xref: /aosp_15_r20/external/ComputeLibrary/examples/graph_googlenet.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
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 Googlenet's network using the Compute Library's graph API */
35 class GraphGooglenetExample : public Example
36 {
37 public:
GraphGooglenetExample()38     GraphGooglenetExample()
39         : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "GoogleNet")
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         // Checks
59         ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph");
60 
61         // Print parameter values
62         std::cout << common_params << std::endl;
63 
64         // Get trainable parameters data path
65         std::string data_path = common_params.data_path;
66 
67         // Create a preprocessor object
68         const std::array<float, 3> mean_rgb{ { 122.68f, 116.67f, 104.01f } };
69         std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<CaffePreproccessor>(mean_rgb);
70 
71         // Create input descriptor
72         const auto        operation_layout = common_params.data_layout;
73         const TensorShape tensor_shape     = permute_shape(TensorShape(224U, 224U, 3U, common_params.batches), DataLayout::NCHW, operation_layout);
74         TensorDescriptor  input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout);
75 
76         // Set weights trained layout
77         const DataLayout weights_layout = DataLayout::NCHW;
78 
79         graph << common_params.target
80               << common_params.fast_math_hint
81               << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor)))
82               << ConvolutionLayer(
83                   7U, 7U, 64U,
84                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_w.npy", weights_layout),
85                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_b.npy"),
86                   PadStrideInfo(2, 2, 3, 3))
87               .set_name("conv1/7x7_s2")
88               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1/relu_7x7")
89               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool1/3x3_s2")
90               << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)).set_name("pool1/norm1")
91               << ConvolutionLayer(
92                   1U, 1U, 64U,
93                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_w.npy", weights_layout),
94                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_b.npy"),
95                   PadStrideInfo(1, 1, 0, 0))
96               .set_name("conv2/3x3_reduce")
97               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2/relu_3x3_reduce")
98               << ConvolutionLayer(
99                   3U, 3U, 192U,
100                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_w.npy", weights_layout),
101                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_b.npy"),
102                   PadStrideInfo(1, 1, 1, 1))
103               .set_name("conv2/3x3")
104               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2/relu_3x3")
105               << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)).set_name("conv2/norm2")
106               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool2/3x3_s2");
107         graph << get_inception_node(data_path, "inception_3a", weights_layout, 64, std::make_tuple(96U, 128U), std::make_tuple(16U, 32U), 32U).set_name("inception_3a/concat");
108         graph << get_inception_node(data_path, "inception_3b", weights_layout, 128, std::make_tuple(128U, 192U), std::make_tuple(32U, 96U), 64U).set_name("inception_3b/concat");
109         graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool3/3x3_s2");
110         graph << get_inception_node(data_path, "inception_4a", weights_layout, 192, std::make_tuple(96U, 208U), std::make_tuple(16U, 48U), 64U).set_name("inception_4a/concat");
111         graph << get_inception_node(data_path, "inception_4b", weights_layout, 160, std::make_tuple(112U, 224U), std::make_tuple(24U, 64U), 64U).set_name("inception_4b/concat");
112         graph << get_inception_node(data_path, "inception_4c", weights_layout, 128, std::make_tuple(128U, 256U), std::make_tuple(24U, 64U), 64U).set_name("inception_4c/concat");
113         graph << get_inception_node(data_path, "inception_4d", weights_layout, 112, std::make_tuple(144U, 288U), std::make_tuple(32U, 64U), 64U).set_name("inception_4d/concat");
114         graph << get_inception_node(data_path, "inception_4e", weights_layout, 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U).set_name("inception_4e/concat");
115         graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL))).set_name("pool4/3x3_s2");
116         graph << get_inception_node(data_path, "inception_5a", weights_layout, 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U).set_name("inception_5a/concat");
117         graph << get_inception_node(data_path, "inception_5b", weights_layout, 384, std::make_tuple(192U, 384U), std::make_tuple(48U, 128U), 128U).set_name("inception_5b/concat");
118         graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 7, operation_layout, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL))).set_name("pool5/7x7_s1")
119               << FullyConnectedLayer(
120                   1000U,
121                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_w.npy", weights_layout),
122                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_b.npy"))
