Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / c13021e335b3e395c9d1a3a9935baedb42aebf08 / . / examples / graph_inception_v4.cpp
blob: 8f340354cdd2ab0a3bb51e82f9e6eed89261c260 [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 */
Georgios Pinitasd9eb2752018-04-03 13:44:29 +0100[diff] [blame]24#include "arm_compute/graph.h"
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]25#include "support/ToolchainSupport.h"
26#include "utils/GraphUtils.h"
27#include "utils/Utils.h"
28
29#include <cstdlib>
30#include <tuple>
31
32using namespace arm_compute::utils;
Georgios Pinitasd9eb2752018-04-03 13:44:29 +0100[diff] [blame]33using namespace arm_compute::graph::frontend;
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]34using namespace arm_compute::graph_utils;
35
36/** Example demonstrating how to implement InceptionV4's network using the Compute Library's graph API
37 *
38 * @param[in] argc Number of arguments
39 * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels )
40 */
41class InceptionV4Example final : public Example
42{
43public:
44 void do_setup(int argc, char **argv) override
45 {
Pablo Telloeb82fd22018-02-23 13:43:50 +0000[diff] [blame]46 // Disabled the test for now because the process gets killed on Linux Firefly 32 bit even when using ConvolutionMethodHint::DIRECT.
47 // Needs to review/rework to run the code below.
48#if __aarch64__
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]49 std::string data_path; /* Path to the trainable data */
50 std::string image; /* Image data */
51 std::string label; /* Label data */
52
53 // Create a preprocessor object
54 std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<TFPreproccessor>();
55
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]56 // Set target. 0 (NEON), 1 (OpenCL). By default it is NEON
Gian Marco Iodicec13021e2018-05-09 14:11:34 +0100[diff] [blame^]57 const int target = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
58 Target target_hint = set_target_hint(target);
59 ConvolutionMethod convolution_hint = target_hint == Target::NEON ? ConvolutionMethod::GEMM : ConvolutionMethod::DEFAULT;
Georgios Pinitas28705162018-03-21 20:10:53 +0000[diff] [blame]60
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]61 // Parse arguments
62 if(argc < 2)
63 {
64 // Print help
65 std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
66 std::cout << "No data folder provided: using random values\n\n";
67 }
68 else if(argc == 2)
69 {
70 std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
71 std::cout << "No data folder provided: using random values\n\n";
72 }
73 else if(argc == 3)
74 {
75 data_path = argv[2];
76 std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
77 std::cout << "No image provided: using random values\n\n";
78 }
79 else if(argc == 4)
80 {
81 data_path = argv[2];
82 image = argv[3];
83 std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
84 std::cout << "No text file with labels provided: skipping output accessor\n\n";
85 }
86 else
87 {
88 data_path = argv[2];
89 image = argv[3];
90 label = argv[4];
91 }
92
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]93 graph << target_hint << InputLayer(TensorDescriptor(TensorShape(299U, 299U, 3U, 1U), DataType::F32),
94 get_input_accessor(image, std::move(preprocessor), false))
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]95 // Conv2d_1a_3x3
96 << ConvolutionLayer(3U, 3U, 32U,
97 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_weights.npy"),
98 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
99 << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
100 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
101 get_random_accessor(1.f, 1.f),
102 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_1a_3x3_BatchNorm_beta.npy"),
103 0.001f)
104 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
Georgios Pinitas28705162018-03-21 20:10:53 +0000[diff] [blame]105 << convolution_hint
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]106 // Conv2d_2a_3x3
107 << ConvolutionLayer(3U, 3U, 32U,
108 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_weights.npy"),
109 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
110 << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_moving_mean.npy"),
111 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_moving_variance.npy"),
112 get_random_accessor(1.f, 1.f),
113 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2a_3x3_BatchNorm_beta.npy"),
114 0.001f)
115 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
116 // Conv2d_2b_3x3
117 << ConvolutionLayer(3U, 3U, 64U,
118 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_weights.npy"),
119 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
120 << BatchNormalizationLayer(get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_moving_mean.npy"),
121 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_moving_variance.npy"),
122 get_random_accessor(1.f, 1.f),
123 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Conv2d_2b_3x3_BatchNorm_beta.npy"),
124 0.001f)
Georgios Pinitas41c482d2018-04-17 13:23:26 +0100[diff] [blame]125 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]126
Georgios Pinitas41c482d2018-04-17 13:23:26 +0100[diff] [blame]127 graph << get_mixed_3a(data_path);
128 graph << get_mixed_4a(data_path);
129 graph << get_mixed_5a(data_path);
130 // 4 inception A blocks
131 graph << get_inceptionA_block(data_path, "Mixed_5b");
132 graph << get_inceptionA_block(data_path, "Mixed_5c");
133 graph << get_inceptionA_block(data_path, "Mixed_5d");
134 graph << get_inceptionA_block(data_path, "Mixed_5e");
135 // reduction A block
136 graph << get_reductionA_block(data_path);
137 // 7 inception B blocks
138 graph << get_inceptionB_block(data_path, "Mixed_6b");
139 graph << get_inceptionB_block(data_path, "Mixed_6c");
140 graph << get_inceptionB_block(data_path, "Mixed_6d");
141 graph << get_inceptionB_block(data_path, "Mixed_6e");
142 graph << get_inceptionB_block(data_path, "Mixed_6f");
143 graph << get_inceptionB_block(data_path, "Mixed_6g");
