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 Squeezenet's v1.1 network using the Compute Library's graph API */
35 class GraphSqueezenet_v1_1Example : public Example
36 {
37 public:
GraphSqueezenet_v1_1Example()38 GraphSqueezenet_v1_1Example()
39 : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "SqueezeNetV1.1")
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
68 // Create input descriptor
69 const auto operation_layout = common_params.data_layout;
70 const TensorShape tensor_shape = permute_shape(TensorShape(227U, 227U, 3U, common_params.batches), DataLayout::NCHW, operation_layout);
71 TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout);
72
73 // Set weights trained layout
74 const DataLayout weights_layout = DataLayout::NCHW;
75
76 graph << common_params.target
77 << common_params.fast_math_hint
78 << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor)))
79 << ConvolutionLayer(
80 3U, 3U, 64U,
81 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv1_w.npy", weights_layout),
82 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv1_b.npy"),
83 PadStrideInfo(2, 2, 0, 0))
84 .set_name("conv1")
85 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu_conv1")
86 << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool1")
87 << ConvolutionLayer(
88 1U, 1U, 16U,
89 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire2_squeeze1x1_w.npy", weights_layout),
90 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire2_squeeze1x1_b.npy"),
91 PadStrideInfo(1, 1, 0, 0))
92 .set_name("fire2/squeeze1x1")
93 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire2/relu_squeeze1x1");
94 graph << get_expand_fire_node(data_path, "fire2", weights_layout, 64U, 64U).set_name("fire2/concat");
95 graph << ConvolutionLayer(
96 1U, 1U, 16U,
97 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire3_squeeze1x1_w.npy", weights_layout),
98 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire3_squeeze1x1_b.npy"),
99 PadStrideInfo(1, 1, 0, 0))
100 .set_name("fire3/squeeze1x1")
101 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire3/relu_squeeze1x1");
102 graph << get_expand_fire_node(data_path, "fire3", weights_layout, 64U, 64U).set_name("fire3/concat");
103 graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool3")
104 << ConvolutionLayer(
105 1U, 1U, 32U,
106 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire4_squeeze1x1_w.npy", weights_layout),
107 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire4_squeeze1x1_b.npy"),
108 PadStrideInfo(1, 1, 0, 0))
109 .set_name("fire4/squeeze1x1")
110 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire4/relu_squeeze1x1");
111 graph << get_expand_fire_node(data_path, "fire4", weights_layout, 128U, 128U).set_name("fire4/concat");
112 graph << ConvolutionLayer(
113 1U, 1U, 32U,
114 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire5_squeeze1x1_w.npy", weights_layout),
115 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire5_squeeze1x1_b.npy"),
116 PadStrideInfo(1, 1, 0, 0))
117 .set_name("fire5/squeeze1x1")
118 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire5/relu_squeeze1x1");
119 graph << get_expand_fire_node(data_path, "fire5", weights_layout, 128U, 128U).set_name("fire5/concat");
120 graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool5")
121 << ConvolutionLayer(
122 1U, 1U, 48U,
123 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire6_squeeze1x1_w.npy", weights_layout),
124 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire6_squeeze1x1_b.npy"),
125 PadStrideInfo(1, 1, 0, 0))
126 .set_name("fire6/squeeze1x1")
127 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire6/relu_squeeze1x1");
128 graph << get_expand_fire_node(data_path, "fire6", weights_layout, 192U, 192U).set_name("fire6/concat");
129 graph << ConvolutionLayer(
130 1U, 1U, 48U,
131 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire7_squeeze1x1_w.npy", weights_layout),
132 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire7_squeeze1x1_b.npy"),
133 PadStrideInfo(1, 1, 0, 0))
134 .set_name("fire7/squeeze1x1")
135 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire7/relu_squeeze1x1");
136 graph << get_expand_fire_node(data_path, "fire7", weights_layout, 192U, 192U).set_name("fire7/concat");
