1 /*
2 * Copyright (c) 2018-2021 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24 #include "arm_compute/graph.h"
25 #include "support/ToolchainSupport.h"
26 #include "utils/CommonGraphOptions.h"
27 #include "utils/GraphUtils.h"
28 #include "utils/Utils.h"
29
30 using namespace arm_compute::utils;
31 using namespace arm_compute::graph::frontend;
32 using namespace arm_compute::graph_utils;
33
34 /** Example demonstrating how to implement YOLOv3 network using the Compute Library's graph API */
35 class GraphYOLOv3Example : public Example
36 {
37 public:
GraphYOLOv3Example()38 GraphYOLOv3Example()
39 : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "YOLOv3")
40 {
41 }
42
do_setup(int argc,char ** argv)43 bool do_setup(int argc, char **argv) override
44 {
45 // Parse arguments
46 cmd_parser.parse(argc, argv);
47 cmd_parser.validate();
48
49 // Consume common parameters
50 common_params = consume_common_graph_parameters(common_opts);
51
52 // Return when help menu is requested
53 if(common_params.help)
54 {
55 cmd_parser.print_help(argv[0]);
56 return false;
57 }
58
59 // Checks
60 ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph");
61
62 // Print parameter values
63 std::cout << common_params << std::endl;
64
65 // Get trainable parameters data path
66 std::string data_path = common_params.data_path;
67
68 // Create a preprocessor object
69 std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<TFPreproccessor>(0.f);
70
71 // Create input descriptor
72 const TensorShape tensor_shape = permute_shape(TensorShape(608U, 608U, 3U, 1U), DataLayout::NCHW, common_params.data_layout);
73 TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout);
74
75 // Set weights trained layout
76 const DataLayout weights_layout = DataLayout::NCHW;
77
78 graph << common_params.target
79 << common_params.fast_math_hint
80 << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false));
81 std::pair<SubStream, SubStream> intermediate_layers = darknet53(data_path, weights_layout);
82 graph << ConvolutionLayer(
83 1U, 1U, 512U,
84 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_53_w.npy", weights_layout),
85 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
86 PadStrideInfo(1, 1, 0, 0))
87 .set_name("conv2d_53")
88 << BatchNormalizationLayer(
89 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_mean.npy"),
90 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_var.npy"),
91 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_gamma.npy"),
92 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_53_beta.npy"),
93 0.000001f)
94 .set_name("conv2d_53/BatchNorm")
95 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_53/LeakyRelu")
96 << ConvolutionLayer(
97 3U, 3U, 1024U,
98 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_54_w.npy", weights_layout),
99 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
100 PadStrideInfo(1, 1, 1, 1))
101 .set_name("conv2d_54")
102 << BatchNormalizationLayer(
103 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_mean.npy"),
104 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_var.npy"),
105 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_gamma.npy"),
106 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_54_beta.npy"),
107 0.000001f)
108 .set_name("conv2d_54/BatchNorm")
109 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_54/LeakyRelu")
110 << ConvolutionLayer(
111 1U, 1U, 512U,
112 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_55_w.npy", weights_layout),
113 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
114 PadStrideInfo(1, 1, 0, 0))
115 .set_name("conv2d_55")
116 << BatchNormalizationLayer(
117 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_mean.npy"),
118 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_var.npy"),
119 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_gamma.npy"),
120 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_55_beta.npy"),
121 0.000001f)
122 .set_name("conv2d_55/BatchNorm")
123 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_55/LeakyRelu")
124 << ConvolutionLayer(
125 3U, 3U, 1024U,
126 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_56_w.npy", weights_layout),
127 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
128 PadStrideInfo(1, 1, 1, 1))
129 .set_name("conv2d_56")
130 << BatchNormalizationLayer(
131 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_mean.npy"),
132 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_var.npy"),
133 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_gamma.npy"),
134 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_56_beta.npy"),
135 0.000001f)
136 .set_name("conv2d_56/BatchNorm")
137 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_56/LeakyRelu")
138 << ConvolutionLayer(
139 1U, 1U, 512U,
140 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_57_w.npy", weights_layout),
141 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
142 PadStrideInfo(1, 1, 0, 0))
143 .set_name("conv2d_57")
144 << BatchNormalizationLayer(
145 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_mean.npy"),
146 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_var.npy"),
147 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_gamma.npy"),
148 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_57_beta.npy"),
149 0.000001f)
150 .set_name("conv2d_57/BatchNorm")
151 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_57/LeakyRelu");
