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
34 /** Example demonstrating how to implement ResNetV1_50 network using the Compute Library's graph API */
35 class GraphResNetV1_50Example : public Example
36 {
37 public:
GraphResNetV1_50Example()38 GraphResNetV1_50Example()
39 : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ResNetV1_50")
40 {
41 }
do_setup(int argc,char ** argv)42 bool do_setup(int argc, char **argv) override
43 {
44 // Parse arguments
45 cmd_parser.parse(argc, argv);
46 cmd_parser.validate();
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 // Print parameter values
59 std::cout << common_params << std::endl;
60
61 // Get trainable parameters data path
62 std::string data_path = common_params.data_path;
63
64 // Create a preprocessor object
65 const std::array<float, 3> mean_rgb{ { 122.68f, 116.67f, 104.01f } };
66 std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<CaffePreproccessor>(mean_rgb,
67 false /* Do not convert to BGR */);
68
69 // Create input descriptor
70 const auto operation_layout = common_params.data_layout;
71 const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout);
72 TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_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
79 << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false /* Do not convert to BGR */))
80 << ConvolutionLayer(
81 7U, 7U, 64U,
82 get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_weights.npy", weights_layout),
83 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
84 PadStrideInfo(2, 2, 3, 3))
85 .set_name("conv1/convolution")
86 << BatchNormalizationLayer(
87 get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_moving_mean.npy"),
88 get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_moving_variance.npy"),
89 get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_gamma.npy"),
90 get_weights_accessor(data_path, "/cnn_data/resnet50_model/conv1_BatchNorm_beta.npy"),
91 0.0000100099996416f)
92 .set_name("conv1/BatchNorm")
93 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1/Relu")
94 << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))).set_name("pool1/MaxPool");
95
96 add_residual_block(data_path, "block1", weights_layout, 64, 3, 2);
97 add_residual_block(data_path, "block2", weights_layout, 128, 4, 2);
98 add_residual_block(data_path, "block3", weights_layout, 256, 6, 2);
99 add_residual_block(data_path, "block4", weights_layout, 512, 3, 1);
100
101 graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("pool5")
102 << ConvolutionLayer(
103 1U, 1U, 1000U,
104 get_weights_accessor(data_path, "/cnn_data/resnet50_model/logits_weights.npy", weights_layout),
105 get_weights_accessor(data_path, "/cnn_data/resnet50_model/logits_biases.npy"),
106 PadStrideInfo(1, 1, 0, 0))
107 .set_name("logits/convolution")
108 << FlattenLayer().set_name("predictions/Reshape")
109 << SoftmaxLayer().set_name("predictions/Softmax")
110 << OutputLayer(get_output_accessor(common_params, 5));
111
112 // Finalize graph
113 GraphConfig config;
114 config.num_threads = common_params.threads;
115 config.use_tuner = common_params.enable_tuner;
116 config.tuner_mode = common_params.tuner_mode;
117 config.tuner_file = common_params.tuner_file;
118 config.mlgo_file = common_params.mlgo_file;
119 config.use_synthetic_type = arm_compute::is_data_type_quantized(common_params.data_type);
120 config.synthetic_type = common_params.data_type;
121
122 graph.finalize(common_params.target, config);
123
124 return true;
125 }
126
do_run()127 void do_run() override
128 {
129 // Run graph
130 graph.run();
131 }
132
133 private:
134 CommandLineParser cmd_parser;
135 CommonGraphOptions common_opts;
136 CommonGraphParams common_params;
137 Stream graph;
138
add_residual_block(const std::string & data_path,const std::string & name,DataLayout weights_layout,unsigned int base_depth,unsigned int num_units,unsigned int stride)139 void add_residual_block(const std::string &data_path, const std::string &name, DataLayout weights_layout,
140 unsigned int base_depth, unsigned int num_units, unsigned int stride)
141 {
142 for(unsigned int i = 0; i < num_units; ++i)
143 {
144 std::stringstream unit_path_ss;
145 unit_path_ss << "/cnn_data/resnet50_model/" << name << "_unit_" << (i + 1) << "_bottleneck_v1_";
146 std::stringstream unit_name_ss;
147 unit_name_ss << name << "/unit" << (i + 1) << "/bottleneck_v1/";
148
149 std::string unit_path = unit_path_ss.str();
