examples/graph_vgg19.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 VGG19'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 GraphVGG19Example : 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         // Create a preprocessor object
  51         const std::array<float, 3> mean_rgb{ { 123.68f, 116.779f, 103.939f } };
  52         std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<CaffePreproccessor>(mean_rgb);
  53
  54         // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
  55         const int             int_target_hint  = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
  56         TargetHint            target_hint      = set_target_hint(int_target_hint);
  57         ConvolutionMethodHint convolution_hint = ConvolutionMethodHint::DIRECT;
  58
  59         // Parse arguments
  60         if(argc < 2)
  61         {
  62             // Print help
  63             std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
  64             std::cout << "No data folder provided: using random values\n\n";
  65         }
  66         else if(argc == 2)
  67         {
  68             std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
  69             std::cout << "No data folder provided: using random values\n\n";
  70         }
  71         else if(argc == 3)
  72         {
  73             data_path = argv[2];
  74             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
  75             std::cout << "No image provided: using random values\n\n";
  76         }
  77         else if(argc == 4)
  78         {
  79             data_path = argv[2];
  80             image     = argv[3];
  81             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
  82             std::cout << "No text file with labels provided: skipping output accessor\n\n";
  83         }
  84         else
  85         {
  86             data_path = argv[2];
  87             image     = argv[3];
  88             label     = argv[4];
  89         }
  90
  91         graph << target_hint
  92               << convolution_hint
  93               << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32),
  94                         get_input_accessor(image, std::move(preprocessor)))
  95               // Layer 1
  96               << ConvolutionLayer(
  97                   3U, 3U, 64U,
  98                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_1_w.npy"),
  99                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_1_b.npy"),
 100                   PadStrideInfo(1, 1, 1, 1))
 101               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 102               << ConvolutionLayer(
 103                   3U, 3U, 64U,
 104                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_2_w.npy"),
 105                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_2_b.npy"),
 106                   PadStrideInfo(1, 1, 1, 1))
 107               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 108               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
 109               // Layer 2
 110               << ConvolutionLayer(
 111                   3U, 3U, 128U,
 112                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_1_w.npy"),
 113                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_1_b.npy"),
 114                   PadStrideInfo(1, 1, 1, 1))
 115               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 116               << ConvolutionLayer(
 117                   3U, 3U, 128U,
 118                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_2_w.npy"),
 119                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_2_b.npy"),
 120                   PadStrideInfo(1, 1, 1, 1))
 121               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 122               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
 123               // Layer 3
 124               << ConvolutionLayer(
 125                   3U, 3U, 256U,
 126                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_1_w.npy"),
 127                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_1_b.npy"),
 128                   PadStrideInfo(1, 1, 1, 1))
 129               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 130               << ConvolutionLayer(
 131                   3U, 3U, 256U,
 132                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_2_w.npy"),
 133                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_2_b.npy"),
 134                   PadStrideInfo(1, 1, 1, 1))
 135               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 136               << ConvolutionLayer(
 137                   3U, 3U, 256U,
 138                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_3_w.npy"),
 139                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_3_b.npy"),
 140                   PadStrideInfo(1, 1, 1, 1))
 141               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 142               << ConvolutionLayer(
 143                   3U, 3U, 256U,
 144                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_4_w.npy"),
 145                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_4_b.npy"),
 146                   PadStrideInfo(1, 1, 1, 1))
 147               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 148               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
 149               // Layer 4
 150               << ConvolutionLayer(
 151                   3U, 3U, 512U,
 152                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_1_w.npy"),
 153                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_1_b.npy"),
 154                   PadStrideInfo(1, 1, 1, 1))
 155               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 156               << ConvolutionLayer(
 157                   3U, 3U, 512U,
 158                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_2_w.npy"),
 159                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_2_b.npy"),
 160                   PadStrideInfo(1, 1, 1, 1))
 161               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 162               << ConvolutionLayer(
 163                   3U, 3U, 512U,
 164                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_3_w.npy"),
 165                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_3_b.npy"),
 166                   PadStrideInfo(1, 1, 1, 1))
 167               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 168               << ConvolutionLayer(
 169                   3U, 3U, 512U,
 170                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_4_w.npy"),
 171                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_4_b.npy"),
 172                   PadStrideInfo(1, 1, 1, 1))
 173               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 174               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
 175               // Layer 5
 176               << ConvolutionLayer(
 177                   3U, 3U, 512U,
 178                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_1_w.npy"),
 179                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_1_b.npy"),
 180                   PadStrideInfo(1, 1, 1, 1))
 181               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 182               << ConvolutionLayer(
 183                   3U, 3U, 512U,
 184                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_2_w.npy"),
 185                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_2_b.npy"),
 186                   PadStrideInfo(1, 1, 1, 1))
 187               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 188               << ConvolutionLayer(
 189                   3U, 3U, 512U,
 190                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_3_w.npy"),
 191                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_3_b.npy"),
 192                   PadStrideInfo(1, 1, 1, 1))
 193               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 194               << ConvolutionLayer(
 195                   3U, 3U, 512U,
 196                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_4_w.npy"),
 197                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_4_b.npy"),
 198                   PadStrideInfo(1, 1, 1, 1))
 199               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 200               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
 201               // Layer 6
 202               << FullyConnectedLayer(
 203                   4096U,
 204                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc6_w.npy"),
 205                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc6_b.npy"))
 206               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 207               // Layer 7
 208               << FullyConnectedLayer(
 209                   4096U,
 210                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc7_w.npy"),
 211                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc7_b.npy"))
 212               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 213               // Layer 8
 214               << FullyConnectedLayer(
 215                   1000U,
 216                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc8_w.npy"),
 217                   get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc8_b.npy"))
 218               // Softmax
 219               << SoftmaxLayer()
 220               << Tensor(get_output_accessor(label, 5));
 221
 222         // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
 223         graph.graph_init(int_target_hint == 2);
 224     }
 225     void do_run() override
 226     {
 227         // Run graph
 228         graph.run();
 229     }
 230
 231 private:
 232     Graph graph{};
 233 };
 234
 235 /** Main program for VGG19
 236  *
 237  * @param[in] argc Number of arguments
 238  * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
 239  */
 240 int main(int argc, char **argv)
 241 {
 242     return arm_compute::utils::run_example<GraphVGG19Example>(argc, argv);
 243 }