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.h"
  25 #include "support/ToolchainSupport.h"
  26 #include "utils/GraphUtils.h"
  27 #include "utils/Utils.h"
  28
  29 #include <cstdlib>
  30
  31 using namespace arm_compute::utils;
  32 using namespace arm_compute::graph::frontend;
  33 using namespace arm_compute::graph_utils;
  34
  35 /** Example demonstrating how to implement VGG16'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, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
  39  */
  40 class GraphVGG16Example : public Example
  41 {
  42 public:
  43     void do_setup(int argc, char **argv) override
  44     {
  45         std::string data_path; /* Path to the trainable data */
  46         std::string image;     /* Image data */
  47         std::string label;     /* Label data */
  48
  49         // Create a preprocessor object
  50         const std::array<float, 3> mean_rgb{ { 123.68f, 116.779f, 103.939f } };
  51         std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<CaffePreproccessor>(mean_rgb);
  52
  53         // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
  54         const int  target      = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
  55         Target     target_hint = set_target_hint(target);
  56         const bool is_opencl   = target_hint == Target::CL;
  57
  58         ConvolutionMethod first_convolution3x3_hint = is_opencl ? ConvolutionMethod::DIRECT : ConvolutionMethod::GEMM;
  59         ConvolutionMethod convolution3x3_hint       = ConvolutionMethod::DEFAULT;
  60         FastMathHint      fast_math_hint            = FastMathHint::DISABLED;
  61
  62         // Parse arguments
  63         if(argc < 2)
  64         {
  65             // Print help
  66             std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
  67             std::cout << "No data folder provided: using random values\n\n";
  68         }
  69         else if(argc == 2)
  70         {
  71             std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
  72             std::cout << "No data folder provided: using random values\n\n";
  73         }
  74         else if(argc == 3)
  75         {
  76             data_path = argv[2];
  77             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels] [fast_math_hint]\n\n";
  78             std::cout << "No image provided: using random values\n\n";
  79         }
  80         else if(argc == 4)
  81         {
  82             data_path = argv[2];
  83             image     = argv[3];
  84             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
  85             std::cout << "No text file with labels provided: skipping output accessor\n\n";
  86         }
  87         else if(argc == 5)
  88         {
  89             data_path = argv[2];
  90             image     = argv[3];
  91             label     = argv[4];
  92             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
  93             std::cout << "No fast math info provided: disabling fast math\n\n";
  94         }
  95         else
  96         {
  97             data_path      = argv[2];
  98             image          = argv[3];
  99             label          = argv[4];
 100             fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
 101         }
 102
 103         graph << target_hint
 104               << fast_math_hint
 105               << first_convolution3x3_hint
 106               << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32),
 107                             get_input_accessor(image, std::move(preprocessor)))
 108               // Layer 1
 109               << ConvolutionLayer(
 110                   3U, 3U, 64U,
 111                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_1_w.npy"),
 112                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_1_b.npy"),
 113                   PadStrideInfo(1, 1, 1, 1))
 114               .set_name("conv1_1")
 115               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1_1/Relu")
 116               << convolution3x3_hint
 117               // Layer 2
 118               << ConvolutionLayer(
 119                   3U, 3U, 64U,
 120                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_2_w.npy"),
 121                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv1_2_b.npy"),
 122                   PadStrideInfo(1, 1, 1, 1))
 123               .set_name("conv1_2")
 124               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv1_2/Relu")
 125               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool1")
 126               // Layer 3
 127               << ConvolutionLayer(
 128                   3U, 3U, 128U,
 129                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_1_w.npy"),
 130                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_1_b.npy"),
 131                   PadStrideInfo(1, 1, 1, 1))
 132               .set_name("conv2_1")
 133               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2_1/Relu")
 134               // Layer 4
 135               << ConvolutionLayer(
 136                   3U, 3U, 128U,
 137                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_2_w.npy"),
 138                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv2_2_b.npy"),
 139                   PadStrideInfo(1, 1, 1, 1))
 140               .set_name("conv2_2")
 141               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv2_2/Relu")
 142               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool2")
 143               // Layer 5
 144               << ConvolutionLayer(
 145                   3U, 3U, 256U,
 146                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_1_w.npy"),
 147                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_1_b.npy"),
 148                   PadStrideInfo(1, 1, 1, 1))
