Compute Library
 21.02
graph_resnext50.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 ResNeXt50 network using the Compute Library's graph API */
35 class GraphResNeXt50Example : public Example
36 {
37 public:
38  GraphResNeXt50Example()
39  : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "ResNeXt50")
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  // 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 input descriptor
68  const auto operation_layout = common_params.data_layout;
69  const TensorShape tensor_shape = permute_shape(TensorShape(224U, 224U, 3U, 1U), DataLayout::NCHW, operation_layout);
70  TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout);
71 
72  // Set weights trained layout
73  const DataLayout weights_layout = DataLayout::NCHW;
74 
75  graph << common_params.target
76  << common_params.fast_math_hint
77  << InputLayer(input_descriptor, get_input_accessor(common_params))
78  << ScaleLayer(get_weights_accessor(data_path, "/cnn_data/resnext50_model/bn_data_mul.npy"),
79  get_weights_accessor(data_path, "/cnn_data/resnext50_model/bn_data_add.npy"))
80  .set_name("bn_data/Scale")
82  7U, 7U, 64U,
83  get_weights_accessor(data_path, "/cnn_data/resnext50_model/conv0_weights.npy", weights_layout),
84  get_weights_accessor(data_path, "/cnn_data/resnext50_model/conv0_biases.npy"),
85  PadStrideInfo(2, 2, 2, 3, 2, 3, DimensionRoundingType::FLOOR))
86  .set_name("conv0/Convolution")
87  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv0/Relu")
88  << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, operation_layout, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))).set_name("pool0");
89 
90  add_residual_block(data_path, weights_layout, /*ofm*/ 256, /*stage*/ 1, /*num_unit*/ 3, /*stride_conv_unit1*/ 1);
91  add_residual_block(data_path, weights_layout, 512, 2, 4, 2);
92  add_residual_block(data_path, weights_layout, 1024, 3, 6, 2);
93  add_residual_block(data_path, weights_layout, 2048, 4, 3, 2);
94 
95  graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, operation_layout)).set_name("pool1")
96  << FlattenLayer().set_name("predictions/Reshape")
97  << OutputLayer(get_npy_output_accessor(common_params.labels, TensorShape(2048U), DataType::F32));
98 
99  // Finalize graph
100  GraphConfig config;
101  config.num_threads = common_params.threads;
102  config.use_tuner = common_params.enable_tuner;
103  config.tuner_mode = common_params.tuner_mode;
104  config.tuner_file = common_params.tuner_file;
105  config.mlgo_file = common_params.mlgo_file;
106 
107  graph.finalize(common_params.target, config);
108 
109  return true;
110  }
111 
112  void do_run() override
113  {
114  // Run graph
115  graph.run();
116  }
117 
118 private:
119  CommandLineParser cmd_parser;
120  CommonGraphOptions common_opts;
121  CommonGraphParams common_params;
122  Stream graph;
123 
124  void add_residual_block(const std::string &data_path, DataLayout weights_layout,
125  unsigned int base_depth, unsigned int stage, unsigned int num_units, unsigned int stride_conv_unit1)
126  {
127  for(unsigned int i = 0; i < num_units; ++i)
128  {
129  std::stringstream unit_path_ss;
130  unit_path_ss << "/cnn_data/resnext50_model/stage" << stage << "_unit" << (i + 1) << "_";
131  std::string unit_path = unit_path_ss.str();
132 
133  std::stringstream unit_name_ss;
134  unit_name_ss << "stage" << stage << "/unit" << (i + 1) << "/";
135  std::string unit_name = unit_name_ss.str();
136 
137  PadStrideInfo pad_grouped_conv(1, 1, 1, 1);
138  if(i == 0)
139  {
140  pad_grouped_conv = (stage == 1) ? PadStrideInfo(stride_conv_unit1, stride_conv_unit1, 1, 1) : PadStrideInfo(stride_conv_unit1, stride_conv_unit1, 0, 1, 0, 1, DimensionRoundingType::FLOOR);
141  }
142 
143  SubStream right(graph);
