1 /*
2 * Copyright (c) 2017-2021 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 /** Example demonstrating how to implement VGG19's network using the Compute Library's graph API */
34 class GraphVGG19Example : public Example
35 {
36 public:
GraphVGG19Example()37 GraphVGG19Example()
38 : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "VGG19")
39 {
40 }
do_setup(int argc,char ** argv)41 bool do_setup(int argc, char **argv) override
42 {
43 // Parse arguments
44 cmd_parser.parse(argc, argv);
45 cmd_parser.validate();
46
47 // Consume common parameters
48 common_params = consume_common_graph_parameters(common_opts);
49
50 // Return when help menu is requested
51 if(common_params.help)
52 {
53 cmd_parser.print_help(argv[0]);
54 return false;
55 }
56
57 // Print parameter values
58 std::cout << common_params << std::endl;
59
60 // Get trainable parameters data path
61 std::string data_path = common_params.data_path;
62
63 // Create a preprocessor object
64 const std::array<float, 3> mean_rgb{ { 123.68f, 116.779f, 103.939f } };
65 std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<CaffePreproccessor>(mean_rgb);
66
67 // Create input descriptor
68 const auto operation_layout = common_params.data_layout;
69 const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout);
70 TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout);
71
72 // Set weights trained layout
73 const DataLayout weights_layout = DataLayout::NCHW;
74
75 graph << common_params.target
76 << common_params.fast_math_hint
77 << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor)))
78 // Layer 1
79 << ConvolutionLayer(
80 3U, 3U, 64U,
81 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_1_w.npy", weights_layout),
82 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_1_b.npy"),
83 PadStrideInfo(1, 1, 1, 1))
84 .set_name("conv1_1")
85 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1_1/Relu")
86 << ConvolutionLayer(
87 3U, 3U, 64U,
88 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_2_w.npy", weights_layout),
89 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_2_b.npy"),
90 PadStrideInfo(1, 1, 1, 1))
91 .set_name("conv1_2")
92 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1_2/Relu")
93 << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool1")
94 // Layer 2
95 << ConvolutionLayer(
96 3U, 3U, 128U,
97 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_1_w.npy", weights_layout),
98 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_1_b.npy"),
99 PadStrideInfo(1, 1, 1, 1))
100 .set_name("conv2_1")
101 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2_1/Relu")
102 << ConvolutionLayer(
103 3U, 3U, 128U,
104 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_2_w.npy", weights_layout),
105 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_2_b.npy"),
106 PadStrideInfo(1, 1, 1, 1))
107 .set_name("conv2_2")
108 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2_2/Relu")
109 << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool2")
110 // Layer 3
111 << ConvolutionLayer(
112 3U, 3U, 256U,
113 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_1_w.npy", weights_layout),
114 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_1_b.npy"),
115 PadStrideInfo(1, 1, 1, 1))
116 .set_name("conv3_1")
117 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_1/Relu")
118 << ConvolutionLayer(
119 3U, 3U, 256U,
120 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_2_w.npy", weights_layout),
121 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_2_b.npy"),
122 PadStrideInfo(1, 1, 1, 1))
123 .set_name("conv3_2")
124 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_2/Relu")
125 << ConvolutionLayer(
126 3U, 3U, 256U,
127 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_3_w.npy", weights_layout),
128 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_3_b.npy"),
129 PadStrideInfo(1, 1, 1, 1))
130 .set_name("conv3_3")
131 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_3/Relu")
132 << ConvolutionLayer(
133 3U, 3U, 256U,
134 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_4_w.npy", weights_layout),
135 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_4_b.npy"),
136 PadStrideInfo(1, 1, 1, 1))
137 .set_name("conv3_4")
138 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_4/Relu")
139 << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool3")
140 // Layer 4
141 << ConvolutionLayer(
142 3U, 3U, 512U,
143 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_1_w.npy", weights_layout),
144 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_1_b.npy"),
145 PadStrideInfo(1, 1, 1, 1))
146 .set_name("conv4_1")
147 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_1/Relu")
148 << ConvolutionLayer(