144 graph << get_inceptionB_block(data_path, "Mixed_6h");
145 // reduction B block
146 graph << get_reductionB_block(data_path);
147 // 3 inception C blocks
148 graph << get_inceptionC_block(data_path, "Mixed_7b");
149 graph << get_inceptionC_block(data_path, "Mixed_7c");
150 graph << get_inceptionC_block(data_path, "Mixed_7d");
151 graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]152 << FlattenLayer()
153 << FullyConnectedLayer(
154 1001U,
155 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_weights.npy"),
156 get_weights_accessor(data_path, "/cnn_data/inceptionv4_model/Logits_Logits_biases.npy"))
157 << SoftmaxLayer()
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]158 << OutputLayer(get_output_accessor(label, 5));
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]159
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]160 // Finalize graph
Georgios Pinitas9a8c6722018-03-21 17:52:35 +0000[diff] [blame]161 GraphConfig config;
162 config.use_function_memory_manager = true;
163 config.use_tuner = (target == 2);
164 graph.finalize(target_hint, config);
Pablo Telloeb82fd22018-02-23 13:43:50 +0000[diff] [blame]165#else /* __aarch64__ */
166 using namespace arm_compute;
167 ARM_COMPUTE_UNUSED(argc);
168 ARM_COMPUTE_UNUSED(argv);
169#endif /* __aarch64__ */
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]170 }
171
172 void do_run() override
173 {
Pablo Telloeb82fd22018-02-23 13:43:50 +0000[diff] [blame]174#if __aarch64__
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]175 graph.run();
Pablo Telloeb82fd22018-02-23 13:43:50 +0000[diff] [blame]176#endif /* __aarch64__ */
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]177 }
178
179private:
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]180 Stream graph{ 0, "InceptionV4" };
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]181
182private:
183 BranchLayer get_mixed_3a(const std::string &data_path)
184 {
185 std::string total_path = "/cnn_data/inceptionv4_model/Mixed_3a_";
186
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]187 SubStream i_a(graph);
188 i_a << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true));
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]189
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]190 SubStream i_b(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]191 i_b << ConvolutionLayer(3U, 3U, 96U,
192 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_weights.npy"),
193 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
194 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_moving_mean.npy"),
195 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_moving_variance.npy"),
196 get_random_accessor(1.f, 1.f),
197 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_3x3_BatchNorm_beta.npy"),
198 0.001f)
199 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
200
201 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
202 }
203
204 BranchLayer get_mixed_4a(const std::string &data_path)
205 {
206 std::string total_path = "/cnn_data/inceptionv4_model/Mixed_4a_";
207
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]208 SubStream i_a(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]209 i_a << ConvolutionLayer(1U, 1U, 64U,
210 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
211 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
212 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
213 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
214 get_random_accessor(1.f, 1.f),
215 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
216 0.001f)
217 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
218 << ConvolutionLayer(3U, 3U, 96U,
219 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"),
220 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
221 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
222 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
223 get_random_accessor(1.f, 1.f),
224 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"),
225 0.001f)
226 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
227
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]228 SubStream i_b(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]229 i_b << ConvolutionLayer(1U, 1U, 64U,
230 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
231 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
232 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
233 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
234 get_random_accessor(1.f, 1.f),
235 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
236 0.001f)
237 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
238 << ConvolutionLayer(7U, 1U, 64U,
239 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"),
240 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0))
241 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"),
242 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"),
243 get_random_accessor(1.f, 1.f),
244 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"),
245 0.001f)
246 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
247 << ConvolutionLayer(1U, 7U, 64U,
248 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"),
249 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3))
250 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"),
251 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"),
252 get_random_accessor(1.f, 1.f),
253 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"),
254 0.001f)
255 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
256 << ConvolutionLayer(3U, 3U, 96U,
257 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy"),
258 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
259 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
260 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
261 get_random_accessor(1.f, 1.f),
262 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"),
263 0.001f)
264 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
265
266 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
267 }
268
269 BranchLayer get_mixed_5a(const std::string &data_path)