150 std::string unit_name = unit_name_ss.str();
151
152 unsigned int middle_stride = 1;
153
154 if(i == (num_units - 1))
155 {
156 middle_stride = stride;
157 }
158
159 SubStream right(graph);
160 right << ConvolutionLayer(
161 1U, 1U, base_depth,
162 get_weights_accessor(data_path, unit_path + "conv1_weights.npy", weights_layout),
163 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
164 PadStrideInfo(1, 1, 0, 0))
165 .set_name(unit_name + "conv1/convolution")
166 << BatchNormalizationLayer(
167 get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_mean.npy"),
168 get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_variance.npy"),
169 get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_gamma.npy"),
170 get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_beta.npy"),
171 0.0000100099996416f)
172 .set_name(unit_name + "conv1/BatchNorm")
173 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu")
174
175 << ConvolutionLayer(
176 3U, 3U, base_depth,
177 get_weights_accessor(data_path, unit_path + "conv2_weights.npy", weights_layout),
178 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
179 PadStrideInfo(middle_stride, middle_stride, 1, 1))
180 .set_name(unit_name + "conv2/convolution")
181 << BatchNormalizationLayer(
182 get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_mean.npy"),
183 get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_variance.npy"),
184 get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_gamma.npy"),
185 get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_beta.npy"),
186 0.0000100099996416f)
187 .set_name(unit_name + "conv2/BatchNorm")
188 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu")
189
190 << ConvolutionLayer(
191 1U, 1U, base_depth * 4,
192 get_weights_accessor(data_path, unit_path + "conv3_weights.npy", weights_layout),
193 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
194 PadStrideInfo(1, 1, 0, 0))
195 .set_name(unit_name + "conv3/convolution")
196 << BatchNormalizationLayer(
197 get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_moving_mean.npy"),
198 get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_moving_variance.npy"),
199 get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_gamma.npy"),
200 get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_beta.npy"),
201 0.0000100099996416f)
202 .set_name(unit_name + "conv2/BatchNorm");
203
204 if(i == 0)
205 {
206 SubStream left(graph);
207 left << ConvolutionLayer(
208 1U, 1U, base_depth * 4,
209 get_weights_accessor(data_path, unit_path + "shortcut_weights.npy", weights_layout),
210 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
211 PadStrideInfo(1, 1, 0, 0))
212 .set_name(unit_name + "shortcut/convolution")
213 << BatchNormalizationLayer(
214 get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_moving_mean.npy"),
215 get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_moving_variance.npy"),
216 get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_gamma.npy"),
217 get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_beta.npy"),
218 0.0000100099996416f)
219 .set_name(unit_name + "shortcut/BatchNorm");
220
221 graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(unit_name + "add");
222 }
223 else if(middle_stride > 1)
224 {
225 SubStream left(graph);
226 left << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 1, common_params.data_layout, PadStrideInfo(middle_stride, middle_stride, 0, 0), true)).set_name(unit_name + "shortcut/MaxPool");
227
228 graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(unit_name + "add");
229 }
230 else
231 {
232 SubStream left(graph);
233 graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(unit_name + "add");
234 }
235
236 graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu");
237 }
238 }
239 };
240
241 /** Main program for ResNetV1_50
242 *
243 * Model is based on:
244 * https://arxiv.org/abs/1512.03385
245 * "Deep Residual Learning for Image Recognition"
246 * Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
247 *
248 * Provenance: download.tensorflow.org/models/resnet_v1_50_2016_08_28.tar.gz
249 *
250 * @note To list all the possible arguments execute the binary appended with the --help option
251 *
252 * @param[in] argc Number of arguments
253 * @param[in] argv Arguments
254 */
main(int argc,char ** argv)255 int main(int argc, char **argv)
256 {
257 return arm_compute::utils::run_example<GraphResNetV1_50Example>(argc, argv);
258 }
259