examples/graph_vgg19.cpp

   1 /*
   2  * Copyright (c) 2017 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::graph;
  33 using namespace arm_compute::graph_utils;
  34
  35 /** Example demonstrating how to implement VGG19's network using the Compute Library's graph API
  36  *
  37  * @param[in] argc Number of arguments
  38  * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
  39  */
  40 void main_graph_vgg19(int argc, const char **argv)
  41 {
  42     std::string data_path; /* Path to the trainable data */
  43     std::string image;     /* Image data */
  44     std::string label;     /* Label data */
  45
  46     constexpr float mean_r = 123.68f;  /* Mean value to subtract from red channel */
  47     constexpr float mean_g = 116.779f; /* Mean value to subtract from green channel */
  48     constexpr float mean_b = 103.939f; /* Mean value to subtract from blue channel */
  49
  50     // Set target. 0 (NEON), 1 (OpenCL). By default it is NEON
  51     TargetHint            target_hint      = set_target_hint(argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0);
  52     ConvolutionMethodHint convolution_hint = ConvolutionMethodHint::DIRECT;
  53
  54     // Parse arguments
  55     if(argc < 2)
  56     {
  57         // Print help
  58         std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
  59         std::cout << "No data folder provided: using random values\n\n";
  60     }
  61     else if(argc == 2)
  62     {
  63         std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
  64         std::cout << "No data folder provided: using random values\n\n";
  65     }
  66     else if(argc == 3)
  67     {
  68         data_path = argv[2];
  69         std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
  70         std::cout << "No image provided: using random values\n\n";
  71     }
  72     else if(argc == 4)
  73     {
  74         data_path = argv[2];
  75         image     = argv[3];
  76         std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
  77         std::cout << "No text file with labels provided: skipping output accessor\n\n";
  78     }
  79     else
  80     {
  81         data_path = argv[2];
  82         image     = argv[3];
  83         label     = argv[4];
  84     }
  85
  86     Graph graph;
  87
  88     graph << target_hint
  89           << convolution_hint
  90           << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32),
  91                     get_input_accessor(image, mean_r, mean_g, mean_b))
  92           // Layer 1
  93           << ConvolutionLayer(
  94               3U, 3U, 64U,
  95               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_1_w.npy"),
  96               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_1_b.npy"),
  97               PadStrideInfo(1, 1, 1, 1))
  98           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
  99           << ConvolutionLayer(
 100               3U, 3U, 64U,
 101               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_2_w.npy"),
 102               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv1_2_b.npy"),
 103               PadStrideInfo(1, 1, 1, 1))
 104           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 105           << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
 106           // Layer 2
 107           << ConvolutionLayer(
 108               3U, 3U, 128U,
 109               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_1_w.npy"),
 110               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_1_b.npy"),
 111               PadStrideInfo(1, 1, 1, 1))
 112           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 113           << ConvolutionLayer(
 114               3U, 3U, 128U,
 115               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_2_w.npy"),
 116               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv2_2_b.npy"),
 117               PadStrideInfo(1, 1, 1, 1))
 118           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 119           << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
 120           // Layer 3
 121           << ConvolutionLayer(
 122               3U, 3U, 256U,
 123               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_1_w.npy"),
 124               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_1_b.npy"),
 125               PadStrideInfo(1, 1, 1, 1))
 126           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 127           << ConvolutionLayer(
 128               3U, 3U, 256U,
 129               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_2_w.npy"),
 130               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_2_b.npy"),
 131               PadStrideInfo(1, 1, 1, 1))
 132           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 133           << ConvolutionLayer(
 134               3U, 3U, 256U,
 135               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_3_w.npy"),
 136               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_3_b.npy"),
 137               PadStrideInfo(1, 1, 1, 1))
 138           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 139           << ConvolutionLayer(
 140               3U, 3U, 256U,
 141               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_4_w.npy"),
 142               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv3_4_b.npy"),
 143               PadStrideInfo(1, 1, 1, 1))
 144           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 145           << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
 146           // Layer 4
 147           << ConvolutionLayer(
 148               3U, 3U, 512U,
 149               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_1_w.npy"),
 150               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_1_b.npy"),
 151               PadStrideInfo(1, 1, 1, 1))
 152           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 153           << ConvolutionLayer(
 154               3U, 3U, 512U,
 155               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_2_w.npy"),
 156               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_2_b.npy"),
 157               PadStrideInfo(1, 1, 1, 1))
 158           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 159           << ConvolutionLayer(
 160               3U, 3U, 512U,
 161               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_3_w.npy"),
 162               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_3_b.npy"),
 163               PadStrideInfo(1, 1, 1, 1))
 164           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 165           << ConvolutionLayer(
 166               3U, 3U, 512U,
 167               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_4_w.npy"),
 168               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv4_4_b.npy"),
 169               PadStrideInfo(1, 1, 1, 1))
 170           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 171           << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
 172           // Layer 5
 173           << ConvolutionLayer(
 174               3U, 3U, 512U,
 175               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_1_w.npy"),
 176               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_1_b.npy"),
 177               PadStrideInfo(1, 1, 1, 1))
 178           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 179           << ConvolutionLayer(
 180               3U, 3U, 512U,
 181               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_2_w.npy"),
 182               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_2_b.npy"),
 183               PadStrideInfo(1, 1, 1, 1))
 184           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 185           << ConvolutionLayer(
 186               3U, 3U, 512U,
 187               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_3_w.npy"),
 188               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_3_b.npy"),
 189               PadStrideInfo(1, 1, 1, 1))
 190           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 191           << ConvolutionLayer(
 192               3U, 3U, 512U,
 193               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_4_w.npy"),
 194               get_weights_accessor(data_path, "/cnn_data/vgg19_model/conv5_4_b.npy"),
 195               PadStrideInfo(1, 1, 1, 1))
 196           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 197           << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0)))
 198           // Layer 6
 199           << FullyConnectedLayer(
 200               4096U,
 201               get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc6_w.npy"),
 202               get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc6_b.npy"))
 203           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 204           // Layer 7
 205           << FullyConnectedLayer(
 206               4096U,
 207               get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc7_w.npy"),
 208               get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc7_b.npy"))
 209           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 210           // Layer 8
 211           << FullyConnectedLayer(
 212               1000U,
 213               get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc8_w.npy"),
 214               get_weights_accessor(data_path, "/cnn_data/vgg19_model/fc8_b.npy"))
 215           // Softmax
 216           << SoftmaxLayer()
 217           << Tensor(get_output_accessor(label, 5));
 218
 219     // Run graph
 220     graph.run();
 221 }
 222
 223 /** Main program for VGG19
 224  *
 225  * @param[in] argc Number of arguments
 226  * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
 227  */
 228 int main(int argc, const char **argv)
 229 {
 230     return arm_compute::utils::run_example(argc, argv, main_graph_vgg19);
 231 }