examples/graph_vgg16.cpp

   1 /*
   2  * Copyright (c) 2017, 2018 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/Graph.h"
  25 #include "arm_compute/graph/Nodes.h"
  26 #include "support/ToolchainSupport.h"
  27 #include "utils/GraphUtils.h"
  28 #include "utils/Utils.h"
  29
  30 #include <cstdlib>
  31
  32 using namespace arm_compute::utils;
  33 using namespace arm_compute::graph;
  34 using namespace arm_compute::graph_utils;
  35
  36 /** Example demonstrating how to implement VGG16's network using the Compute Library's graph API
  37  *
  38  * @param[in] argc Number of arguments
  39  * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
  40  */
  41 class GraphVGG16Example : public Example
  42 {
  43 public:
  44     void do_setup(int argc, char **argv) override
  45     {
  46         std::string data_path; /* Path to the trainable data */
  47         std::string image;     /* Image data */
  48         std::string label;     /* Label data */
  49
  50         constexpr float mean_r = 123.68f;  /* Mean value to subtract from red channel */
  51         constexpr float mean_g = 116.779f; /* Mean value to subtract from green channel */
  52         constexpr float mean_b = 103.939f; /* Mean value to subtract from blue channel */
  53
  54         // Set target. 0 (NEON), 1 (OpenCL). By default it is NEON
  55         TargetHint            target_hint      = set_target_hint(argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0);
  56         ConvolutionMethodHint convolution_hint = ConvolutionMethodHint::DIRECT;
  57
  58         // Parse arguments
  59         if(argc < 2)
  60         {
  61             // Print help
  62             std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
  63             std::cout << "No data folder provided: using random values\n\n";
  64         }
  65         else if(argc == 2)
  66         {
  67             std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
  68             std::cout << "No data folder provided: using random values\n\n";
  69         }
  70         else if(argc == 3)
  71         {
  72             data_path = argv[2];
  73             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
  74             std::cout << "No image provided: using random values\n\n";
  75         }
  76         else if(argc == 4)
  77         {
  78             data_path = argv[2];
  79             image     = argv[3];
  80             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
  81             std::cout << "No text file with labels provided: skipping output accessor\n\n";
  82         }
  83         else
  84         {
  85             data_path = argv[2];
  86             image     = argv[3];
  87             label     = argv[4];
  88         }
  89
  90         graph << target_hint
  91               << convolution_hint
  92               << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32),
  93                         get_input_accessor(image, mean_r, mean_g, mean_b))
  94               << ConvolutionMethodHint::DIRECT
  95               // Layer 1
  96               << ConvolutionLayer(
  97                   3U, 3U, 64U,
  98                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_1_w.npy"),
  99                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_1_b.npy"),
 100                   PadStrideInfo(1, 1, 1, 1))
 101               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 102               // Layer 2
 103               << ConvolutionLayer(
 104                   3U, 3U, 64U,
 105                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_2_w.npy"),
 106                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_2_b.npy"),
 107                   PadStrideInfo(1, 1, 1, 1))
 108               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 109               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
 110               // Layer 3
 111               << ConvolutionLayer(
 112                   3U, 3U, 128U,
 113                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_1_w.npy"),
 114                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_1_b.npy"),
 115                   PadStrideInfo(1, 1, 1, 1))
 116               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 117               // Layer 4
 118               << ConvolutionLayer(
 119                   3U, 3U, 128U,
 120                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_2_w.npy"),
 121                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_2_b.npy"),
 122                   PadStrideInfo(1, 1, 1, 1))
 123               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 124               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
 125               // Layer 5
 126               << ConvolutionLayer(
 127                   3U, 3U, 256U,
 128                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_1_w.npy"),
 129                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_1_b.npy"),
 130                   PadStrideInfo(1, 1, 1, 1))
 131               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 132               // Layer 6
 133               << ConvolutionLayer(
 134                   3U, 3U, 256U,
 135                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_2_w.npy"),
 136                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_2_b.npy"),
 137                   PadStrideInfo(1, 1, 1, 1))
 138               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 139               // Layer 7
 140               << ConvolutionLayer(
 141                   3U, 3U, 256U,
 142                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_3_w.npy"),
 143                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_3_b.npy"),
 144                   PadStrideInfo(1, 1, 1, 1))
 145               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 146               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
 147               // Layer 8
 148               << ConvolutionLayer(
 149                   3U, 3U, 512U,
 150                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_1_w.npy"),
 151                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_1_b.npy"),
 152                   PadStrideInfo(1, 1, 1, 1))
 153               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 154               // Layer 9
 155               << ConvolutionLayer(
 156                   3U, 3U, 512U,
 157                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_2_w.npy"),
 158                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_2_b.npy"),
 159                   PadStrideInfo(1, 1, 1, 1))
 160               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 161               // Layer 10
 162               << ConvolutionLayer(
 163                   3U, 3U, 512U,
 164                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_3_w.npy"),
 165                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_3_b.npy"),
 166                   PadStrideInfo(1, 1, 1, 1))
 167               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 168               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
 169               // Layer 11
 170               << ConvolutionLayer(
 171                   3U, 3U, 512U,
 172                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_1_w.npy"),
 173                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_1_b.npy"),
 174                   PadStrideInfo(1, 1, 1, 1))
 175               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 176               // Layer 12
 177               << ConvolutionLayer(
 178                   3U, 3U, 512U,
 179                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_2_w.npy"),
 180                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_2_b.npy"),
 181                   PadStrideInfo(1, 1, 1, 1))
 182               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 183               // Layer 13
 184               << ConvolutionLayer(
 185                   3U, 3U, 512U,
 186                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_3_w.npy"),
 187                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_3_b.npy"),
 188                   PadStrideInfo(1, 1, 1, 1))
 189               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 190               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
 191               // Layer 14
 192               << FullyConnectedLayer(
 193                   4096U,
 194                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc6_w.npy"),
 195                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc6_b.npy"))
 196               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 197               // Layer 15
 198               << FullyConnectedLayer(
 199                   4096U,
 200                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc7_w.npy"),
 201                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc7_b.npy"))
 202               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 203               // Layer 16
 204               << FullyConnectedLayer(
 205                   1000U,
 206                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc8_w.npy"),
 207                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc8_b.npy"))
 208               // Softmax
 209               << SoftmaxLayer()
 210               << Tensor(get_output_accessor(label, 5));
 211     }
 212     void do_run() override
 213     {
 214         // Run graph
 215         graph.run();
 216     }
 217
 218 private:
 219     Graph graph{};
 220 };
 221
 222 /** Main program for VGG16
 223  *
 224  * @param[in] argc Number of arguments
 225  * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
 226  */
 227 int main(int argc, char **argv)
 228 {
 229     return arm_compute::utils::run_example<GraphVGG16Example>(argc, argv);
 230 }