152 SubStream route_1(graph);
153 graph << ConvolutionLayer(
154 3U, 3U, 1024U,
155 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_58_w.npy", weights_layout),
156 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
157 PadStrideInfo(1, 1, 1, 1))
158 .set_name("conv2d_58")
159 << BatchNormalizationLayer(
160 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_mean.npy"),
161 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_var.npy"),
162 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_gamma.npy"),
163 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_58_beta.npy"),
164 0.000001f)
165 .set_name("conv2d_58/BatchNorm")
166 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_58/LeakyRelu")
167 << ConvolutionLayer(
168 1U, 1U, 255U,
169 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_59_w.npy", weights_layout),
170 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_59_b.npy", weights_layout),
171 PadStrideInfo(1, 1, 0, 0))
172 .set_name("conv2d_59")
173 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f)).set_name("conv2d_59/Linear")
174 << YOLOLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.1f)).set_name("Yolo1")
175 << OutputLayer(get_output_accessor(common_params, 5));
176 route_1 << ConvolutionLayer(
177 1U, 1U, 256U,
178 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_60_w.npy", weights_layout),
179 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
180 PadStrideInfo(1, 1, 0, 0))
181 .set_name("conv2d_60")
182 << BatchNormalizationLayer(
183 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_mean.npy"),
184 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_var.npy"),
185 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_gamma.npy"),
186 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_59_beta.npy"),
187 0.000001f)
188 .set_name("conv2d_59/BatchNorm")
189 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_60/LeakyRelu")
190 << ResizeLayer(InterpolationPolicy::NEAREST_NEIGHBOR, 2, 2).set_name("Upsample_60");
191 SubStream concat_1(route_1);
192 concat_1 << ConcatLayer(std::move(route_1), std::move(intermediate_layers.second)).set_name("Route1")
193 << ConvolutionLayer(
194 1U, 1U, 256U,
195 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_61_w.npy", weights_layout),
196 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
197 PadStrideInfo(1, 1, 0, 0))
198 .set_name("conv2d_61")
199 << BatchNormalizationLayer(
200 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_mean.npy"),
201 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_var.npy"),
202 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_gamma.npy"),
203 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_60_beta.npy"),
204 0.000001f)
205 .set_name("conv2d_60/BatchNorm")
206 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_61/LeakyRelu")
207 << ConvolutionLayer(
208 3U, 3U, 512U,
209 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_62_w.npy", weights_layout),
210 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
211 PadStrideInfo(1, 1, 1, 1))
212 .set_name("conv2d_62")
213 << BatchNormalizationLayer(
214 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_mean.npy"),
215 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_var.npy"),
216 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_gamma.npy"),
217 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_61_beta.npy"),
218 0.000001f)
219 .set_name("conv2d_61/BatchNorm")
220 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_62/LeakyRelu")
221 << ConvolutionLayer(
222 1U, 1U, 256U,
223 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_63_w.npy", weights_layout),
224 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
225 PadStrideInfo(1, 1, 0, 0))
226 .set_name("conv2d_63")
227 << BatchNormalizationLayer(
228 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_mean.npy"),
229 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_var.npy"),
230 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_gamma.npy"),
231 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_62_beta.npy"),
232 0.000001f)
233 .set_name("conv2d_62/BatchNorm")
234 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_63/LeakyRelu")
235 << ConvolutionLayer(
236 3U, 3U, 512U,
237 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_64_w.npy", weights_layout),
238 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
239 PadStrideInfo(1, 1, 1, 1))
240 .set_name("conv2d_64")
241 << BatchNormalizationLayer(
242 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_mean.npy"),
243 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_var.npy"),
244 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_gamma.npy"),
245 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_63_beta.npy"),
246 0.000001f)
247 .set_name("conv2d_63/BatchNorm")
248 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_64/LeakyRelu")
249 << ConvolutionLayer(
250 1U, 1U, 256U,
251 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_65_w.npy", weights_layout),
252 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
253 PadStrideInfo(1, 1, 0, 0))
254 .set_name("conv2d_65")
255 << BatchNormalizationLayer(
256 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_mean.npy"),
257 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_var.npy"),
258 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_gamma.npy"),
