionLayer(3U, 3U, 192U,
533 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"),
534 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
535 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
536 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
537 get_random_accessor(1.f, 1.f),
538 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"),
539 0.001f)
540 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
541
542 SubGraph i_b;
543 i_b << ConvolutionLayer(1U, 1U, 256U,
544 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
545 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
546 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
547 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
548 get_random_accessor(1.f, 1.f),
549 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
550 0.001f)
551 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
552 << ConvolutionLayer(7U, 1U, 256U,
553 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"),
554 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0))
555 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"),
556 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"),
557 get_random_accessor(1.f, 1.f),
558 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"),
559 0.001f)
560 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
561 << ConvolutionLayer(1U, 7U, 320U,
562 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"),
563 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3))
564 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"),
565 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"),
566 get_random_accessor(1.f, 1.f),
567 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"),
568 0.001f)
569 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
570 << ConvolutionLayer(3U, 3U, 320U,
571 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy"),
572 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
573 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
574 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
575 get_random_accessor(1.f, 1.f),
576 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"),
577 0.001f)
578 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
579
580 SubGraph i_c;
581 i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true))
582 // TODO (geopin01) : Remove once we understand why a single node graph does not run in CL
583 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f, 0.f));
584 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c));
585 }
586
587 BranchLayer get_inceptionC_block(const std::string &data_path, std::string &&param_path)
588 {
589 std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_";
590
591 SubGraph i_a;
592 i_a << ConvolutionLayer(1U, 1U, 256U,
593 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
594 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
595 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
596 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
597 get_random_accessor(1.f, 1.f),
598 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
599 0.001f)
600 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
601
602 SubGraph i_b1;
603 i_b1 << ConvolutionLayer(
604 3U, 1U, 256U,
605 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_weights.npy"),
606 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
607 PadStrideInfo(1, 1, 1, 0))
608 << BatchNormalizationLayer(
609 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"),
610 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"),
611 get_random_accessor(1.f, 1.f),
612 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"),
613 0.001f)
614 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
615
616 SubGraph i_b2;
617 i_b2 << ConvolutionLayer(
618 1U, 3U, 256U,
619 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_weights.npy"),
620 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
621 PadStrideInfo(1, 1, 0, 1))
622 << BatchNormalizationLayer(
623 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_mean.npy"),
624 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_variance.npy"),
625 get_random_accessor(1.f, 1.f),
626 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_beta.npy"),
627 0.001f)
628 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
629
630 SubGraph i_b;
631 i_b << ConvolutionLayer(
632 1U, 1U, 384U,
633 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
634 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
635 PadStrideInfo(1, 1, 0, 0))
636 << BatchNormalizationLayer(
637 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
638 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
639 get_random_accessor(1.f, 1.f),
640 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
641 0.001f)
642 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
643 << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_b1), std::move(i_b2));
644
645 SubGraph i_c1;
646 i_c1 << ConvolutionLayer(
647 3U, 1U, 256U,
648 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_weights.npy"),
649 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
650 PadStrideInfo(1, 1, 1, 0))
651 << BatchNormalizationLayer(
652 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_moving_mean.npy"),
653 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_moving_variance.npy"),
654 get_random_accessor(1.f, 1.f),
655 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_beta.npy"),
656 0.001f)
657 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
658
659 SubGraph i_c2;
660 i_c2 << ConvolutionLayer(
661 1U, 3U, 256U,
662 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_weights.npy"),
663 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
664 PadStrideInfo(1, 1, 0, 1))
665 << BatchNormalizationLayer(
666 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_moving_mean.npy"),
667 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_moving_variance.npy"),
668 get_random_accessor(1.f, 1.f),
669 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_beta.npy"),
670 0.001f)
671 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
672
673 SubGraph i_c;
674 i_c << ConvolutionLayer(
675 1U, 1U, 384U,
676 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"),
677 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
678 PadStrideInfo(1, 1, 0, 0))
679 << BatchNormalizationLayer(
680 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
681 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
682 get_random_accessor(1.f, 1.f),
683 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"),
684 0.001f)
685 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
686 << ConvolutionLayer(
687 1U, 3U, 448U,
688 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_weights.npy"),
689 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
690 PadStrideInfo(1, 1, 0, 1))
691 << BatchNormalizationLayer(
692 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_BatchNorm_moving_mean.npy"),
693 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_BatchNorm_moving_variance.npy"),
694 get_random_accessor(1.f, 1.f),
695 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_BatchNorm_beta.npy"),
696 0.001f)
697 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
698 << ConvolutionLayer(
699 3U, 1U, 512U,
700 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_weights.npy"),
701 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
702 PadStrideInfo(1, 1, 1, 0))
703 << BatchNormalizationLayer(
704 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_mean.npy"),
705 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_variance.npy"),
706 get_random_accessor(1.f, 1.f),
707 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_beta.npy"),
708 0.001f)
709 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
710 << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_c1), std::move(i_c2));
711
712 SubGraph i_d;
713 i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true))
714 << ConvolutionLayer(1U, 1U, 256U,
715 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"),
716 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
717 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"),
718 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"),
719 get_random_accessor(1.f, 1.f),
720 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"),
721 0.001f)
722 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
723
724 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d));
725 }
726};
727
728/** Main program for Inception V4
729 *
730 * @param[in] argc Number of arguments
731 * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels )
732 */
733int main(int argc, char **argv)
734{
735 return arm_compute::utils::run_example<InceptionV4Example>(argc, argv);
736}