Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 050471e40fc58cb5ea745701a43ec5b2b9586b81 / . / examples / graph_yolov3.cpp
blob: c0a97da0605670ac1e6cf6467fd886256ac2102f [file] [log] [blame]
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]1/*
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]2 * Copyright (c) 2018-2019 ARM Limited.
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24#include "arm_compute/graph.h"
25#include "support/ToolchainSupport.h"
26#include "utils/CommonGraphOptions.h"
27#include "utils/GraphUtils.h"
28#include "utils/Utils.h"
29
30using namespace arm_compute::utils;
31using namespace arm_compute::graph::frontend;
32using namespace arm_compute::graph_utils;
33
Georgios Pinitas108ab0b2018-09-14 18:35:11 +0100[diff] [blame]34/** Example demonstrating how to implement YOLOv3 network using the Compute Library's graph API */
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]35class GraphYOLOv3Example : public Example
36{
37public:
38 GraphYOLOv3Example()
39 : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "YOLOv3")
40 {
41 }
Michalis Spyroue22aa132018-09-13 10:35:33 +0100[diff] [blame]42
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]43 bool do_setup(int argc, char **argv) override
44 {
45 // Parse arguments
46 cmd_parser.parse(argc, argv);
47
48 // Consume common parameters
49 common_params = consume_common_graph_parameters(common_opts);
50
51 // Return when help menu is requested
52 if(common_params.help)
53 {
54 cmd_parser.print_help(argv[0]);
55 return false;
56 }
57
58 // Checks
59 ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph");
60
61 // Print parameter values
62 std::cout << common_params << std::endl;
63
64 // Get trainable parameters data path
65 std::string data_path = common_params.data_path;
66
67 // Create a preprocessor object
68 std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<TFPreproccessor>(0.f);
69
70 // Create input descriptor
71 const TensorShape tensor_shape = permute_shape(TensorShape(608U, 608U, 3U, 1U), DataLayout::NCHW, common_params.data_layout);
72 TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout);
73
74 // Set weights trained layout
75 const DataLayout weights_layout = DataLayout::NCHW;
76
77 graph << common_params.target
78 << common_params.fast_math_hint
Michalis Spyroue22aa132018-09-13 10:35:33 +0100[diff] [blame]79 << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false));
80 std::pair<SubStream, SubStream> intermediate_layers = darknet53(data_path, weights_layout);
81 graph << ConvolutionLayer(
82 1U, 1U, 512U,
83 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_53_w.npy", weights_layout),
84 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
85 PadStrideInfo(1, 1, 0, 0))
86 .set_name("conv2d_53")
87 << BatchNormalizationLayer(
88 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_mean.npy"),
89 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_var.npy"),
90 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_gamma.npy"),
91 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_beta.npy"),
92 0.000001f)
93 .set_name("conv2d_53/BatchNorm")
94 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_53/LeakyRelu")
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]95 << ConvolutionLayer(
Michalis Spyroue22aa132018-09-13 10:35:33 +0100[diff] [blame]96 3U, 3U, 1024U,
97 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_54_w.npy", weights_layout),
98 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
99 PadStrideInfo(1, 1, 1, 1))
100 .set_name("conv2d_54")
101 << BatchNormalizationLayer(
102 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_mean.npy"),
103 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_var.npy"),
104 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_gamma.npy"),
105 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_beta.npy"),
106 0.000001f)
107 .set_name("conv2d_54/BatchNorm")
108 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_54/LeakyRelu")
109 << ConvolutionLayer(
110 1U, 1U, 512U,
111 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_55_w.npy", weights_layout),
112 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
113 PadStrideInfo(1, 1, 0, 0))
114 .set_name("conv2d_55")
115 << BatchNormalizationLayer(
116 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_mean.npy"),
117 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_var.npy"),
118 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_gamma.npy"),
119 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_beta.npy"),
120 0.000001f)
121 .set_name("conv2d_55/BatchNorm")
122 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_55/LeakyRelu")
123 << ConvolutionLayer(
124 3U, 3U, 1024U,
125 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_56_w.npy", weights_layout),
126 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
127 PadStrideInfo(1, 1, 1, 1))
128 .set_name("conv2d_56")
129 << BatchNormalizationLayer(
130 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_mean.npy"),