270 {
271 std::string total_path = "/cnn_data/inceptionv4_model/Mixed_5a_";
272
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]273 SubStream i_a(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]274 i_a << ConvolutionLayer(3U, 3U, 192U,
275 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"),
276 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
277 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
278 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
279 get_random_accessor(1.f, 1.f),
280 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"),
281 0.001f)
282 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
283
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]284 SubStream i_b(graph);
285 i_b << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true));
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]286
287 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
288 }
289
290 BranchLayer get_inceptionA_block(const std::string &data_path, std::string &&param_path)
291 {
292 std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_";
293
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]294 SubStream i_a(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]295 i_a << ConvolutionLayer(1U, 1U, 96U,
296 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
297 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
298 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
299 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
300 get_random_accessor(1.f, 1.f),
301 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
302 0.001f)
303 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
304
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]305 SubStream i_b(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]306 i_b << ConvolutionLayer(1U, 1U, 64U,
307 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
308 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
309 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
310 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
311 get_random_accessor(1.f, 1.f),
312 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
313 0.001f)
314 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
315 << ConvolutionLayer(3U, 3U, 96U,
316 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy"),
317 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
318 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
319 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
320 get_random_accessor(1.f, 1.f),
321 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"),
322 0.001f)
323 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
324
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]325 SubStream i_c(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]326 i_c << ConvolutionLayer(1U, 1U, 64U,
327 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"),
328 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
329 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
330 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
331 get_random_accessor(1.f, 1.f),
332 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"),
333 0.001f)
334 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
335 << ConvolutionLayer(3U, 3U, 96U,
336 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_weights.npy"),
337 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
338 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
339 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
340 get_random_accessor(1.f, 1.f),
341 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x3_BatchNorm_beta.npy"),
342 0.001f)
343 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
344 << ConvolutionLayer(3U, 3U, 96U,
345 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_weights.npy"),
346 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
347 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_mean.npy"),
348 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_moving_variance.npy"),
349 get_random_accessor(1.f, 1.f),
350 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_3x3_BatchNorm_beta.npy"),
351 0.001f)
352 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
353
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]354 SubStream i_d(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]355 i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true))
356 << ConvolutionLayer(1U, 1U, 96U,
357 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"),
358 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
359 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"),
360 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"),
361 get_random_accessor(1.f, 1.f),
362 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"),
363 0.001f)
364 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
365
366 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d));
367 }
368
369 BranchLayer get_reductionA_block(const std::string &data_path)
370 {
371 std::string total_path = "/cnn_data/inceptionv4_model/Mixed_6a_";
372
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]373 SubStream i_a(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]374 i_a << ConvolutionLayer(3U, 3U, 384U,
375 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"),
376 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
377 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
378 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
379 get_random_accessor(1.f, 1.f),
380 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"),
381 0.001f)
382 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
383
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]384 SubStream i_b(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]385 i_b << ConvolutionLayer(1U, 1U, 192U,
386 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
387 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
388 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
389 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