259 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_64_beta.npy"),
260 0.000001f)
261 .set_name("conv2d_65/BatchNorm")
262 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_65/LeakyRelu");
263 SubStream route_2(concat_1);
264 concat_1 << ConvolutionLayer(
265 3U, 3U, 512U,
266 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_66_w.npy", weights_layout),
267 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
268 PadStrideInfo(1, 1, 1, 1))
269 .set_name("conv2d_66")
270 << BatchNormalizationLayer(
271 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_mean.npy"),
272 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_var.npy"),
273 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_gamma.npy"),
274 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_65_beta.npy"),
275 0.000001f)
276 .set_name("conv2d_65/BatchNorm")
277 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_66/LeakyRelu")
278 << ConvolutionLayer(
279 1U, 1U, 255U,
280 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_67_w.npy", weights_layout),
281 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_67_b.npy", weights_layout),
282 PadStrideInfo(1, 1, 0, 0))
283 .set_name("conv2d_67")
284 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f)).set_name("conv2d_67/Linear")
285 << YOLOLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.1f)).set_name("Yolo2")
286 << OutputLayer(get_output_accessor(common_params, 5));
287 route_2 << ConvolutionLayer(
288 1U, 1U, 128U,
289 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_68_w.npy", weights_layout),
290 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
291 PadStrideInfo(1, 1, 0, 0))
292 .set_name("conv2d_68")
293 << BatchNormalizationLayer(
294 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_mean.npy"),
295 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_var.npy"),
296 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_gamma.npy"),
297 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_66_beta.npy"),
298 0.000001f)
299 .set_name("conv2d_66/BatchNorm")
300 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_68/LeakyRelu")
301 << ResizeLayer(InterpolationPolicy::NEAREST_NEIGHBOR, 2, 2).set_name("Upsample_68");
302 SubStream concat_2(route_2);
303 concat_2 << ConcatLayer(std::move(route_2), std::move(intermediate_layers.first)).set_name("Route2")
304 << ConvolutionLayer(
305 1U, 1U, 128U,
306 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_69_w.npy", weights_layout),
307 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
308 PadStrideInfo(1, 1, 0, 0))
309 .set_name("conv2d_69")
310 << BatchNormalizationLayer(
311 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_mean.npy"),
312 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_var.npy"),
313 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_gamma.npy"),
314 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_67_beta.npy"),
315 0.000001f)
316 .set_name("conv2d_67/BatchNorm")
317 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_69/LeakyRelu")
318 << ConvolutionLayer(
319 3U, 3U, 256U,
320 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_70_w.npy", weights_layout),
321 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
322 PadStrideInfo(1, 1, 1, 1))
323 .set_name("conv2d_70")
324 << BatchNormalizationLayer(
325 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_mean.npy"),
326 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_var.npy"),
327 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_gamma.npy"),
328 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_68_beta.npy"),
329 0.000001f)
330 .set_name("conv2d_68/BatchNorm")
331 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_70/LeakyRelu")
332 << ConvolutionLayer(
333 1U, 1U, 128U,
334 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_71_w.npy", weights_layout),
335 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
336 PadStrideInfo(1, 1, 0, 0))
337 .set_name("conv2d_71")
338 << BatchNormalizationLayer(
339 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_mean.npy"),
340 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_var.npy"),
341 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_gamma.npy"),
342 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_69_beta.npy"),
343 0.000001f)
344 .set_name("conv2d_69/BatchNorm")
345 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_71/LeakyRelu")
346 << ConvolutionLayer(
347 3U, 3U, 256U,
348 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_72_w.npy", weights_layout),
349 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
350 PadStrideInfo(1, 1, 1, 1))
351 .set_name("conv2d_72")
352 << BatchNormalizationLayer(
353 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_mean.npy"),
354 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_var.npy"),
355 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_gamma.npy"),
356 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_70_beta.npy"),
357 0.000001f)
358 .set_name("conv2d_70/BatchNorm")
359 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_72/LeakyRelu")
360 << ConvolutionLayer(
361 1U, 1U, 128U,
362 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_73_w.npy", weights_layout),
363 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
364 PadStrideInfo(1, 1, 0, 0))
365 .set_name("conv2d_73")
366 << BatchNormalizationLayer(