131 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_var.npy"),
132 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_gamma.npy"),
133 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_beta.npy"),
134 0.000001f)
135 .set_name("conv2d_56/BatchNorm")
136 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_56/LeakyRelu")
137 << ConvolutionLayer(
138 1U, 1U, 512U,
139 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_57_w.npy", weights_layout),
140 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
141 PadStrideInfo(1, 1, 0, 0))
142 .set_name("conv2d_57")
143 << BatchNormalizationLayer(
144 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_mean.npy"),
145 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_var.npy"),
146 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_gamma.npy"),
147 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_beta.npy"),
148 0.000001f)
149 .set_name("conv2d_57/BatchNorm")
150 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_57/LeakyRelu");
151 SubStream route_1(graph);
152 graph << ConvolutionLayer(
153 3U, 3U, 1024U,
154 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_58_w.npy", weights_layout),
155 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
156 PadStrideInfo(1, 1, 1, 1))
157 .set_name("conv2d_58")
158 << BatchNormalizationLayer(
159 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_mean.npy"),
160 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_var.npy"),
161 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_gamma.npy"),
162 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_beta.npy"),
163 0.000001f)
164 .set_name("conv2d_58/BatchNorm")
165 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_58/LeakyRelu")
166 << ConvolutionLayer(
167 1U, 1U, 255U,
168 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_59_w.npy", weights_layout),
169 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_59_b.npy", weights_layout),
170 PadStrideInfo(1, 1, 0, 0))
171 .set_name("conv2d_59")
172 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f)).set_name("conv2d_59/Linear")
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]173 << YOLOLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.1f), 80).set_name("Yolo1")
Michalis Spyroue22aa132018-09-13 10:35:33 +0100[diff] [blame]174 << OutputLayer(get_output_accessor(common_params, 5));
175 route_1 << ConvolutionLayer(
176 1U, 1U, 256U,
177 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_60_w.npy", weights_layout),
178 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
179 PadStrideInfo(1, 1, 0, 0))
180 .set_name("conv2d_60")
181 << BatchNormalizationLayer(
182 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_mean.npy"),
183 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_var.npy"),
184 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_gamma.npy"),
185 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_beta.npy"),
186 0.000001f)
187 .set_name("conv2d_59/BatchNorm")
188 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_60/LeakyRelu")
189 << UpsampleLayer(Size2D(2, 2), InterpolationPolicy::NEAREST_NEIGHBOR).set_name("Upsample_60");
190 SubStream concat_1(route_1);
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]191 concat_1 << ConcatLayer(std::move(route_1), std::move(intermediate_layers.second)).set_name("Route1")
Michalis Spyroue22aa132018-09-13 10:35:33 +0100[diff] [blame]192 << ConvolutionLayer(
193 1U, 1U, 256U,
194 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_61_w.npy", weights_layout),
195 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
196 PadStrideInfo(1, 1, 0, 0))
197 .set_name("conv2d_61")
198 << BatchNormalizationLayer(
199 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_mean.npy"),
200 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_var.npy"),
201 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_gamma.npy"),
202 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_beta.npy"),
203 0.000001f)
204 .set_name("conv2d_60/BatchNorm")
205 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_61/LeakyRelu")
206 << ConvolutionLayer(
207 3U, 3U, 512U,
208 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_62_w.npy", weights_layout),
209 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
210 PadStrideInfo(1, 1, 1, 1))
211 .set_name("conv2d_62")
212 << BatchNormalizationLayer(
213 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_mean.npy"),
214 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_var.npy"),
215 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_gamma.npy"),
216 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_beta.npy"),
217 0.000001f)
218 .set_name("conv2d_61/BatchNorm")
219 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_62/LeakyRelu")
220 << ConvolutionLayer(
221 1U, 1U, 256U,
222 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_63_w.npy", weights_layout),
223 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
224 PadStrideInfo(1, 1, 0, 0))