 149               .set_name("conv3_1")
 150               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_1/Relu")
 151               // Layer 6
 152               << ConvolutionLayer(
 153                   3U, 3U, 256U,
 154                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_2_w.npy"),
 155                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_2_b.npy"),
 156                   PadStrideInfo(1, 1, 1, 1))
 157               .set_name("conv3_2")
 158               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_2/Relu")
 159               // Layer 7
 160               << ConvolutionLayer(
 161                   3U, 3U, 256U,
 162                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_3_w.npy"),
 163                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv3_3_b.npy"),
 164                   PadStrideInfo(1, 1, 1, 1))
 165               .set_name("conv3_3")
 166               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv3_3/Relu")
 167               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool3")
 168               // Layer 8
 169               << ConvolutionLayer(
 170                   3U, 3U, 512U,
 171                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_1_w.npy"),
 172                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_1_b.npy"),
 173                   PadStrideInfo(1, 1, 1, 1))
 174               .set_name("conv4_1")
 175               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_1/Relu")
 176               // Layer 9
 177               << ConvolutionLayer(
 178                   3U, 3U, 512U,
 179                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_2_w.npy"),
 180                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_2_b.npy"),
 181                   PadStrideInfo(1, 1, 1, 1))
 182               .set_name("conv4_2")
 183               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_2/Relu")
 184               // Layer 10
 185               << ConvolutionLayer(
 186                   3U, 3U, 512U,
 187                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_3_w.npy"),
 188                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv4_3_b.npy"),
 189                   PadStrideInfo(1, 1, 1, 1))
 190               .set_name("conv4_3")
 191               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv4_3/Relu")
 192               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool4")
 193               // Layer 11
 194               << ConvolutionLayer(
 195                   3U, 3U, 512U,
 196                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_1_w.npy"),
 197                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_1_b.npy"),
 198                   PadStrideInfo(1, 1, 1, 1))
 199               .set_name("conv5_1")
 200               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_1/Relu")
 201               // Layer 12
 202               << ConvolutionLayer(
 203                   3U, 3U, 512U,
 204                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_2_w.npy"),
 205                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_2_b.npy"),
 206                   PadStrideInfo(1, 1, 1, 1))
 207               .set_name("conv5_2")
 208               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_2/Relu")
 209               // Layer 13
 210               << ConvolutionLayer(
 211                   3U, 3U, 512U,
 212                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_3_w.npy"),
 213                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/conv5_3_b.npy"),
 214                   PadStrideInfo(1, 1, 1, 1))
 215               .set_name("conv5_3")
 216               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("conv5_3/Relu")
 217               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2, 2, 0, 0))).set_name("pool5")
 218               // Layer 14
 219               << FullyConnectedLayer(
 220                   4096U,
 221                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc6_w.npy"),
 222                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc6_b.npy"))
 223               .set_name("fc6")
 224               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu")
 225               // Layer 15
 226               << FullyConnectedLayer(
 227                   4096U,
 228                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc7_w.npy"),
 229                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc7_b.npy"))
 230               .set_name("fc7")
 231               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("Relu_1")
 232               // Layer 16
 233               << FullyConnectedLayer(
 234                   1000U,
 235                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc8_w.npy"),
 236                   get_weights_accessor(data_path, "/cnn_data/vgg16_model/fc8_b.npy"))
 237               .set_name("fc8")
 238               // Softmax
 239               << SoftmaxLayer().set_name("prob")
 240               << OutputLayer(get_output_accessor(label, 5));
 241
 242         // Finalize graph
 243         GraphConfig config;
 244         config.use_tuner = (target == 2);
 245         graph.finalize(target_hint, config);
 246     }
 247     void do_run() override
 248     {
 249         // Run graph
 250         graph.run();
 251     }
 252
 253 private:
 254     Stream graph{ 0, "VGG16" };
 255 };
 256
 257 /** Main program for VGG16
 258  *
 259  * @param[in] argc Number of arguments
 260  * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
 261  */
 262 int main(int argc, char **argv)
 263 {
 264     return arm_compute::utils::run_example<GraphVGG16Example>(argc, argv);
 265 }