390 get_random_accessor(1.f, 1.f),
391 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
392 0.001f)
393 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
394 << ConvolutionLayer(3U, 3U, 224U,
395 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_weights.npy"),
396 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 1, 1))
397 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_mean.npy"),
398 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_moving_variance.npy"),
399 get_random_accessor(1.f, 1.f),
400 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_3x3_BatchNorm_beta.npy"),
401 0.001f)
402 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
403 << ConvolutionLayer(3U, 3U, 256U,
404 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy"),
405 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
406 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
407 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
408 get_random_accessor(1.f, 1.f),
409 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"),
410 0.001f)
411 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
412
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]413 SubStream i_c(graph);
414 i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true));
415
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]416 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c));
417 }
418
419 BranchLayer get_inceptionB_block(const std::string &data_path, std::string &&param_path)
420 {
421 std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_";
422
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]423 SubStream i_a(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]424 i_a << ConvolutionLayer(1U, 1U, 384U,
425 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
426 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
427 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
428 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
429 get_random_accessor(1.f, 1.f),
430 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
431 0.001f)
432 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
433
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]434 SubStream i_b(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]435 i_b << ConvolutionLayer(1U, 1U, 192U,
436 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
437 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
438 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
439 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
440 get_random_accessor(1.f, 1.f),
441 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
442 0.001f)
443 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
444 << ConvolutionLayer(7U, 1U, 224U,
445 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"),
446 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0))
447 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"),
448 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"),
449 get_random_accessor(1.f, 1.f),
450 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"),
451 0.001f)
452 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
453 << ConvolutionLayer(1U, 7U, 256U,
454 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"),
455 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3))
456 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"),
457 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"),
458 get_random_accessor(1.f, 1.f),
459 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"),
460 0.001f)
461 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
462
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]463 SubStream i_c(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]464 i_c << ConvolutionLayer(1U, 1U, 192U,
465 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"),
466 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
467 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
468 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
469 get_random_accessor(1.f, 1.f),
470 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"),
471 0.001f)
472 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
473 << ConvolutionLayer(1U, 7U, 192U,
474 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_weights.npy"),
475 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3))
476 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_mean.npy"),
477 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_moving_variance.npy"),
478 get_random_accessor(1.f, 1.f),
479 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_7x1_BatchNorm_beta.npy"),
480 0.001f)
481 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
482 << ConvolutionLayer(7U, 1U, 224U,
483 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_weights.npy"),
484 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0))
485 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_mean.npy"),
486 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_moving_variance.npy"),
487 get_random_accessor(1.f, 1.f),
488 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x7_BatchNorm_beta.npy"),
489 0.001f)
490 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
491 << ConvolutionLayer(1U, 7U, 224U,
492 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_weights.npy"),
493 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3))
494 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_mean.npy"),
495 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_moving_variance.npy"),
496 get_random_accessor(1.f, 1.f),
497 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_7x1_BatchNorm_beta.npy"),
498 0.001f)
499 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
500 << ConvolutionLayer(7U, 1U, 256U,
501 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_weights.npy"),
502 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0))
503 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_mean.npy"),
504 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_moving_variance.npy"),
505 get_random_accessor(1.f, 1.f),