225 .set_name("conv2d_63")
226 << BatchNormalizationLayer(
227 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_mean.npy"),
228 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_var.npy"),
229 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_gamma.npy"),
230 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_beta.npy"),
231 0.000001f)
232 .set_name("conv2d_62/BatchNorm")
233 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_63/LeakyRelu")
234 << ConvolutionLayer(
235 3U, 3U, 512U,
236 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_64_w.npy", weights_layout),
237 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
238 PadStrideInfo(1, 1, 1, 1))
239 .set_name("conv2d_64")
240 << BatchNormalizationLayer(
241 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_mean.npy"),
242 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_var.npy"),
243 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_gamma.npy"),
244 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_beta.npy"),
245 0.000001f)
246 .set_name("conv2d_63/BatchNorm")
247 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_64/LeakyRelu")
248 << ConvolutionLayer(
249 1U, 1U, 256U,
250 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_65_w.npy", weights_layout),
251 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
252 PadStrideInfo(1, 1, 0, 0))
253 .set_name("conv2d_65")
254 << BatchNormalizationLayer(
255 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_mean.npy"),
256 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_var.npy"),
257 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_gamma.npy"),
258 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_beta.npy"),
259 0.000001f)
260 .set_name("conv2d_65/BatchNorm")
261 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_65/LeakyRelu");
262 SubStream route_2(concat_1);
263 concat_1 << ConvolutionLayer(
264 3U, 3U, 512U,
265 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_66_w.npy", weights_layout),
266 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
267 PadStrideInfo(1, 1, 1, 1))
268 .set_name("conv2d_66")
269 << BatchNormalizationLayer(
270 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_mean.npy"),
271 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_var.npy"),
272 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_gamma.npy"),
273 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_beta.npy"),
274 0.000001f)
275 .set_name("conv2d_65/BatchNorm")
276 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_66/LeakyRelu")
277 << ConvolutionLayer(
278 1U, 1U, 255U,
279 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_67_w.npy", weights_layout),
280 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_67_b.npy", weights_layout),
281 PadStrideInfo(1, 1, 0, 0))
282 .set_name("conv2d_67")
283 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f)).set_name("conv2d_67/Linear")
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]284 << YOLOLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.1f), 80).set_name("Yolo2")
Michalis Spyroue22aa132018-09-13 10:35:33 +0100[diff] [blame]285 << OutputLayer(get_output_accessor(common_params, 5));
286 route_2 << ConvolutionLayer(
287 1U, 1U, 128U,
288 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_68_w.npy", weights_layout),
289 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
290 PadStrideInfo(1, 1, 0, 0))
291 .set_name("conv2d_68")
292 << BatchNormalizationLayer(
293 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_mean.npy"),
294 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_var.npy"),
295 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_gamma.npy"),
296 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_beta.npy"),
297 0.000001f)
298 .set_name("conv2d_66/BatchNorm")
299 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_68/LeakyRelu")
300 << UpsampleLayer(Size2D(2, 2), InterpolationPolicy::NEAREST_NEIGHBOR).set_name("Upsample_68");
301 SubStream concat_2(route_2);
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]302 concat_2 << ConcatLayer(std::move(route_2), std::move(intermediate_layers.first)).set_name("Route2")
Michalis Spyroue22aa132018-09-13 10:35:33 +0100[diff] [blame]303 << ConvolutionLayer(
304 1U, 1U, 128U,
305 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_69_w.npy", weights_layout),
306 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
307 PadStrideInfo(1, 1, 0, 0))
308 .set_name("conv2d_69")
309 << BatchNormalizationLayer(
310 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_mean.npy"),
311 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_var.npy"),
312 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_gamma.npy"),
313 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_beta.npy"),
314 0.000001f)
315 .set_name("conv2d_67/BatchNorm")
316 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_69/LeakyRelu")
317 << ConvolutionLayer(
318 3U, 3U, 256U,
319 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_70_w.npy", weights_layout),
