Compute Library
 21.02
graph_resnet50.cpp
Go to the documentation of this file.
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"
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:
38  GraphResNetV1_50Example()
39  : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ResNetV1_50")
40  {
41  }
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, 1U), 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 */))
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")
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.convert_to_uint8 = (common_params.data_type == DataType::QASYMM8);
120 
121  graph.finalize(common_params.target, config);
122 
123  return true;
124  }
125 
126  void do_run() override
127  {
128  // Run graph
129  graph.run();
130  }
131 
132 private:
133  CommandLineParser cmd_parser;
134  CommonGraphOptions common_opts;
135  CommonGraphParams common_params;
136  Stream graph;
137 
138  void add_residual_block(const std::string &data_path, const std::string &name, DataLayout weights_layout,
139  unsigned int base_depth, unsigned int num_units, unsigned int stride)
140  {
141  for(unsigned int i = 0; i < num_units; ++i)
142  {
143  std::stringstream unit_path_ss;
144  unit_path_ss << "/cnn_data/resnet50_model/" << name << "_unit_" << (i + 1) << "_bottleneck_v1_";
145  std::stringstream unit_name_ss;
146  unit_name_ss << name << "/unit" << (i + 1) << "/bottleneck_v1/";
147 
148  std::string unit_path = unit_path_ss.str();
149  std::string unit_name = unit_name_ss.str();
150 
151  unsigned int middle_stride = 1;
152 
153  if(i == (num_units - 1))
154  {
155  middle_stride = stride;
156  }
157 
158  SubStream right(graph);
159  right << ConvolutionLayer(
160  1U, 1U, base_depth,
161  get_weights_accessor(data_path, unit_path + "conv1_weights.npy", weights_layout),
162  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
163  PadStrideInfo(1, 1, 0, 0))
164  .set_name(unit_name + "conv1/convolution")
166  get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_mean.npy"),
167  get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_variance.npy"),
168  get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_gamma.npy"),
169  get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_beta.npy"),
170  0.0000100099996416f)
171  .set_name(unit_name + "conv1/BatchNorm")
172  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu")
173 
174  << ConvolutionLayer(
175  3U, 3U, base_depth,
176  get_weights_accessor(data_path, unit_path + "conv2_weights.npy", weights_layout),
177  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
178  PadStrideInfo(middle_stride, middle_stride, 1, 1))
179  .set_name(unit_name + "conv2/convolution")
181  get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_mean.npy"),
182  get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_variance.npy"),
183  get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_gamma.npy"),
184  get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_beta.npy"),
185  0.0000100099996416f)
186  .set_name(unit_name + "conv2/BatchNorm")
187  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu")
188 
189  << ConvolutionLayer(
190  1U, 1U, base_depth * 4,
191  get_weights_accessor(data_path, unit_path + "conv3_weights.npy", weights_layout),
192  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
193  PadStrideInfo(1, 1, 0, 0))
194  .set_name(unit_name + "conv3/convolution")
196  get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_moving_mean.npy"),
197  get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_moving_variance.npy"),
198  get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_gamma.npy"),
199  get_weights_accessor(data_path, unit_path + "conv3_BatchNorm_beta.npy"),
200  0.0000100099996416f)
201  .set_name(unit_name + "conv2/BatchNorm");
202 
203  if(i == 0)
204  {
205  SubStream left(graph);
206  left << ConvolutionLayer(
207  1U, 1U, base_depth * 4,
208  get_weights_accessor(data_path, unit_path + "shortcut_weights.npy", weights_layout),
209  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
210  PadStrideInfo(1, 1, 0, 0))
211  .set_name(unit_name + "shortcut/convolution")
213  get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_moving_mean.npy"),
214  get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_moving_variance.npy"),
