Compute Library
 21.02
graph_resnet_v2_50.cpp
Go to the documentation of this file.
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"
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 ResNetV2_50 network using the Compute Library's graph API */
35 class GraphResNetV2_50Example : public Example
36 {
37 public:
38  GraphResNetV2_50Example()
39  : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ResNetV2_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  std::string model_path = "/cnn_data/resnet_v2_50_model/";
64  if(!data_path.empty())
65  {
66  data_path += model_path;
67  }
68 
69  // Create a preprocessor object
70  std::unique_ptr<IPreprocessor> preprocessor = std::make_unique<TFPreproccessor>();
71 
72  // Create input descriptor
73  const auto operation_layout = common_params.data_layout;
74  const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, 1U), DataLayout::NCHW, operation_layout);
75  TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout);
76 
77  // Set weights trained layout
78  const DataLayout weights_layout = DataLayout::NCHW;
79 
80  graph << common_params.target
81  << common_params.fast_math_hint
82  << InputLayer(input_descriptor, get_input_accessor(common_params, std::move(preprocessor), false /* Do not convert to BGR */))
84  7U, 7U, 64U,
85  get_weights_accessor(data_path, "conv1_weights.npy", weights_layout),
86  get_weights_accessor(data_path, "conv1_biases.npy", weights_layout),
87  PadStrideInfo(2, 2, 3, 3))
88  .set_name("conv1/convolution")
89  << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))).set_name("pool1/MaxPool");
90 
91  add_residual_block(data_path, "block1", weights_layout, 64, 3, 2);
92  add_residual_block(data_path, "block2", weights_layout, 128, 4, 2);
93  add_residual_block(data_path, "block3", weights_layout, 256, 6, 2);
94  add_residual_block(data_path, "block4", weights_layout, 512, 3, 1);
95 
96  graph << BatchNormalizationLayer(
97  get_weights_accessor(data_path, "postnorm_moving_mean.npy"),
98  get_weights_accessor(data_path, "postnorm_moving_variance.npy"),
99  get_weights_accessor(data_path, "postnorm_gamma.npy"),
100  get_weights_accessor(data_path, "postnorm_beta.npy"),
101  0.000009999999747378752f)
102  .set_name("postnorm/BatchNorm")
103  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("postnorm/Relu")
104  << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("pool5")
105  << ConvolutionLayer(
106  1U, 1U, 1001U,
107  get_weights_accessor(data_path, "logits_weights.npy", weights_layout),
108  get_weights_accessor(data_path, "logits_biases.npy"),
109  PadStrideInfo(1, 1, 0, 0))
110  .set_name("logits/convolution")
111  << FlattenLayer().set_name("predictions/Reshape")
112  << SoftmaxLayer().set_name("predictions/Softmax")
113  << OutputLayer(get_output_accessor(common_params, 5));
114 
115  // Finalize graph
116  GraphConfig config;
117  config.num_threads = common_params.threads;
118  config.use_tuner = common_params.enable_tuner;
119  config.tuner_mode = common_params.tuner_mode;
120  config.tuner_file = common_params.tuner_file;
121  config.mlgo_file = common_params.mlgo_file;
122  config.convert_to_uint8 = (common_params.data_type == DataType::QASYMM8);
123 
124  graph.finalize(common_params.target, config);
125 
126  return true;
127  }
128 
129  void do_run() override
130  {
131  // Run graph
132  graph.run();
133  }
134 
135 private:
136  CommandLineParser cmd_parser;
137  CommonGraphOptions common_opts;
138  CommonGraphParams common_params;
139  Stream graph;
140 
141  void add_residual_block(const std::string &data_path, const std::string &name, DataLayout weights_layout,
142  unsigned int base_depth, unsigned int num_units, unsigned int stride)
143  {
144  for(unsigned int i = 0; i < num_units; ++i)
145  {
146  // Generate unit names
147  std::stringstream unit_path_ss;
148  unit_path_ss << name << "_unit_" << (i + 1) << "_bottleneck_v2_";
149  std::stringstream unit_name_ss;
150  unit_name_ss << name << "/unit" << (i + 1) << "/bottleneck_v2/";
151 
152  std::string unit_path = unit_path_ss.str();
153  std::string unit_name = unit_name_ss.str();
154 
155  const TensorShape last_shape = graph.graph().node(graph.tail_node())->output(0)->desc().shape;
156  unsigned int depth_in = last_shape[arm_compute::get_data_layout_dimension_index(common_params.data_layout, DataLayoutDimension::CHANNEL)];
157  unsigned int depth_out = base_depth * 4;
158 
159  // All units have stride 1 apart from last one
160  unsigned int middle_stride = (i == (num_units - 1)) ? stride : 1;
161 
162  // Preact
163  SubStream preact(graph);
164  preact << BatchNormalizationLayer(
165  get_weights_accessor(data_path, unit_path + "preact_moving_mean.npy"),
