Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 177a9a54d6b556edf39bb3ec968c074d1b865964 / . / examples / graph_yolov3.cpp
blob: d2f7ae5e59d9768cae7d7d9aa6a61259c3034c4e [file] [log] [blame]
Michalis Spyrou177a9a52018-09-06 15:10:22 +0100[diff] [blame^]1/*
2 * Copyright (c) 2018 ARM Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24#include "arm_compute/graph.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
34/** Example demonstrating how to implement YOLOv3 network using the Compute Library's graph API
35 *
36 * @param[in] argc Number of arguments
37 * @param[in] argv Arguments
38 */
39class GraphYOLOv3Example : public Example
40{
41public:
42 GraphYOLOv3Example()
43 : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "YOLOv3")
44 {
45 }
46 bool do_setup(int argc, char **argv) override
47 {
48 // Parse arguments
49 cmd_parser.parse(argc, argv);
50
51 // Consume common parameters
52 common_params = consume_common_graph_parameters(common_opts);
53
54 // Return when help menu is requested
55 if(common_params.help)
56 {
57 cmd_parser.print_help(argv[0]);
58 return false;
59 }
60
61 // Checks
62 ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph");
63
64 // Print parameter values
65 std::cout << common_params << std::endl;
66
67 // Get trainable parameters data path
68 std::string data_path = common_params.data_path;
69
70 // Create a preprocessor object
71 std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<TFPreproccessor>(0.f);
72
73 // Create input descriptor
74 const TensorShape tensor_shape = permute_shape(TensorShape(608U, 608U, 3U, 1U), DataLayout::NCHW, common_params.data_layout);
75 TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout);
76
77 // Set weights trained layout
78 const DataLayout weights_layout = DataLayout::NCHW;
79
80 graph << common_params.target
81 << common_params.fast_math_hint
82 << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false))
83 // Layer 1
84 << ConvolutionLayer(
85 3U, 3U, 32U,
86 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_1_w.npy", weights_layout),
87 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
88 PadStrideInfo(1, 1, 1, 1))
89 .set_name("conv2d_1")
90 << BatchNormalizationLayer(
91 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_mean.npy"),
92 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_var.npy"),
93 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_gamma.npy"),
94 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_1_beta.npy"),
95 0.000001f)
96 .set_name("conv2d_1/BatchNorm")
97 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_1/LeakyRelu")
98
99 // Layer 2
100 << ConvolutionLayer(
101 3U, 3U, 64U,
102 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_2_w.npy", weights_layout),
103 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
104 PadStrideInfo(2, 2, 1, 1))
105 .set_name("conv2d_2")
106 << BatchNormalizationLayer(
107 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_mean.npy"),
108 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_var.npy"),
109 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_gamma.npy"),
110 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_2_beta.npy"),
111 0.000001f)
112 .set_name("conv2d_2/BatchNorm")
113 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_2/LeakyRelu");
114 darknet53_block(data_path, "3", weights_layout, 32U);
115 graph << ConvolutionLayer(
116 3U, 3U, 128U,
117 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_5_w.npy", weights_layout),
118 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
119 PadStrideInfo(2, 2, 1, 1))
120 .set_name("conv2d_5")
121 << BatchNormalizationLayer(
122 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_mean.npy"),
123 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_var.npy"),
124 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_gamma.npy"),
125 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_5_beta.npy"),
126 0.000001f)
127 .set_name("conv2d_5/BatchNorm")
128 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_5/LeakyRelu");
129 darknet53_block(data_path, "6", weights_layout, 64U);
130 darknet53_block(data_path, "8", weights_layout, 64U);
131 graph << ConvolutionLayer(
132 3U, 3U, 256U,
133 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_10_w.npy", weights_layout),
134 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
135 PadStrideInfo(2, 2, 1, 1))
136 .set_name("conv2d_10")
137 << BatchNormalizationLayer(
138 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_mean.npy"),
139 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_var.npy"),
140 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_gamma.npy"),
141 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_10_beta.npy"),
142 0.000001f)
143 .set_name("conv2d_10/BatchNorm")
144 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_10/LeakyRelu");
145 darknet53_block(data_path, "11", weights_layout, 128U);
146 darknet53_block(data_path, "13", weights_layout, 128U);
147 darknet53_block(data_path, "15", weights_layout, 128U);
148 darknet53_block(data_path, "17", weights_layout, 128U);
149 darknet53_block(data_path, "19", weights_layout, 128U);