367 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_mean.npy"),
368 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_var.npy"),
369 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_gamma.npy"),
370 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_71_beta.npy"),
371 0.000001f)
372 .set_name("conv2d_71/BatchNorm")
373 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_73/LeakyRelu")
374 << ConvolutionLayer(
375 3U, 3U, 256U,
376 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_74_w.npy", weights_layout),
377 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
378 PadStrideInfo(1, 1, 1, 1))
379 .set_name("conv2d_74")
380 << BatchNormalizationLayer(
381 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_mean.npy"),
382 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_var.npy"),
383 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_gamma.npy"),
384 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_72_beta.npy"),
385 0.000001f)
386 .set_name("conv2d_72/BatchNorm")
387 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_74/LeakyRelu")
388 << ConvolutionLayer(
389 1U, 1U, 255U,
390 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_75_w.npy", weights_layout),
391 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_75_b.npy", weights_layout),
392 PadStrideInfo(1, 1, 0, 0))
393 .set_name("conv2d_75")
394 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 1.f)).set_name("conv2d_75/Linear")
395 << YOLOLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.1f)).set_name("Yolo3")
396 << OutputLayer(get_output_accessor(common_params, 5));
397
398 // Finalize graph
399 GraphConfig config;
400 config.num_threads = common_params.threads;
401 config.use_tuner = common_params.enable_tuner;
402 config.tuner_mode = common_params.tuner_mode;
403 config.tuner_file = common_params.tuner_file;
404 config.mlgo_file = common_params.mlgo_file;
405
406 graph.finalize(common_params.target, config);
407
408 return true;
409 }
do_run()410 void do_run() override
411 {
412 // Run graph
413 graph.run();
414 }
415
416 private:
417 CommandLineParser cmd_parser;
418 CommonGraphOptions common_opts;
419 CommonGraphParams common_params;
420 Stream graph;
421
darknet53(const std::string & data_path,DataLayout weights_layout)422 std::pair<SubStream, SubStream> darknet53(const std::string &data_path, DataLayout weights_layout)
423 {
424 graph << ConvolutionLayer(
425 3U, 3U, 32U,
426 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_1_w.npy", weights_layout),
427 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
428 PadStrideInfo(1, 1, 1, 1))
429 .set_name("conv2d_1/Conv2D")
430 << BatchNormalizationLayer(
431 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_mean.npy"),
432 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_var.npy"),
433 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_gamma.npy"),
434 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_beta.npy"),
435 0.000001f)
436 .set_name("conv2d_1/BatchNorm")
437 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_1/LeakyRelu")
438 << ConvolutionLayer(
439 3U, 3U, 64U,
440 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_2_w.npy", weights_layout),
441 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
442 PadStrideInfo(2, 2, 1, 1))
443 .set_name("conv2d_2/Conv2D")
444 << BatchNormalizationLayer(
445 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_mean.npy"),
446 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_var.npy"),
447 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_gamma.npy"),
448 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_beta.npy"),
449 0.000001f)
450 .set_name("conv2d_2/BatchNorm")
451 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_2/LeakyRelu");
452 darknet53_block(data_path, "3", weights_layout, 32U);
453 graph << ConvolutionLayer(
454 3U, 3U, 128U,
455 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_5_w.npy", weights_layout),
456 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
457 PadStrideInfo(2, 2, 1, 1))
458 .set_name("conv2d_5/Conv2D")
459 << BatchNormalizationLayer(
460 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_mean.npy"),
461 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_var.npy"),
462 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_gamma.npy"),
463 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_beta.npy"),
464 0.000001f)
465 .set_name("conv2d_5/BatchNorm")
466 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_5/LeakyRelu");
467 darknet53_block(data_path, "6", weights_layout, 64U);
468 darknet53_block(data_path, "8", weights_layout, 64U);
469 graph << ConvolutionLayer(
470 3U, 3U, 256U,
471 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_10_w.npy", weights_layout),
472 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
473 PadStrideInfo(2, 2, 1, 1))
474 .set_name("conv2d_10/Conv2D")
475 << BatchNormalizationLayer(
476 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_mean.npy"),
477 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_var.npy"),
478 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_gamma.npy"),
479 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_beta.npy"),
480 0.000001f)
481 .set_name("conv2d_10/BatchNorm")
482 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_10/LeakyRelu");
483 darknet53_block(data_path, "11", weights_layout, 128U);
