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 #ifdef ARM_COMPUTE_CL
26 #include "arm_compute/runtime/CL/Utils.h"
27 #endif /* ARM_COMPUTE_CL */
28 #include "support/ToolchainSupport.h"
29 #include "utils/CommonGraphOptions.h"
30 #include "utils/GraphUtils.h"
31 #include "utils/Utils.h"
32
33 using namespace arm_compute;
34 using namespace arm_compute::utils;
35 using namespace arm_compute::graph::frontend;
36 using namespace arm_compute::graph_utils;
37
38 /** Example demonstrating how to implement AlexNet's network using the Compute Library's graph API */
39 class GraphAlexnetExample : public Example
40 {
41 public:
GraphAlexnetExample()42 GraphAlexnetExample()
43 : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "AlexNet")
44 {
45 }
do_setup(int argc,char ** argv)46 bool do_setup(int argc, char **argv) override
47 {
48 // Parse arguments
49 cmd_parser.parse(argc, argv);
50 cmd_parser.validate();
51
52 // Consume common parameters
53 common_params = consume_common_graph_parameters(common_opts);
54
55 // Return when help menu is requested
56 if(common_params.help)
57 {
58 cmd_parser.print_help(argv[0]);
59 return false;
60 }
61
62 // Checks
63 ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph");
64
65 // Print parameter values
66 std::cout << common_params << std::endl;
67
68 // Get trainable parameters data path
69 std::string data_path = common_params.data_path;
70
71 // Create a preprocessor object
72 const std::array<float, 3> mean_rgb{ { 122.68f, 116.67f, 104.01f } };
73 std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<CaffePreproccessor>(mean_rgb);
74
75 // Create input descriptor
76 const auto operation_layout = common_params.data_layout;
77 const TensorShape tensor_shape = permute_shape(TensorShape(227U, 227U, 3U, common_params.batches), DataLayout::NCHW, operation_layout);
78 TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout);
79
80 // Set weights trained layout
81 const DataLayout weights_layout = DataLayout::NCHW;
82
83 graph << common_params.target
84 << common_params.fast_math_hint
85 << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor)))
86 // Layer 1
87 << ConvolutionLayer(
88 11U, 11U, 96U,
89 get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv1_w.npy", weights_layout),
90 get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv1_b.npy"),
91 PadStrideInfo(4, 4, 0, 0))
92 .set_name("conv1")
93 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu1")
94 << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)).set_name("norm1")
95 << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool1")
96 // Layer 2
97 << ConvolutionLayer(
98 5U, 5U, 256U,
99 get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv2_w.npy", weights_layout),
100 get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv2_b.npy"),
101 PadStrideInfo(1, 1, 2, 2), 2)
102 .set_name("conv2")
103 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu2")
104 << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f)).set_name("norm2")
105 << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool2")
106 // Layer 3
107 << ConvolutionLayer(
108 3U, 3U, 384U,
109 get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv3_w.npy", weights_layout),
110 get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv3_b.npy"),
111 PadStrideInfo(1, 1, 1, 1))
112 .set_name("conv3")
113 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu3")
114 // Layer 4
115 << ConvolutionLayer(
116 3U, 3U, 384U,
117 get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv4_w.npy", weights_layout),
118 get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv4_b.npy"),
119 PadStrideInfo(1, 1, 1, 1), 2)
120 .set_name("conv4")
121 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu4")
122 // Layer 5
123 << ConvolutionLayer(
124 3U, 3U, 256U,
125 get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv5_w.npy", weights_layout),
126 get_weights_accessor(data_path, "/cnn_data/alexnet_model/conv5_b.npy"),
127 PadStrideInfo(1, 1, 1, 1), 2)
128 .set_name("conv5")
129 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu5")
130 << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool5")
131 // Layer 6
132 << FullyConnectedLayer(
133 4096U,
134 get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc6_w.npy", weights_layout),
135 get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc6_b.npy"))
136 .set_name("fc6")
137 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu6")
138 // Layer 7
139 << FullyConnectedLayer(
140 4096U,
141 get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc7_w.npy", weights_layout),
142 get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc7_b.npy"))
143 .set_name("fc7")
144 << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu7")
145 // Layer 8
146 << FullyConnectedLayer(
147 1000U,
148 get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc8_w.npy", weights_layout),
149 get_weights_accessor(data_path, "/cnn_data/alexnet_model/fc8_b.npy"))
150 .set_name("fc8")
151 // Softmax
152 << SoftmaxLayer().set_name("prob")
153 << OutputLayer(get_output_accessor(common_params, 5));
154
155 // Finalize graph
156 GraphConfig config;
157
158 config.num_threads = common_params.threads;
159 config.use_tuner = common_params.enable_tuner;
160 config.tuner_mode = common_params.tuner_mode;
161 config.tuner_file = common_params.tuner_file;
162 config.mlgo_file = common_params.mlgo_file;
163
164 // Load the precompiled kernels from a file into the kernel library, in this way the next time they are needed
165 // compilation won't be required.
166 if(common_params.enable_cl_cache)
167 {
168 #ifdef ARM_COMPUTE_CL
169 restore_program_cache_from_file();
170 #endif /* ARM_COMPUTE_CL */
171 }
172
173 graph.finalize(common_params.target, config);
174
175 // Save the opencl kernels to a file
176 if(common_opts.enable_cl_cache)
177 {
178 #ifdef ARM_COMPUTE_CL
179 save_program_cache_to_file();
180 #endif /* ARM_COMPUTE_CL */
181 }
182
183 return true;
184 }
do_run()185 void do_run() override
186 {
187 // Run graph
188 graph.run();
189 }
190
191 private:
192 CommandLineParser cmd_parser;
193 CommonGraphOptions common_opts;
194 CommonGraphParams common_params;
195 Stream graph;
196 };
197
198 /** Main program for AlexNet
199 *
200 * Model is based on:
201 * https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks
202 * "ImageNet Classification with Deep Convolutional Neural Networks"
203 * Alex Krizhevsky and Sutskever, Ilya and Hinton, Geoffrey E
204 *
205 * Provenance: https://github.com/BVLC/caffe/tree/master/models/bvlc_alexnet
206 *
207 * @note To list all the possible arguments execute the binary appended with the --help option
208 *
209 * @param[in] argc Number of arguments
210 * @param[in] argv Arguments
211 *
212 * @return Return code
213 */
main(int argc,char ** argv)214 int main(int argc, char **argv)
215 {
216 return arm_compute::utils::run_example<GraphAlexnetExample>(argc, argv);
217 }
218