320 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
321 PadStrideInfo(1, 1, 1, 1))
322 .set_name("conv2d_70")
323 << BatchNormalizationLayer(
324 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_mean.npy"),
325 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_var.npy"),
326 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_gamma.npy"),
327 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_beta.npy"),
328 0.000001f)
329 .set_name("conv2d_68/BatchNorm")
330 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_70/LeakyRelu")
331 << ConvolutionLayer(
332 1U, 1U, 128U,
333 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_71_w.npy", weights_layout),
334 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
335 PadStrideInfo(1, 1, 0, 0))
336 .set_name("conv2d_71")
337 << BatchNormalizationLayer(
338 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_mean.npy"),
339 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_var.npy"),
340 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_gamma.npy"),
341 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_beta.npy"),
342 0.000001f)
343 .set_name("conv2d_69/BatchNorm")
344 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_71/LeakyRelu")
345 << ConvolutionLayer(
346 3U, 3U, 256U,
347 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_72_w.npy", weights_layout),
348 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
349 PadStrideInfo(1, 1, 1, 1))
350 .set_name("conv2d_72")
351 << BatchNormalizationLayer(
352 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_mean.npy"),
353 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_var.npy"),
354 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_gamma.npy"),
355 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_beta.npy"),
356 0.000001f)
357 .set_name("conv2d_70/BatchNorm")
358 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_72/LeakyRelu")
359 << ConvolutionLayer(
360 1U, 1U, 128U,
361 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_73_w.npy", weights_layout),
362 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
363 PadStrideInfo(1, 1, 0, 0))
364 .set_name("conv2d_73")
365 << BatchNormalizationLayer(
366 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_mean.npy"),
367 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_var.npy"),
368 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_gamma.npy"),
369 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_beta.npy"),
370 0.000001f)
371 .set_name("conv2d_71/BatchNorm")
372 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_73/LeakyRelu")
373 << ConvolutionLayer(
374 3U, 3U, 256U,
375 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_74_w.npy", weights_layout),
376 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
377 PadStrideInfo(1, 1, 1, 1))
378 .set_name("conv2d_74")
379 << BatchNormalizationLayer(
380 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_mean.npy"),
381 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_var.npy"),
382 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_gamma.npy"),
383 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_beta.npy"),
384 0.000001f)
385 .set_name("conv2d_72/BatchNorm")
386 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_74/LeakyRelu")
387 << ConvolutionLayer(
388 1U, 1U, 255U,
389 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_75_w.npy", weights_layout),
390 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_75_b.npy", weights_layout),
391 PadStrideInfo(1, 1, 0, 0))
392 .set_name("conv2d_75")
393 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f)).set_name("conv2d_75/Linear")
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]394 << YOLOLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.1f), 80).set_name("Yolo3")
Michalis Spyroue22aa132018-09-13 10:35:33 +0100[diff] [blame]395 << OutputLayer(get_output_accessor(common_params, 5));
396
397 // Finalize graph
398 GraphConfig config;
399 config.num_threads = common_params.threads;
400 config.use_tuner = common_params.enable_tuner;
Vidhya Sudhan Loganathan050471e2019-04-25 09:27:24 +0100[diff] [blame^]401 config.tuner_mode = common_params.tuner_mode;
Michalis Spyroue22aa132018-09-13 10:35:33 +0100[diff] [blame]402 config.tuner_file = common_params.tuner_file;
403
404 graph.finalize(common_params.target, config);
405
406 return true;
407 }
408 void do_run() override
409 {
410 // Run graph
411 graph.run();
412 }
413
414private:
415 CommandLineParser cmd_parser;
416 CommonGraphOptions common_opts;
417 CommonGraphParams common_params;
418 Stream graph;
419
420 std::pair<SubStream, SubStream> darknet53(const std::string &data_path, DataLayout weights_layout)
421 {
422 graph << ConvolutionLayer(
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]423 3U, 3U, 32U,
424 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_1_w.npy", weights_layout),
425 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
426 PadStrideInfo(1, 1, 1, 1))