484 darknet53_block(data_path, "13", weights_layout, 128U);
485 darknet53_block(data_path, "15", weights_layout, 128U);
486 darknet53_block(data_path, "17", weights_layout, 128U);
487 darknet53_block(data_path, "19", weights_layout, 128U);
488 darknet53_block(data_path, "21", weights_layout, 128U);
489 darknet53_block(data_path, "23", weights_layout, 128U);
490 darknet53_block(data_path, "25", weights_layout, 128U);
491 SubStream layer_36(graph);
492 graph << ConvolutionLayer(
493 3U, 3U, 512U,
494 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_27_w.npy", weights_layout),
495 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
496 PadStrideInfo(2, 2, 1, 1))
497 .set_name("conv2d_27/Conv2D")
498 << BatchNormalizationLayer(
499 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_mean.npy"),
500 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_var.npy"),
501 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_gamma.npy"),
502 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_beta.npy"),
503 0.000001f)
504 .set_name("conv2d_27/BatchNorm")
505 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_27/LeakyRelu");
506 darknet53_block(data_path, "28", weights_layout, 256U);
507 darknet53_block(data_path, "30", weights_layout, 256U);
508 darknet53_block(data_path, "32", weights_layout, 256U);
509 darknet53_block(data_path, "34", weights_layout, 256U);
510 darknet53_block(data_path, "36", weights_layout, 256U);
511 darknet53_block(data_path, "38", weights_layout, 256U);
512 darknet53_block(data_path, "40", weights_layout, 256U);
513 darknet53_block(data_path, "42", weights_layout, 256U);
514 SubStream layer_61(graph);
515 graph << ConvolutionLayer(
516 3U, 3U, 1024U,
517 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_44_w.npy", weights_layout),
518 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
519 PadStrideInfo(2, 2, 1, 1))
520 .set_name("conv2d_44/Conv2D")
521 << BatchNormalizationLayer(
522 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_mean.npy"),
523 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_var.npy"),
524 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_gamma.npy"),
525 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_beta.npy"),
526 0.000001f)
527 .set_name("conv2d_44/BatchNorm")
528 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_44/LeakyRelu");
529 darknet53_block(data_path, "45", weights_layout, 512U);
530 darknet53_block(data_path, "47", weights_layout, 512U);
531 darknet53_block(data_path, "49", weights_layout, 512U);
532 darknet53_block(data_path, "51", weights_layout, 512U);
533
534 return std::pair<SubStream, SubStream>(layer_36, layer_61);
535 }
536
darknet53_block(const std::string & data_path,std::string && param_path,DataLayout weights_layout,unsigned int filter_size)537 void darknet53_block(const std::string &data_path, std::string &¶m_path, DataLayout weights_layout,
538 unsigned int filter_size)
539 {
540 std::string total_path = "/cnn_data/yolov3_model/";
541 std::string param_path2 = arm_compute::support::cpp11::to_string(arm_compute::support::cpp11::stoi(param_path) + 1);
542 SubStream i_a(graph);
543 SubStream i_b(graph);
544 i_a << ConvolutionLayer(
545 1U, 1U, filter_size,
546 get_weights_accessor(data_path, total_path + "conv2d_" + param_path + "_w.npy", weights_layout),
547 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
548 PadStrideInfo(1, 1, 0, 0))
549 .set_name("conv2d_" + param_path + "/Conv2D")
550 << BatchNormalizationLayer(
551 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_mean.npy"),
552 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_var.npy"),
553 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_gamma.npy"),
554 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_beta.npy"),
555 0.000001f)
556 .set_name("conv2d_" + param_path + "/BatchNorm")
557 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_" + param_path + "/LeakyRelu")
558 << ConvolutionLayer(
559 3U, 3U, filter_size * 2,
560 get_weights_accessor(data_path, total_path + "conv2d_" + param_path2 + "_w.npy", weights_layout),
561 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
562 PadStrideInfo(1, 1, 1, 1))
563 .set_name("conv2d_" + param_path2 + "/Conv2D")
564 << BatchNormalizationLayer(
565 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_mean.npy"),
566 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_var.npy"),
567 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_gamma.npy"),
568 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_beta.npy"),
569 0.000001f)
570 .set_name("conv2d_" + param_path2 + "/BatchNorm")
571 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_" + param_path2 + "/LeakyRelu");
572
573 graph << EltwiseLayer(std::move(i_a), std::move(i_b), EltwiseOperation::Add).set_name("").set_name("add_" + param_path + "_" + param_path2);
574 }
575 };
576
577 /** Main program for YOLOv3
578 *
579 * Model is based on:
580 * https://arxiv.org/abs/1804.02767
581 * "YOLOv3: An Incremental Improvement"
582 * Joseph Redmon, Ali Farhadi
583 *
584 * @note To list all the possible arguments execute the binary appended with the --help option
585 *
586 * @param[in] argc Number of arguments
587 * @param[in] argv Arguments
588 *
589 * @return Return code
590 */
main(int argc,char ** argv)591 int main(int argc, char **argv)
592 {
593 return arm_compute::utils::run_example<GraphYOLOv3Example>(argc, argv);
594 }
595