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 ShuffleNet network using the Compute Library's graph API */
35 class ShuffleNetExample : public Example
36 {
37 public:
ShuffleNetExample()38 ShuffleNetExample()
39 : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ShuffleNet")
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 // Set default layout if needed (Single kernel grouped convolution not yet supported int NHWC)
59 if(!common_opts.data_layout->is_set())
60 {
61 common_params.data_layout = DataLayout::NHWC;
62 }
63
64 // Checks
65 ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph");
66
67 // Print parameter values
68 std::cout << common_params << std::endl;
69 std::cout << "Model: Shufflenet_1_g4" << std::endl;
70
71 // Create model path
72 std::string model_path = "/cnn_data/shufflenet_model/";
73
74 // Get trainable parameters data path
75 std::string data_path = common_params.data_path;
76
77 // Add model path to data path
78 if(!data_path.empty())
79 {
80 data_path += model_path;
81 }
82
83 // Create input descriptor
84 const auto operation_layout = common_params.data_layout;
85 const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout);
86 TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout);
87
88 // Set weights trained layout
89 const DataLayout weights_layout = DataLayout::NCHW;
90
91 // Create preprocessor
92 std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<TFPreproccessor>(0);
93
94 graph << common_params.target
95 << common_params.fast_math_hint
96 << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false /* Do not convert to BGR */))
97 << ConvolutionLayer(
98 3U, 3U, 24U,
99 get_weights_accessor(data_path, "conv3_0_w_0.npy", weights_layout),
100 get_weights_accessor(data_path, "conv3_0_b_0.npy", weights_layout),
101 PadStrideInfo(2, 2, 1, 1))
102 .set_name("Conv1/convolution")
103 << BatchNormalizationLayer(
104 get_weights_accessor(data_path, "conv3_0_bn_rm_0.npy"),
105 get_weights_accessor(data_path, "conv3_0_bn_riv_0.npy"),
106 get_weights_accessor(data_path, "conv3_0_bn_s_0.npy"),
107 get_weights_accessor(data_path, "conv3_0_bn_b_0.npy"),
108 1e-5f)
109 .set_name("Conv1/BatchNorm")
110 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Conv1/Relu")
111 << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 1, 1))).set_name("pool1/MaxPool");
112
113 // Stage 2
114 add_residual_block(data_path, DataLayout::NCHW, 0U /* unit */, 112U /* depth */, 2U /* stride */);
115 add_residual_block(data_path, DataLayout::NCHW, 1U /* unit */, 136U /* depth */, 1U /* stride */);
116 add_residual_block(data_path, DataLayout::NCHW, 2U /* unit */, 136U /* depth */, 1U /* stride */);
117 add_residual_block(data_path, DataLayout::NCHW, 3U /* unit */, 136U /* depth */, 1U /* stride */);
118
119 // Stage 3
120 add_residual_block(data_path, DataLayout::NCHW, 4U /* unit */, 136U /* depth */, 2U /* stride */);
121 add_residual_block(data_path, DataLayout::NCHW, 5U /* unit */, 272U /* depth */, 1U /* stride */);
122 add_residual_block(data_path, DataLayout::NCHW, 6U /* unit */, 272U /* depth */, 1U /* stride */);
123 add_residual_block(data_path, DataLayout::NCHW, 7U /* unit */, 272U /* depth */, 1U /* stride */);
124 add_residual_block(data_path, DataLayout::NCHW, 8U /* unit */, 272U /* depth */, 1U /* stride */);
125 add_residual_block(data_path, DataLayout::NCHW, 9U /* unit */, 272U /* depth */, 1U /* stride */);
126 add_residual_block(data_path, DataLayout::NCHW, 10U /* unit */, 272U /* depth */, 1U /* stride */);
127 add_residual_block(data_path, DataLayout::NCHW, 11U /* unit */, 272U /* depth */, 1U /* stride */);
128
129 // Stage 4
130 add_residual_block(data_path, DataLayout::NCHW, 12U /* unit */, 272U /* depth */, 2U /* stride */);
131 add_residual_block(data_path, DataLayout::NCHW, 13U /* unit */, 544U /* depth */, 1U /* stride */);
132 add_residual_block(data_path, DataLayout::NCHW, 14U /* unit */, 544U /* depth */, 1U /* stride */);
133 add_residual_block(data_path, DataLayout::NCHW, 15U /* unit */, 544U /* depth */, 1U /* stride */);
134
135 graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("predictions/AvgPool")
136 << FlattenLayer().set_name("predictions/Reshape")
137 << FullyConnectedLayer(
138 1000U,
139 get_weights_accessor(data_path, "pred_w_0.npy", weights_layout),
140 get_weights_accessor(data_path, "pred_b_0.npy"))
141 .set_name("predictions/FC")
142 << SoftmaxLayer().set_name("predictions/Softmax")
143 << OutputLayer(get_output_accessor(common_params, 5));
144
145 // Finalize graph
146 GraphConfig config;
147 config.num_threads = common_params.threads;
148 config.use_tuner = common_params.enable_tuner;
149 config.tuner_mode = common_params.tuner_mode;
150 config.tuner_file = common_params.tuner_file;