149 3U, 3U, 512U,
150 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_2_w.npy", weights_layout),
151 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_2_b.npy"),
152 PadStrideInfo(1, 1, 1, 1))
153 .set_name("conv4_2")
154 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_2/Relu")
155 << ConvolutionLayer(
156 3U, 3U, 512U,
157 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_3_w.npy", weights_layout),
158 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_3_b.npy"),
159 PadStrideInfo(1, 1, 1, 1))
160 .set_name("conv4_3")
161 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_3/Relu")
162 << ConvolutionLayer(
163 3U, 3U, 512U,
164 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_4_w.npy", weights_layout),
165 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_4_b.npy"),
166 PadStrideInfo(1, 1, 1, 1))
167 .set_name("conv4_4")
168 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_4/Relu")
169 << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool4")
170 // Layer 5
171 << ConvolutionLayer(
172 3U, 3U, 512U,
173 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_1_w.npy", weights_layout),
174 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_1_b.npy"),
175 PadStrideInfo(1, 1, 1, 1))
176 .set_name("conv5_1")
177 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_1/Relu")
178 << ConvolutionLayer(
179 3U, 3U, 512U,
180 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_2_w.npy", weights_layout),
181 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_2_b.npy"),
182 PadStrideInfo(1, 1, 1, 1))
183 .set_name("conv5_2")
184 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_2/Relu")
185 << ConvolutionLayer(
186 3U, 3U, 512U,
187 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_3_w.npy", weights_layout),
188 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_3_b.npy"),
189 PadStrideInfo(1, 1, 1, 1))
190 .set_name("conv5_3")
191 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_3/Relu")
192 << ConvolutionLayer(
193 3U, 3U, 512U,
194 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_4_w.npy", weights_layout),
195 get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_4_b.npy"),
196 PadStrideInfo(1, 1, 1, 1))
197 .set_name("conv5_4")
198 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_4/Relu")
199 << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool5")
200 // Layer 6
201 << FullyConnectedLayer(
202 4096U,
203 get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc6_w.npy", weights_layout),
204 get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc6_b.npy"))
205 .set_name("fc6")
206 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu")
207 // Layer 7
208 << FullyConnectedLayer(
209 4096U,
210 get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc7_w.npy", weights_layout),
211 get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc7_b.npy"))
212 .set_name("fc7")
213 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu_1")
214 // Layer 8
215 << FullyConnectedLayer(
216 1000U,
217 get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc8_w.npy", weights_layout),
218 get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc8_b.npy"))
219 .set_name("fc8")
220 // Softmax
221 << SoftmaxLayer().set_name("prob")
222 << OutputLayer(get_output_accessor(common_params, 5));
223
224 // Finalize graph
225 GraphConfig config;
226 config.num_threads = common_params.threads;
227 config.use_tuner = common_params.enable_tuner;
228 config.tuner_mode = common_params.tuner_mode;
229 config.tuner_file = common_params.tuner_file;
230 config.mlgo_file = common_params.mlgo_file;
231 config.use_synthetic_type = arm_compute::is_data_type_quantized(common_params.data_type);
232 config.synthetic_type = common_params.data_type;
233
234 graph.finalize(common_params.target, config);
235
236 return true;
237 }
do_run()238 void do_run() override
239 {
240 // Run graph
241 graph.run();
242 }
243
244 private:
245 CommandLineParser cmd_parser;
246 CommonGraphOptions common_opts;
247 CommonGraphParams common_params;
248 Stream graph;
249 };
250
251 /** Main program for VGG19
252 *
253 * Model is based on:
254 * https://arxiv.org/abs/1409.1556
255 * "Very Deep Convolutional Networks for Large-Scale Image Recognition"
256 * Karen Simonyan, Andrew Zisserman
257 *
258 * Provenance: www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_19_layers.caffemodel
259 *
260 * @note To list all the possible arguments execute the binary appended with the --help option
261 *
262 * @param[in] argc Number of arguments
263 * @param[in] argv Arguments
264 */
main(int argc,char ** argv)265 int main(int argc, char **argv)
266 {
267 return arm_compute::utils::run_example<GraphVGG19Example>(argc, argv);
268 }
269