506 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_1x7_BatchNorm_beta.npy"),
507 0.001f)
508 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
509
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]510 SubStream i_d(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]511 i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true))
512 << ConvolutionLayer(1U, 1U, 128U,
513 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"),
514 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
515 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"),
516 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"),
517 get_random_accessor(1.f, 1.f),
518 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"),
519 0.001f)
520 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
521
522 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d));
523 }
524
525 BranchLayer get_reductionB_block(const std::string &data_path)
526 {
527 std::string total_path = "/cnn_data/inceptionv4_model/Mixed_7a_";
528
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]529 SubStream i_a(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]530 i_a << ConvolutionLayer(1U, 1U, 192U,
531 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
532 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
533 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
534 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
535 get_random_accessor(1.f, 1.f),
536 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
537 0.001f)
538 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
539 << ConvolutionLayer(3U, 3U, 192U,
540 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_weights.npy"),
541 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
542 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
543 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
544 get_random_accessor(1.f, 1.f),
545 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_1a_3x3_BatchNorm_beta.npy"),
546 0.001f)
547 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
548
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]549 SubStream i_b(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]550 i_b << ConvolutionLayer(1U, 1U, 256U,
551 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
552 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
553 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
554 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
555 get_random_accessor(1.f, 1.f),
556 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
557 0.001f)
558 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
559 << ConvolutionLayer(7U, 1U, 256U,
560 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_weights.npy"),
561 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 3, 0))
562 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_mean.npy"),
563 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_moving_variance.npy"),
564 get_random_accessor(1.f, 1.f),
565 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x7_BatchNorm_beta.npy"),
566 0.001f)
567 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
568 << ConvolutionLayer(1U, 7U, 320U,
569 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_weights.npy"),
570 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 3))
571 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_mean.npy"),
572 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_moving_variance.npy"),
573 get_random_accessor(1.f, 1.f),
574 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_7x1_BatchNorm_beta.npy"),
575 0.001f)
576 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
577 << ConvolutionLayer(3U, 3U, 320U,
578 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_weights.npy"),
579 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(2, 2, 0, 0))
580 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_mean.npy"),
581 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_moving_variance.npy"),
582 get_random_accessor(1.f, 1.f),
583 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_1a_3x3_BatchNorm_beta.npy"),
584 0.001f)
585 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
586
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]587 SubStream i_c(graph);
588 i_c << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), true));
589
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]590 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c));
591 }
592
593 BranchLayer get_inceptionC_block(const std::string &data_path, std::string &&param_path)
594 {
595 std::string total_path = "/cnn_data/inceptionv4_model/" + param_path + "_";
596
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]597 SubStream i_a(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]598 i_a << ConvolutionLayer(1U, 1U, 256U,
599 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_weights.npy"),
600 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
601 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
602 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
603 get_random_accessor(1.f, 1.f),
604 get_weights_accessor(data_path, total_path + "Branch_0_Conv2d_0a_1x1_BatchNorm_beta.npy"),
605 0.001f)
606 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
607
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]608 SubStream i_b(graph);
609 i_b << ConvolutionLayer(
610 1U, 1U, 384U,
611 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_weights.npy"),
612 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
613 PadStrideInfo(1, 1, 0, 0))
614 << BatchNormalizationLayer(
615 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
616 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
617 get_random_accessor(1.f, 1.f),
618 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0a_1x1_BatchNorm_beta.npy"),
619 0.001f)
620 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
621
622 SubStream i_b1(static_cast<IStream &>(i_b));
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]623 i_b1 << ConvolutionLayer(
624 3U, 1U, 256U,