215  get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_gamma.npy"),
216  get_weights_accessor(data_path, unit_path + "shortcut_BatchNorm_beta.npy"),
217  0.0000100099996416f)
218  .set_name(unit_name + "shortcut/BatchNorm");
219 
220  graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(unit_name + "add");
221  }
222  else if(middle_stride > 1)
223  {
224  SubStream left(graph);
225  left << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 1, common_params.data_layout, PadStrideInfo(middle_stride, middle_stride, 0, 0), true)).set_name(unit_name + "shortcut/MaxPool");
226 
227  graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(unit_name + "add");
228  }
229  else
230  {
231  SubStream left(graph);
232  graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(unit_name + "add");
233  }
234 
235  graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu");
236  }
237  }
238 };
239 
240 /** Main program for ResNetV1_50
241  *
242  * Model is based on:
243  * https://arxiv.org/abs/1512.03385
244  * "Deep Residual Learning for Image Recognition"
245  * Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
246  *
247  * Provenance: download.tensorflow.org/models/resnet_v1_50_2016_08_28.tar.gz
248  *
249  * @note To list all the possible arguments execute the binary appended with the --help option
250  *
251  * @param[in] argc Number of arguments
252  * @param[in] argv Arguments
253  */
254 int main(int argc, char **argv)
255 {
256  return arm_compute::utils::run_example<GraphResNetV1_50Example>(argc, argv);
257 }
Graph configuration structure Device target types.
Definition: Types.h:80
CLTunerMode tuner_mode
Tuner mode to be used by the CL tuner.
Definition: Types.h:87
std::unique_ptr< graph::ITensorAccessor > get_input_accessor(const arm_compute::utils::CommonGraphParams &graph_parameters, std::unique_ptr< IPreprocessor > preprocessor=nullptr, bool bgr=true)
Generates appropriate input accessor according to the specified graph parameters. ...
Definition: GraphUtils.h:497
bool convert_to_uint8
Convert graph to a synthetic uint8 graph.
Definition: Types.h:86
void consume_common_graph_parameters(CommonGraphValidateOptions &options, CommonParams &common_params)
Consumes the consume_common_graph_parameters graph options and creates a structure containing any inf...
Includes all the Graph headers at once.
int main(int argc, char **argv)
Main program for ResNetV1_50.
Common command line options used to configure the graph examples.
Class to parse command line arguments.
std::string mlgo_file
Filename to load MLGO heuristics from.
Definition: Types.h:90
std::string tuner_file
File to load/store tuning values from.
Definition: Types.h:89
quantized, asymmetric fixed-point 8-bit number unsigned
Abstract Example class.
Definition: Utils.h:78
Num samples, channels, height, width.
const char * name
TensorShape permute_shape(TensorShape tensor_shape, DataLayout in_data_layout, DataLayout out_data_layout)
Permutes a given tensor shape given the input and output data layout.
Definition: GraphUtils.h:664
TensorDescriptor & set_layout(DataLayout data_layout)
Sets tensor descriptor data layout.
Structure holding all the common graph parameters.
std::unique_ptr< graph::ITensorAccessor > get_output_accessor(const arm_compute::utils::CommonGraphParams &graph_parameters, size_t top_n=5, bool is_validation=false, std::ostream &output_stream=std::cout)
Generates appropriate output accessor according to the specified graph parameters.
Definition: GraphUtils.h:543
bool use_tuner
Use a tuner in tunable backends.
Definition: Types.h:85
std::unique_ptr< graph::ITensorAccessor > get_weights_accessor(const std::string &path, const std::string &data_file, DataLayout file_layout=DataLayout::NCHW)
Generates appropriate weights accessor according to the specified path.
Definition: GraphUtils.h:475
int num_threads
Number of threads to use (thread capable backends), if 0 the backend will auto-initialize, if -1 the backend will stay as it is.
Definition: Types.h:88
Stream frontend class to construct simple graphs in a stream fashion.
Definition: Stream.h:45
DataLayout
[DataLayout enum definition]
Definition: Types.h:120
ILayer & set_name(std::string name)
Sets the name of the layer.
Definition: ILayer.h:55