137 graph << ConvolutionLayer(
138 1U, 1U, 64U,
139 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire8_squeeze1x1_w.npy", weights_layout),
140 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire8_squeeze1x1_b.npy"),
141 PadStrideInfo(1, 1, 0, 0))
142 .set_name("fire8/squeeze1x1")
143 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire8/relu_squeeze1x1");
144 graph << get_expand_fire_node(data_path, "fire8", weights_layout, 256U, 256U).set_name("fire8/concat");
145 graph << ConvolutionLayer(
146 1U, 1U, 64U,
147 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire9_squeeze1x1_w.npy", weights_layout),
148 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire9_squeeze1x1_b.npy"),
149 PadStrideInfo(1, 1, 0, 0))
150 .set_name("fire9/squeeze1x1")
151 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("fire9/relu_squeeze1x1");
152 graph << get_expand_fire_node(data_path, "fire9", weights_layout, 256U, 256U).set_name("fire9/concat");
153 graph << ConvolutionLayer(
154 1U, 1U, 1000U,
155 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv10_w.npy", weights_layout),
156 get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv10_b.npy"),
157 PadStrideInfo(1, 1, 0, 0))
158 .set_name("conv10")
159 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu_conv10")
160 << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("pool10")
161 << FlattenLayer().set_name("flatten")
162 << SoftmaxLayer().set_name("prob")
163 << OutputLayer(get_output_accessor(common_params, 5));
164
165 // Finalize graph
166 GraphConfig config;
167 config.num_threads = common_params.threads;
168 config.use_tuner = common_params.enable_tuner;
169 config.tuner_mode = common_params.tuner_mode;
170 config.tuner_file = common_params.tuner_file;
171 config.mlgo_file = common_params.mlgo_file;
172 config.use_synthetic_type = arm_compute::is_data_type_quantized(common_params.data_type);
173 config.synthetic_type = common_params.data_type;
174
175 graph.finalize(common_params.target, config);
176
177 return true;
178 }
do_run()179 void do_run() override
180 {
181 // Run graph
182 graph.run();
183 }
184
185 private:
186 CommandLineParser cmd_parser;
187 CommonGraphOptions common_opts;
188 CommonGraphParams common_params;
189 Stream graph;
190
get_expand_fire_node(const std::string & data_path,std::string && param_path,DataLayout weights_layout,unsigned int expand1_filt,unsigned int expand3_filt)191 ConcatLayer get_expand_fire_node(const std::string &data_path, std::string &¶m_path, DataLayout weights_layout,
192 unsigned int expand1_filt, unsigned int expand3_filt)
193 {
194 std::string total_path = "/cnn_data/squeezenet_v1_1_model/" + param_path + "_";
195 SubStream i_a(graph);
196 i_a << ConvolutionLayer(
197 1U, 1U, expand1_filt,
198 get_weights_accessor(data_path, total_path + "expand1x1_w.npy", weights_layout),
199 get_weights_accessor(data_path, total_path + "expand1x1_b.npy"),
200 PadStrideInfo(1, 1, 0, 0))
201 .set_name(param_path + "/expand1x1")
202 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_expand1x1");
203
204 SubStream i_b(graph);
205 i_b << ConvolutionLayer(
206 3U, 3U, expand3_filt,
207 get_weights_accessor(data_path, total_path + "expand3x3_w.npy", weights_layout),
208 get_weights_accessor(data_path, total_path + "expand3x3_b.npy"),
209 PadStrideInfo(1, 1, 1, 1))
210 .set_name(param_path + "/expand3x3")
211 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_expand3x3");
212
213 return ConcatLayer(std::move(i_a), std::move(i_b));
214 }
215 };
216
217 /** Main program for Squeezenet v1.1
218 *
219 * Model is based on:
220 * https://arxiv.org/abs/1602.07360
221 * "SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size"
222 * Forrest N. Iandola, Song Han, Matthew W. Moskewicz, Khalid Ashraf, William J. Dally, Kurt Keutzer
223 *
224 * Provenance: https://github.com/DeepScale/SqueezeNet/blob/master/SqueezeNet_v1.1/squeezenet_v1.1.caffemodel
225 *
226 * @note To list all the possible arguments execute the binary appended with the --help option
227 *
228 * @param[in] argc Number of arguments
229 * @param[in] argv Arguments
230 */
main(int argc,char ** argv)231 int main(int argc, char **argv)
232 {
233 return arm_compute::utils::run_example<GraphSqueezenet_v1_1Example>(argc, argv);
234 }
235