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]427 .set_name("conv2d_1/Conv2D")
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]428 << BatchNormalizationLayer(
429 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_mean.npy"),
430 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_var.npy"),
431 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_gamma.npy"),
432 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_beta.npy"),
433 0.000001f)
434 .set_name("conv2d_1/BatchNorm")
435 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_1/LeakyRelu")
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]436 << ConvolutionLayer(
437 3U, 3U, 64U,
438 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_2_w.npy", weights_layout),
439 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
440 PadStrideInfo(2, 2, 1, 1))
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]441 .set_name("conv2d_2/Conv2D")
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]442 << BatchNormalizationLayer(
443 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_mean.npy"),
444 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_var.npy"),
445 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_gamma.npy"),
446 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_beta.npy"),
447 0.000001f)
448 .set_name("conv2d_2/BatchNorm")
449 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_2/LeakyRelu");
450 darknet53_block(data_path, "3", weights_layout, 32U);
451 graph << ConvolutionLayer(
452 3U, 3U, 128U,
453 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_5_w.npy", weights_layout),
454 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
455 PadStrideInfo(2, 2, 1, 1))
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]456 .set_name("conv2d_5/Conv2D")
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]457 << BatchNormalizationLayer(
458 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_mean.npy"),
459 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_var.npy"),
460 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_gamma.npy"),
461 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_beta.npy"),
462 0.000001f)
463 .set_name("conv2d_5/BatchNorm")
464 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_5/LeakyRelu");
465 darknet53_block(data_path, "6", weights_layout, 64U);
466 darknet53_block(data_path, "8", weights_layout, 64U);
467 graph << ConvolutionLayer(
468 3U, 3U, 256U,
469 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_10_w.npy", weights_layout),
470 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
471 PadStrideInfo(2, 2, 1, 1))
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]472 .set_name("conv2d_10/Conv2D")
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]473 << BatchNormalizationLayer(
474 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_mean.npy"),
475 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_var.npy"),
476 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_gamma.npy"),
477 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_beta.npy"),
478 0.000001f)
479 .set_name("conv2d_10/BatchNorm")
480 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_10/LeakyRelu");
481 darknet53_block(data_path, "11", weights_layout, 128U);
482 darknet53_block(data_path, "13", weights_layout, 128U);
483 darknet53_block(data_path, "15", weights_layout, 128U);
484 darknet53_block(data_path, "17", weights_layout, 128U);
485 darknet53_block(data_path, "19", weights_layout, 128U);
486 darknet53_block(data_path, "21", weights_layout, 128U);
487 darknet53_block(data_path, "23", weights_layout, 128U);
488 darknet53_block(data_path, "25", weights_layout, 128U);
Michalis Spyroue22aa132018-09-13 10:35:33 +0100[diff] [blame]489 SubStream layer_36(graph);
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]490 graph << ConvolutionLayer(
491 3U, 3U, 512U,
492 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_27_w.npy", weights_layout),
493 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
494 PadStrideInfo(2, 2, 1, 1))
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]495 .set_name("conv2d_27/Conv2D")
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]496 << BatchNormalizationLayer(
497 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_mean.npy"),
498 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_var.npy"),
499 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_gamma.npy"),
500 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_beta.npy"),
501 0.000001f)
502 .set_name("conv2d_27/BatchNorm")
503 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_27/LeakyRelu");
504 darknet53_block(data_path, "28", weights_layout, 256U);
505 darknet53_block(data_path, "30", weights_layout, 256U);
506 darknet53_block(data_path, "32", weights_layout, 256U);
507 darknet53_block(data_path, "34", weights_layout, 256U);
508 darknet53_block(data_path, "36", weights_layout, 256U);
509 darknet53_block(data_path, "38", weights_layout, 256U);
510 darknet53_block(data_path, "40", weights_layout, 256U);