151 config.mlgo_file = common_params.mlgo_file;
152
153 graph.finalize(common_params.target, config);
154
155 return true;
156 }
157
do_run()158 void do_run() override
159 {
160 // Run graph
161 graph.run();
162 }
163
164 private:
165 CommandLineParser cmd_parser;
166 CommonGraphOptions common_opts;
167 CommonGraphParams common_params;
168 Stream graph;
169
add_residual_block(const std::string & data_path,DataLayout weights_layout,unsigned int unit,unsigned int depth,unsigned int stride)170 void add_residual_block(const std::string &data_path, DataLayout weights_layout,
171 unsigned int unit, unsigned int depth, unsigned int stride)
172 {
173 PadStrideInfo dwc_info = PadStrideInfo(1, 1, 1, 1);
174 const unsigned int gconv_id = unit * 2;
175 const unsigned int num_groups = 4;
176 const std::string unit_id_name = arm_compute::support::cpp11::to_string(unit);
177 const std::string gconv_id_name = arm_compute::support::cpp11::to_string(gconv_id);
178 const std::string gconv_id_1_name = arm_compute::support::cpp11::to_string(gconv_id + 1);
179 const std::string unit_name = "unit" + unit_id_name;
180
181 SubStream left_ss(graph);
182 SubStream right_ss(graph);
183
184 if(stride == 2)
185 {
186 right_ss << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 3, common_params.data_layout, PadStrideInfo(2, 2, 1, 1))).set_name(unit_name + "/pool_1/AveragePool");
187 dwc_info = PadStrideInfo(2, 2, 1, 1);
188 }
189
190 left_ss << ConvolutionLayer(
191 1U, 1U, depth,
192 get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_w_0.npy", weights_layout),
193 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
194 PadStrideInfo(1, 1, 0, 0), num_groups)
195 .set_name(unit_name + "/gconv1_" + gconv_id_name + "/convolution")
196 << BatchNormalizationLayer(
197 get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_rm_0.npy"),
198 get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_riv_0.npy"),
199 get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_s_0.npy"),
200 get_weights_accessor(data_path, "gconv1_" + gconv_id_name + "_bn_b_0.npy"),
201 1e-5f)
202 .set_name(unit_name + "/gconv1_" + gconv_id_name + "/BatchNorm")
203 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "/gconv1_" + gconv_id_name + "/Relu")
204 << ChannelShuffleLayer(num_groups).set_name(unit_name + "/shuffle_0/ChannelShufle")
205 << DepthwiseConvolutionLayer(
206 3U, 3U,
207 get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_w_0.npy", weights_layout),
208 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
209 dwc_info)
210 .set_name(unit_name + "/gconv3_" + unit_id_name + "/depthwise")
211 << BatchNormalizationLayer(
212 get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_rm_0.npy"),
213 get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_riv_0.npy"),
214 get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_s_0.npy"),
215 get_weights_accessor(data_path, "gconv3_" + unit_id_name + "_bn_b_0.npy"),
216 1e-5f)
217 .set_name(unit_name + "/gconv3_" + unit_id_name + "/BatchNorm")
218 << ConvolutionLayer(
219 1U, 1U, depth,
220 get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_w_0.npy", weights_layout),
221 std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
222 PadStrideInfo(1, 1, 0, 0), num_groups)
223 .set_name(unit_name + "/gconv1_" + gconv_id_1_name + "/convolution")
224 << BatchNormalizationLayer(
225 get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_rm_0.npy"),
226 get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_riv_0.npy"),
227 get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_s_0.npy"),
228 get_weights_accessor(data_path, "gconv1_" + gconv_id_1_name + "_bn_b_0.npy"),
229 1e-5f)
230 .set_name(unit_name + "/gconv1_" + gconv_id_1_name + "/BatchNorm");
231
232 if(stride == 2)
233 {
234 graph << ConcatLayer(std::move(left_ss), std::move(right_ss)).set_name(unit_name + "/Concat");
235 }
236 else
237 {
238 graph << EltwiseLayer(std::move(left_ss), std::move(right_ss), EltwiseOperation::Add).set_name(unit_name + "/Add");
239 }
240 graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "/Relu");
241 }
242 };
243
244 /** Main program for ShuffleNet
245 *
246 * Model is based on:
247 * https://arxiv.org/abs/1707.01083
248 * "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices"
249 * Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, Jian Sun
250 *
251 * Provenance: https://s3.amazonaws.com/download.onnx/models/opset_9/shufflenet.tar.gz
252 *
253 * @note To list all the possible arguments execute the binary appended with the --help option
254 *
255 * @param[in] argc Number of arguments
256 * @param[in] argv Arguments
257 */
main(int argc,char ** argv)258 int main(int argc, char **argv)
259 {
260 return arm_compute::utils::run_example<ShuffleNetExample>(argc, argv);
261 }
262