625 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_weights.npy"),
626 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
627 PadStrideInfo(1, 1, 1, 0))
628 << BatchNormalizationLayer(
629 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_mean.npy"),
630 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_moving_variance.npy"),
631 get_random_accessor(1.f, 1.f),
632 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0b_1x3_BatchNorm_beta.npy"),
633 0.001f)
634 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
635
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]636 SubStream i_b2(static_cast<IStream &>(i_b));
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]637 i_b2 << ConvolutionLayer(
638 1U, 3U, 256U,
639 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_weights.npy"),
640 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
641 PadStrideInfo(1, 1, 0, 1))
642 << BatchNormalizationLayer(
643 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_mean.npy"),
644 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_moving_variance.npy"),
645 get_random_accessor(1.f, 1.f),
646 get_weights_accessor(data_path, total_path + "Branch_1_Conv2d_0c_3x1_BatchNorm_beta.npy"),
647 0.001f)
648 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
649
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]650 // Merge b1 and b2
651 i_b << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_b1), std::move(i_b2));
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]652
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]653 SubStream i_c(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]654 i_c << ConvolutionLayer(
655 1U, 1U, 384U,
656 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_weights.npy"),
657 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
658 PadStrideInfo(1, 1, 0, 0))
659 << BatchNormalizationLayer(
660 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_mean.npy"),
661 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_moving_variance.npy"),
662 get_random_accessor(1.f, 1.f),
663 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0a_1x1_BatchNorm_beta.npy"),
664 0.001f)
665 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
666 << ConvolutionLayer(
667 1U, 3U, 448U,
668 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_weights.npy"),
669 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
670 PadStrideInfo(1, 1, 0, 1))
671 << BatchNormalizationLayer(
672 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_BatchNorm_moving_mean.npy"),
673 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_BatchNorm_moving_variance.npy"),
674 get_random_accessor(1.f, 1.f),
675 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0b_3x1_BatchNorm_beta.npy"),
676 0.001f)
677 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
678 << ConvolutionLayer(
679 3U, 1U, 512U,
680 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_weights.npy"),
681 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
682 PadStrideInfo(1, 1, 1, 0))
683 << BatchNormalizationLayer(
684 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_mean.npy"),
685 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_moving_variance.npy"),
686 get_random_accessor(1.f, 1.f),
687 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0c_1x3_BatchNorm_beta.npy"),
688 0.001f)
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]689 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]690
Georgios Pinitasd8734b52017-12-22 15:27:52 +0000[diff] [blame]691 SubStream i_c1(static_cast<IStream &>(i_c));
692 i_c1 << ConvolutionLayer(
693 3U, 1U, 256U,
694 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_weights.npy"),
695 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
696 PadStrideInfo(1, 1, 1, 0))
697 << BatchNormalizationLayer(
698 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_moving_mean.npy"),
699 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_moving_variance.npy"),
700 get_random_accessor(1.f, 1.f),
701 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0d_1x3_BatchNorm_beta.npy"),
702 0.001f)
703 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
704
705 SubStream i_c2(static_cast<IStream &>(i_c));
706 i_c2 << ConvolutionLayer(
707 1U, 3U, 256U,
708 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_weights.npy"),
709 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
710 PadStrideInfo(1, 1, 0, 1))
711 << BatchNormalizationLayer(
712 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_moving_mean.npy"),
713 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_moving_variance.npy"),
714 get_random_accessor(1.f, 1.f),
715 get_weights_accessor(data_path, total_path + "Branch_2_Conv2d_0e_3x1_BatchNorm_beta.npy"),
716 0.001f)
717 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
718
719 // Merge i_c1 and i_c2
720 i_c << BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_c1), std::move(i_c2));
721
722 SubStream i_d(graph);
Georgios Pinitas240cfa62018-02-26 19:58:04 +0000[diff] [blame]723 i_d << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL), true))
724 << ConvolutionLayer(1U, 1U, 256U,
725 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_weights.npy"),
726 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr), PadStrideInfo(1, 1, 0, 0))
727 << BatchNormalizationLayer(get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_mean.npy"),
728 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_moving_variance.npy"),
729 get_random_accessor(1.f, 1.f),
730 get_weights_accessor(data_path, total_path + "Branch_3_Conv2d_0b_1x1_BatchNorm_beta.npy"),
731 0.001f)
732 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
733
734 return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d));
735 }
736};
737
738/** Main program for Inception V4
739 *
740 * @param[in] argc Number of arguments
741 * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels )
742 */
743int main(int argc, char **argv)
744{
745 return arm_compute::utils::run_example<InceptionV4Example>(argc, argv);
746}