511 darknet53_block(data_path, "42", weights_layout, 256U);
Michalis Spyroue22aa132018-09-13 10:35:33 +0100[diff] [blame]512 SubStream layer_61(graph);
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]513 graph << ConvolutionLayer(
514 3U, 3U, 1024U,
515 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_44_w.npy", weights_layout),
516 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
517 PadStrideInfo(2, 2, 1, 1))
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]518 .set_name("conv2d_44/Conv2D")
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]519 << BatchNormalizationLayer(
520 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_mean.npy"),
521 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_var.npy"),
522 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_gamma.npy"),
523 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_beta.npy"),
524 0.000001f)
525 .set_name("conv2d_44/BatchNorm")
526 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_44/LeakyRelu");
527 darknet53_block(data_path, "45", weights_layout, 512U);
528 darknet53_block(data_path, "47", weights_layout, 512U);
529 darknet53_block(data_path, "49", weights_layout, 512U);
530 darknet53_block(data_path, "51", weights_layout, 512U);
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]531
Michalis Spyroue22aa132018-09-13 10:35:33 +0100[diff] [blame]532 return std::pair<SubStream, SubStream>(layer_36, layer_61);
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]533 }
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]534
535 void darknet53_block(const std::string &data_path, std::string &&param_path, DataLayout weights_layout,
536 unsigned int filter_size)
537 {
538 std::string total_path = "/cnn_data/yolov3_model/";
Michalis Spyrouc98b9902018-09-12 11:28:39 +0100[diff] [blame]539 std::string param_path2 = arm_compute::support::cpp11::to_string(arm_compute::support::cpp11::stoi(param_path) + 1);
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]540 SubStream i_a(graph);
541 SubStream i_b(graph);
542 i_a << ConvolutionLayer(
543 1U, 1U, filter_size,
544 get_weights_accessor(data_path, total_path + "conv2d_" + param_path + "_w.npy", weights_layout),
545 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
546 PadStrideInfo(1, 1, 0, 0))
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]547 .set_name("conv2d_" + param_path + "/Conv2D")
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]548 << BatchNormalizationLayer(
549 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_mean.npy"),
550 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_var.npy"),
551 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_gamma.npy"),
552 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_beta.npy"),
553 0.000001f)
Michalis Spyroue22aa132018-09-13 10:35:33 +0100[diff] [blame]554 .set_name("conv2d_" + param_path + "/BatchNorm")
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]555 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_" + param_path + "/LeakyRelu")
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]556 << ConvolutionLayer(
557 3U, 3U, filter_size * 2,
558 get_weights_accessor(data_path, total_path + "conv2d_" + param_path2 + "_w.npy", weights_layout),
559 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
560 PadStrideInfo(1, 1, 1, 1))
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]561 .set_name("conv2d_" + param_path2 + "/Conv2D")
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]562 << BatchNormalizationLayer(
563 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_mean.npy"),
564 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_var.npy"),
565 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_gamma.npy"),
566 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_beta.npy"),
567 0.000001f)
Michalis Spyroue22aa132018-09-13 10:35:33 +0100[diff] [blame]568 .set_name("conv2d_" + param_path2 + "/BatchNorm")
569 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_" + param_path2 + "/LeakyRelu");
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]570
Georgios Pinitas62c36392019-01-31 12:53:10 +0000[diff] [blame]571 graph << EltwiseLayer(std::move(i_a), std::move(i_b), EltwiseOperation::Add).set_name("").set_name("add_" + param_path + "_" + param_path2);
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]572 }
573};
574
575/** Main program for YOLOv3
576 *
Georgios Pinitasbdbbbe82018-11-07 16:06:47 +0000[diff] [blame]577 * Model is based on:
578 * https://arxiv.org/abs/1804.02767
579 * "YOLOv3: An Incremental Improvement"
580 * Joseph Redmon, Ali Farhadi
581 *
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame]582 * @note To list all the possible arguments execute the binary appended with the --help option
583 *
584 * @param[in] argc Number of arguments
585 * @param[in] argv Arguments
586 *
587 * @return Return code
588 */
589int main(int argc, char **argv)
590{
591 return arm_compute::utils::run_example<GraphYOLOv3Example>(argc, argv);
592}