examples/graph_googlenet.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 #include <tuple>
  31
  32 using namespace arm_compute::utils;
  33 using namespace arm_compute::graph::frontend;
  34 using namespace arm_compute::graph_utils;
  35
  36 /** Example demonstrating how to implement Googlenet'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, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
  40  */
  41 class GraphGooglenetExample : 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{ { 122.68f, 116.67f, 104.01f } };
  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    target         = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
  56         Target       target_hint    = set_target_hint(target);
  57         FastMathHint fast_math_hint = FastMathHint::DISABLED;
  58
  59         // Parse arguments
  60         if(argc < 2)
  61         {
  62             // Print help
  63             std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\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] [fast_math_hint]\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] [fast_math_hint]\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] [fast_math_hint]\n\n";
  82             std::cout << "No text file with labels provided: skipping output accessor\n\n";
  83         }
  84         else if(argc == 5)
  85         {
  86             data_path = argv[2];
  87             image     = argv[3];
  88             label     = argv[4];
  89             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
  90             std::cout << "No fast math info provided: disabling fast math\n\n";
  91         }
  92         else
  93         {
  94             data_path      = argv[2];
  95             image          = argv[3];
  96             label          = argv[4];
  97             fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
  98         }
  99
 100         graph << target_hint
 101               << fast_math_hint
 102               << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32),
 103                             get_input_accessor(image, std::move(preprocessor)))
 104               << ConvolutionLayer(
 105                   7U, 7U, 64U,
 106                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_w.npy"),
 107                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_b.npy"),
 108                   PadStrideInfo(2, 2, 3, 3))
 109               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 110               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 111               << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f))
 112               << ConvolutionLayer(
 113                   1U, 1U, 64U,
 114                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_w.npy"),
 115                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_b.npy"),
 116                   PadStrideInfo(1, 1, 0, 0))
 117               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 118               << ConvolutionLayer(
 119                   3U, 3U, 192U,
 120                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_w.npy"),
 121                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_b.npy"),
 122                   PadStrideInfo(1, 1, 1, 1))
 123               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 124               << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f))
 125               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)));
 126         graph << get_inception_node(data_path, "inception_3a", 64, std::make_tuple(96U, 128U), std::make_tuple(16U, 32U), 32U);
 127         graph << get_inception_node(data_path, "inception_3b", 128, std::make_tuple(128U, 192U), std::make_tuple(32U, 96U), 64U);
 128         graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)));
 129         graph << get_inception_node(data_path, "inception_4a", 192, std::make_tuple(96U, 208U), std::make_tuple(16U, 48U), 64U);
 130         graph << get_inception_node(data_path, "inception_4b", 160, std::make_tuple(112U, 224U), std::make_tuple(24U, 64U), 64U);
 131         graph << get_inception_node(data_path, "inception_4c", 128, std::make_tuple(128U, 256U), std::make_tuple(24U, 64U), 64U);
 132         graph << get_inception_node(data_path, "inception_4d", 112, std::make_tuple(144U, 288U), std::make_tuple(32U, 64U), 64U);
 133         graph << get_inception_node(data_path, "inception_4e", 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U);
 134         graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)));
 135         graph << get_inception_node(data_path, "inception_5a", 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U);
 136         graph << get_inception_node(data_path, "inception_5b", 384, std::make_tuple(192U, 384U), std::make_tuple(48U, 128U), 128U);
 137         graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 7, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL)))
 138               << FullyConnectedLayer(
 139                   1000U,
 140                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_w.npy"),
 141                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_b.npy"))
 142               << SoftmaxLayer()
 143               << OutputLayer(get_output_accessor(label, 5));
 144
 145         // Finalize graph
 146         GraphConfig config;
 147         config.use_tuner = (target == 2);
 148         graph.finalize(target_hint, config);
 149     }
 150     void do_run() override
 151     {
 152         // Run graph
 153         graph.run();
 154     }
 155
 156 private:
 157     Stream graph{ 0, "GoogleNet" };
 158
 159     BranchLayer get_inception_node(const std::string &data_path, std::string &&param_path,
 160                                    unsigned int a_filt,
 161                                    std::tuple<unsigned int, unsigned int> b_filters,
 162                                    std::tuple<unsigned int, unsigned int> c_filters,
 163                                    unsigned int d_filt)
 164     {
 165         std::string total_path = "/cnn_data/googlenet_model/" + param_path + "/" + param_path + "_";
 166         SubStream   i_a(graph);
 167         i_a << ConvolutionLayer(
 168                 1U, 1U, a_filt,
 169                 get_weights_accessor(data_path, total_path + "1x1_w.npy"),
 170                 get_weights_accessor(data_path, total_path + "1x1_b.npy"),
 171                 PadStrideInfo(1, 1, 0, 0))
 172             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 173
 174         SubStream i_b(graph);
 175         i_b << ConvolutionLayer(
 176                 1U, 1U, std::get<0>(b_filters),
 177                 get_weights_accessor(data_path, total_path + "3x3_reduce_w.npy"),
 178                 get_weights_accessor(data_path, total_path + "3x3_reduce_b.npy"),
 179                 PadStrideInfo(1, 1, 0, 0))
 180             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 181             << ConvolutionLayer(
 182                 3U, 3U, std::get<1>(b_filters),
 183                 get_weights_accessor(data_path, total_path + "3x3_w.npy"),
 184                 get_weights_accessor(data_path, total_path + "3x3_b.npy"),
 185                 PadStrideInfo(1, 1, 1, 1))
 186             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 187
 188         SubStream i_c(graph);
 189         i_c << ConvolutionLayer(
 190                 1U, 1U, std::get<0>(c_filters),
 191                 get_weights_accessor(data_path, total_path + "5x5_reduce_w.npy"),
 192                 get_weights_accessor(data_path, total_path + "5x5_reduce_b.npy"),
 193                 PadStrideInfo(1, 1, 0, 0))
 194             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 195             << ConvolutionLayer(
 196                 5U, 5U, std::get<1>(c_filters),
 197                 get_weights_accessor(data_path, total_path + "5x5_w.npy"),
 198                 get_weights_accessor(data_path, total_path + "5x5_b.npy"),
 199                 PadStrideInfo(1, 1, 2, 2))
 200             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 201
 202         SubStream i_d(graph);
 203         i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL)))
 204             << ConvolutionLayer(
 205                 1U, 1U, d_filt,
 206                 get_weights_accessor(data_path, total_path + "pool_proj_w.npy"),
 207                 get_weights_accessor(data_path, total_path + "pool_proj_b.npy"),
 208                 PadStrideInfo(1, 1, 0, 0))
 209             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 210
 211         return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d));
 212     }
 213 };
 214
 215 /** Main program for Googlenet
 216  *
 217  * @param[in] argc Number of arguments
 218  * @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) )
 219  */
 220 int main(int argc, char **argv)
 221 {
 222     return arm_compute::utils::run_example<GraphGooglenetExample>(argc, argv);
 223 }