123               .set_name("loss3/classifier")
124               << SoftmaxLayer().set_name("prob")
125               << OutputLayer(get_output_accessor(common_params, 5));
126 
127         // Finalize graph
128         GraphConfig config;
129         config.num_threads = common_params.threads;
130         config.use_tuner   = common_params.enable_tuner;
131         config.tuner_mode  = common_params.tuner_mode;
132         config.tuner_file  = common_params.tuner_file;
133         config.mlgo_file   = common_params.mlgo_file;
134 
135         graph.finalize(common_params.target, config);
136 
137         return true;
138     }
do_run()139     void do_run() override
140     {
141         // Run graph
142         graph.run();
143     }
144 
145 private:
146     CommandLineParser  cmd_parser;
147     CommonGraphOptions common_opts;
148     CommonGraphParams  common_params;
149     Stream             graph;
150 
get_inception_node(const std::string & data_path,std::string && param_path,DataLayout weights_layout,unsigned int a_filt,std::tuple<unsigned int,unsigned int> b_filters,std::tuple<unsigned int,unsigned int> c_filters,unsigned int d_filt)151     ConcatLayer get_inception_node(const std::string &data_path, std::string &&param_path, DataLayout weights_layout,
152                                    unsigned int a_filt,
153                                    std::tuple<unsigned int, unsigned int> b_filters,
154                                    std::tuple<unsigned int, unsigned int> c_filters,
155                                    unsigned int d_filt)
156     {
157         std::string total_path = "/cnn_data/googlenet_model/" + param_path + "/" + param_path + "_";
158         SubStream   i_a(graph);
159         i_a << ConvolutionLayer(
160                 1U, 1U, a_filt,
161                 get_weights_accessor(data_path, total_path + "1x1_w.npy", weights_layout),
162                 get_weights_accessor(data_path, total_path + "1x1_b.npy"),
163                 PadStrideInfo(1, 1, 0, 0))
164             .set_name(param_path + "/1x1")
165             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_1x1");
166 
167         SubStream i_b(graph);
168         i_b << ConvolutionLayer(
169                 1U, 1U, std::get<0>(b_filters),
170                 get_weights_accessor(data_path, total_path + "3x3_reduce_w.npy", weights_layout),
171                 get_weights_accessor(data_path, total_path + "3x3_reduce_b.npy"),
172                 PadStrideInfo(1, 1, 0, 0))
173             .set_name(param_path + "/3x3_reduce")
174             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_3x3_reduce")
175             << ConvolutionLayer(
176                 3U, 3U, std::get<1>(b_filters),
177                 get_weights_accessor(data_path, total_path + "3x3_w.npy", weights_layout),
178                 get_weights_accessor(data_path, total_path + "3x3_b.npy"),
179                 PadStrideInfo(1, 1, 1, 1))
180             .set_name(param_path + "/3x3")
181             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_3x3");
182 
183         SubStream i_c(graph);
184         i_c << ConvolutionLayer(
185                 1U, 1U, std::get<0>(c_filters),
186                 get_weights_accessor(data_path, total_path + "5x5_reduce_w.npy", weights_layout),
187                 get_weights_accessor(data_path, total_path + "5x5_reduce_b.npy"),
188                 PadStrideInfo(1, 1, 0, 0))
189             .set_name(param_path + "/5x5_reduce")
190             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_5x5_reduce")
191             << ConvolutionLayer(
192                 5U, 5U, std::get<1>(c_filters),
193                 get_weights_accessor(data_path, total_path + "5x5_w.npy", weights_layout),
194                 get_weights_accessor(data_path, total_path + "5x5_b.npy"),
195                 PadStrideInfo(1, 1, 2, 2))
196             .set_name(param_path + "/5x5")
197             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_5x5");
198 
199         SubStream i_d(graph);
200         i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, common_params.data_layout, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL))).set_name(param_path + "/pool")
201             << ConvolutionLayer(
202                 1U, 1U, d_filt,
203                 get_weights_accessor(data_path, total_path + "pool_proj_w.npy", weights_layout),
204                 get_weights_accessor(data_path, total_path + "pool_proj_b.npy"),
205                 PadStrideInfo(1, 1, 0, 0))
206             .set_name(param_path + "/pool_proj")
207             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(param_path + "/relu_pool_proj");
208 
209         return ConcatLayer(std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d));
210     }
211 };
212 
213 /** Main program for Googlenet
214  *
215  * Model is based on:
216  *      https://arxiv.org/abs/1409.4842
217  *      "Going deeper with convolutions"
218  *      Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, Andrew Rabinovich
219  *
220  * Provenance: https://github.com/BVLC/caffe/tree/master/models/bvlc_googlenet
221  *
222  * @note To list all the possible arguments execute the binary appended with the --help option
223  *
224  * @param[in] argc Number of arguments
225  * @param[in] argv Arguments
226  */
main(int argc,char ** argv)227 int main(int argc, char **argv)
228 {
229     return arm_compute::utils::run_example<GraphGooglenetExample>(argc, argv);
230 }
231