144  right << ConvolutionLayer(
145  1U, 1U, base_depth / 2,
146  get_weights_accessor(data_path, unit_path + "conv1_weights.npy", weights_layout),
147  get_weights_accessor(data_path, unit_path + "conv1_biases.npy"),
148  PadStrideInfo(1, 1, 0, 0))
149  .set_name(unit_name + "conv1/convolution")
150  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv1/Relu")
151 
152  << ConvolutionLayer(
153  3U, 3U, base_depth / 2,
154  get_weights_accessor(data_path, unit_path + "conv2_weights.npy", weights_layout),
155  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
156  pad_grouped_conv, 32)
157  .set_name(unit_name + "conv2/convolution")
158  << ScaleLayer(get_weights_accessor(data_path, unit_path + "bn2_mul.npy"),
159  get_weights_accessor(data_path, unit_path + "bn2_add.npy"))
160  .set_name(unit_name + "conv1/Scale")
161  << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "conv2/Relu")
162 
163  << ConvolutionLayer(
164  1U, 1U, base_depth,
165  get_weights_accessor(data_path, unit_path + "conv3_weights.npy", weights_layout),
166  get_weights_accessor(data_path, unit_path + "conv3_biases.npy"),
167  PadStrideInfo(1, 1, 0, 0))
168  .set_name(unit_name + "conv3/convolution");
169 
170  SubStream left(graph);
171  if(i == 0)
172  {
173  left << ConvolutionLayer(
174  1U, 1U, base_depth,
175  get_weights_accessor(data_path, unit_path + "sc_weights.npy", weights_layout),
176  std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
177  PadStrideInfo(stride_conv_unit1, stride_conv_unit1, 0, 0))
178  .set_name(unit_name + "sc/convolution")
179  << ScaleLayer(get_weights_accessor(data_path, unit_path + "sc_bn_mul.npy"),
180  get_weights_accessor(data_path, unit_path + "sc_bn_add.npy"))
181  .set_name(unit_name + "sc/scale");
182  }
183 
184  graph << EltwiseLayer(std::move(left), std::move(right), EltwiseOperation::Add).set_name(unit_name + "add");
185  graph << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name(unit_name + "Relu");
186  }
187  }
188 };
189 
190 /** Main program for ResNeXt50
191  *
192  * Model is based on:
193  * https://arxiv.org/abs/1611.05431
194  * "Aggregated Residual Transformations for Deep Neural Networks"
195  * Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, Kaiming He
196  *
197  * @note To list all the possible arguments execute the binary appended with the --help option
198  *
199  * @param[in] argc Number of arguments
200  * @param[in] argv Arguments
201  */
202 int main(int argc, char **argv)
203 {
204  return arm_compute::utils::run_example<GraphResNeXt50Example>(argc, argv);
205 }
int main(int argc, char **argv)
Main program for ResNeXt50.
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
1 channel, 1 F32 per channel
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.
std::unique_ptr< graph::ITensorAccessor > get_npy_output_accessor(const std::string &npy_path, TensorShape shape, DataType data_type, DataLayout data_layout=DataLayout::NCHW, std::ostream &output_stream=std::cout)
Generates appropriate npy output accessor according to the specified npy_path.
Definition: GraphUtils.h:611
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
#define ARM_COMPUTE_EXIT_ON_MSG(cond, msg)
If the condition is true, the given message is printed and program exits.
Definition: Error.h:379
Abstract Example class.
Definition: Utils.h:78
Num samples, channels, height, width.
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
bool is_data_type_quantized_asymmetric(DataType dt)
Check if a given data type is of asymmetric quantized type.
Definition: Utils.h:1190
TensorDescriptor & set_layout(DataLayout data_layout)
Sets tensor descriptor data layout.
Structure holding all the common graph parameters.
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