150 darknet53_block(data_path, "21", weights_layout, 128U);
151 darknet53_block(data_path, "23", weights_layout, 128U);
152 darknet53_block(data_path, "25", weights_layout, 128U);
153 graph << ConvolutionLayer(
154 3U, 3U, 512U,
155 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_27_w.npy", weights_layout),
156 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
157 PadStrideInfo(2, 2, 1, 1))
158 .set_name("conv2d_27")
159 << BatchNormalizationLayer(
160 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_mean.npy"),
161 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_var.npy"),
162 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_gamma.npy"),
163 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_27_beta.npy"),
164 0.000001f)
165 .set_name("conv2d_27/BatchNorm")
166 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_27/LeakyRelu");
167 darknet53_block(data_path, "28", weights_layout, 256U);
168 darknet53_block(data_path, "30", weights_layout, 256U);
169 darknet53_block(data_path, "32", weights_layout, 256U);
170 darknet53_block(data_path, "34", weights_layout, 256U);
171 darknet53_block(data_path, "36", weights_layout, 256U);
172 darknet53_block(data_path, "38", weights_layout, 256U);
173 darknet53_block(data_path, "40", weights_layout, 256U);
174 darknet53_block(data_path, "42", weights_layout, 256U);
175 graph << ConvolutionLayer(
176 3U, 3U, 1024U,
177 get_weights_accessor(data_path, "/cnn_data/yolov3_model/conv2d_44_w.npy", weights_layout),
178 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
179 PadStrideInfo(2, 2, 1, 1))
180 .set_name("conv2d_44")
181 << BatchNormalizationLayer(
182 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_mean.npy"),
183 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_var.npy"),
184 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_gamma.npy"),
185 get_weights_accessor(data_path, "/cnn_data/yolov3_model/batch_normalization_44_beta.npy"),
186 0.000001f)
187 .set_name("conv2d_44/BatchNorm")
188 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d_44/LeakyRelu");
189 darknet53_block(data_path, "45", weights_layout, 512U);
190 darknet53_block(data_path, "47", weights_layout, 512U);
191 darknet53_block(data_path, "49", weights_layout, 512U);
192 darknet53_block(data_path, "51", weights_layout, 512U);
193 graph << OutputLayer(get_output_accessor(common_params, 5));
194
195 // Finalize graph
196 GraphConfig config;
197 config.num_threads = common_params.threads;
198 config.use_tuner = common_params.enable_tuner;
199 config.tuner_file = common_params.tuner_file;
200
201 graph.finalize(common_params.target, config);
202
203 return true;
204 }
205 void do_run() override
206 {
207 // Run graph
208 graph.run();
209 }
210
211private:
212 CommandLineParser cmd_parser;
213 CommonGraphOptions common_opts;
214 CommonGraphParams common_params;
215 Stream graph;
216
217 void darknet53_block(const std::string &data_path, std::string &&param_path, DataLayout weights_layout,
218 unsigned int filter_size)
219 {
220 std::string total_path = "/cnn_data/yolov3_model/";
221 std::string param_path2 = std::to_string(std::stoi(param_path) + 1);
222 SubStream i_a(graph);
223 SubStream i_b(graph);
224 i_a << ConvolutionLayer(
225 1U, 1U, filter_size,
226 get_weights_accessor(data_path, total_path + "conv2d_" + param_path + "_w.npy", weights_layout),
227 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
228 PadStrideInfo(1, 1, 0, 0))
229 << BatchNormalizationLayer(
230 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_mean.npy"),
231 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_var.npy"),
232 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_gamma.npy"),
233 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path + "_beta.npy"),
234 0.000001f)
235 .set_name("conv2d" + param_path + "/BatchNorm")
236 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d" + param_path + "/LeakyRelu")
237 << ConvolutionLayer(
238 3U, 3U, filter_size * 2,
239 get_weights_accessor(data_path, total_path + "conv2d_" + param_path2 + "_w.npy", weights_layout),
240 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
241 PadStrideInfo(1, 1, 1, 1))
242 << BatchNormalizationLayer(
243 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_mean.npy"),
244 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_var.npy"),
245 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_gamma.npy"),
246 get_weights_accessor(data_path, total_path + "batch_normalization_" + param_path2 + "_beta.npy"),
247 0.000001f)
248 .set_name("conv2d" + param_path2 + "/BatchNorm")
249 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f)).set_name("conv2d" + param_path2 + "/LeakyRelu");
250
251 graph << EltwiseLayer(std::move(i_a), std::move(i_b), EltwiseOperation::Add);
252 }
253};
254
255/** Main program for YOLOv3
256 *
257 * @note To list all the possible arguments execute the binary appended with the --help option
258 *
259 * @param[in] argc Number of arguments
260 * @param[in] argv Arguments
261 *
262 * @return Return code
263 */
264int main(int argc, char **argv)
265{
266 return arm_compute::utils::run_example<GraphYOLOv3Example>(argc, argv);
267}