166  get_weights_accessor(data_path, unit_path + "preact_moving_variance.npy"),
167  get_weights_accessor(data_path, unit_path + "preact_gamma.npy"),
168  get_weights_accessor(data_path, unit_path + "preact_beta.npy"),
169  0.000009999999747378752f)
170  .set_name(unit_name + "preact/BatchNorm")
171  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "preact/Relu");
172 
173  // Create bottleneck path
174  SubStream shortcut(graph);
175  if(depth_in == depth_out)
176  {
177  if(middle_stride != 1)
178  {
179  shortcut << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 1, common_params.data_layout, PadStrideInfo(middle_stride, middle_stride, 0, 0), true)).set_name(unit_name + "shortcut/MaxPool");
180  }
181  }
182  else
183  {
184  shortcut.forward_tail(preact.tail_node());
185  shortcut << ConvolutionLayer(
186  1U, 1U, depth_out,
187  get_weights_accessor(data_path, unit_path + "shortcut_weights.npy", weights_layout),
188  get_weights_accessor(data_path, unit_path + "shortcut_biases.npy", weights_layout),
189  PadStrideInfo(1, 1, 0, 0))
190  .set_name(unit_name + "shortcut/convolution");
191  }
192 
193  // Create residual path
194  SubStream residual(preact);
195  residual << ConvolutionLayer(
196  1U, 1U, base_depth,
197  get_weights_accessor(data_path, unit_path + "conv1_weights.npy", weights_layout),
198  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
199  PadStrideInfo(1, 1, 0, 0))
200  .set_name(unit_name + "conv1/convolution")
202  get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_mean.npy"),
203  get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_moving_variance.npy"),
204  get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_gamma.npy"),
205  get_weights_accessor(data_path, unit_path + "conv1_BatchNorm_beta.npy"),
206  0.000009999999747378752f)
207  .set_name(unit_name + "conv1/BatchNorm")
208  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu")
209  << ConvolutionLayer(
210  3U, 3U, base_depth,
211  get_weights_accessor(data_path, unit_path + "conv2_weights.npy", weights_layout),
212  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
213  PadStrideInfo(middle_stride, middle_stride, 1, 1))
214  .set_name(unit_name + "conv2/convolution")
216  get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_mean.npy"),
217  get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_moving_variance.npy"),
218  get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_gamma.npy"),
219  get_weights_accessor(data_path, unit_path + "conv2_BatchNorm_beta.npy"),
220  0.000009999999747378752f)
221  .set_name(unit_name + "conv2/BatchNorm")
222  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu")
223  << ConvolutionLayer(
224  1U, 1U, depth_out,
225  get_weights_accessor(data_path, unit_path + "conv3_weights.npy", weights_layout),
226  get_weights_accessor(data_path, unit_path + "conv3_biases.npy", weights_layout),
227  PadStrideInfo(1, 1, 0, 0))
228  .set_name(unit_name + "conv3/convolution");
229 
230  graph << EltwiseLayer(std::move(shortcut), std::move(residual), EltwiseOperation::Add).set_name(unit_name + "add");
231  }
232  }
233 };
234 
235 /** Main program for ResNetV2_50
236  *
237  * Model is based on:
238  * https://arxiv.org/abs/1603.05027
239  * "Identity Mappings in Deep Residual Networks"
240  * Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
241  *
242  * Provenance: download.tensorflow.org/models/resnet_v2_50_2017_04_14.tar.gz
243  *
244  * @note To list all the possible arguments execute the binary appended with the --help option
245  *
246  * @param[in] argc Number of arguments
247  * @param[in] argv Arguments
248  */
249 int main(int argc, char **argv)
250 {
251  return arm_compute::utils::run_example<GraphResNetV2_50Example>(argc, argv);
252 }
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.
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
NodeID tail_node()
Returns the tail node of the Stream.
Definition: IStream.h:65
quantized, asymmetric fixed-point 8-bit number unsigned
int main(int argc, char **argv)
Main program for ResNetV2_50.
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.
const INode * node(NodeID id) const
Get node object given its id.
Definition: Graph.cpp:204
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
Graph & graph() override
Returns the underlying graph.
Definition: Stream.cpp:63
Stream frontend class to construct simple graphs in a stream fashion.
Definition: Stream.h:45
size_t get_data_layout_dimension_index(const DataLayout data_layout, const DataLayoutDimension data_layout_dimension)
Get the index of the given dimension.
Definition: Helpers.inl:193
DataLayout
[DataLayout enum definition]
Definition: Types.h:120
ILayer & set_name(std::string name)
Sets the name of the layer.
Definition: ILayer.h:55