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/Graph.h"
  25 #include "arm_compute/graph/Nodes.h"
  26 #include "arm_compute/graph/SubGraph.h"
  27 #include "support/ToolchainSupport.h"
  28 #include "utils/GraphUtils.h"
  29 #include "utils/Utils.h"
  30
  31 #include <cstdlib>
  32 #include <tuple>
  33
  34 using namespace arm_compute::utils;
  35 using namespace arm_compute::graph;
  36 using namespace arm_compute::graph_utils;
  37
  38 /** Example demonstrating how to implement Googlenet's network using the Compute Library's graph API
  39  *
  40  * @param[in] argc Number of arguments
  41  * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
  42  */
  43 class GraphGooglenetExample : public Example
  44 {
  45 public:
  46     void do_setup(int argc, char **argv) override
  47     {
  48         std::string data_path; /* Path to the trainable data */
  49         std::string image;     /* Image data */
  50         std::string label;     /* Label data */
  51
  52         // Create a preprocessor object
  53         const std::array<float, 3> mean_rgb{ { 122.68f, 116.67f, 104.01f } };
  54         std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<CaffePreproccessor>(mean_rgb);
  55
  56         // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
  57         const int             int_target_hint  = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
  58         TargetHint            target_hint      = set_target_hint(int_target_hint);
  59         ConvolutionMethodHint convolution_hint = ConvolutionMethodHint::GEMM;
  60
  61         // Parse arguments
  62         if(argc < 2)
  63         {
  64             // Print help
  65             std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
  66             std::cout << "No data folder provided: using random values\n\n";
  67         }
  68         else if(argc == 2)
  69         {
  70             std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
  71             std::cout << "No data folder provided: using random values\n\n";
  72         }
  73         else if(argc == 3)
  74         {
  75             data_path = argv[2];
  76             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
  77             std::cout << "No image provided: using random values\n\n";
  78         }
  79         else if(argc == 4)
  80         {
  81             data_path = argv[2];
  82             image     = argv[3];
  83             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
  84             std::cout << "No text file with labels provided: skipping output accessor\n\n";
  85         }
  86         else
  87         {
  88             data_path = argv[2];
  89             image     = argv[3];
  90             label     = argv[4];
  91         }
  92
  93         graph << target_hint
  94               << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32),
  95                         get_input_accessor(image, std::move(preprocessor)))
  96               << ConvolutionLayer(
  97                   7U, 7U, 64U,
  98                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_w.npy"),
  99                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv1/conv1_7x7_s2_b.npy"),
 100                   PadStrideInfo(2, 2, 3, 3))
 101               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 102               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 103               << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f))
 104               << convolution_hint
 105               << ConvolutionLayer(
 106                   1U, 1U, 64U,
 107                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_w.npy"),
 108                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_reduce_b.npy"),
 109                   PadStrideInfo(1, 1, 0, 0))
 110               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 111               << ConvolutionLayer(
 112                   3U, 3U, 192U,
 113                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_w.npy"),
 114                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/conv2/conv2_3x3_b.npy"),
 115                   PadStrideInfo(1, 1, 1, 1))
 116               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 117               << NormalizationLayer(NormalizationLayerInfo(NormType::CROSS_MAP, 5, 0.0001f, 0.75f))
 118               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 119               << get_inception_node(data_path, "inception_3a", 64, std::make_tuple(96U, 128U), std::make_tuple(16U, 32U), 32U)
 120               << get_inception_node(data_path, "inception_3b", 128, std::make_tuple(128U, 192U), std::make_tuple(32U, 96U), 64U)
 121               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 122               << get_inception_node(data_path, "inception_4a", 192, std::make_tuple(96U, 208U), std::make_tuple(16U, 48U), 64U)
 123               << get_inception_node(data_path, "inception_4b", 160, std::make_tuple(112U, 224U), std::make_tuple(24U, 64U), 64U)
 124               << get_inception_node(data_path, "inception_4c", 128, std::make_tuple(128U, 256U), std::make_tuple(24U, 64U), 64U)
 125               << get_inception_node(data_path, "inception_4d", 112, std::make_tuple(144U, 288U), std::make_tuple(32U, 64U), 64U)
 126               << get_inception_node(data_path, "inception_4e", 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U)
 127               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 128               << get_inception_node(data_path, "inception_5a", 256, std::make_tuple(160U, 320U), std::make_tuple(32U, 128U), 128U)
 129               << get_inception_node(data_path, "inception_5b", 384, std::make_tuple(192U, 384U), std::make_tuple(48U, 128U), 128U)
 130               << PoolingLayer(PoolingLayerInfo(PoolingType::AVG, 7, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL)))
 131               << FullyConnectedLayer(
 132                   1000U,
 133                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_w.npy"),
 134                   get_weights_accessor(data_path, "/cnn_data/googlenet_model/loss3/loss3_classifier_b.npy"))
 135               << SoftmaxLayer()
 136               << Tensor(get_output_accessor(label, 5));
 137
 138         // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
 139         graph.graph_init(int_target_hint == 2);
 140     }
 141     void do_run() override
 142     {
 143         // Run graph
 144         graph.run();
 145     }
 146
 147 private:
 148     Graph graph{};
 149
 150     BranchLayer get_inception_node(const std::string &data_path, std::string &&param_path,
 151                                    unsigned int a_filt,
 152                                    std::tuple<unsigned int, unsigned int> b_filters,
 153                                    std::tuple<unsigned int, unsigned int> c_filters,
 154                                    unsigned int d_filt)
 155     {
 156         std::string total_path = "/cnn_data/googlenet_model/" + param_path + "/" + param_path + "_";
 157         SubGraph    i_a;
 158         i_a << ConvolutionLayer(
 159                 1U, 1U, a_filt,
 160                 get_weights_accessor(data_path, total_path + "1x1_w.npy"),
 161                 get_weights_accessor(data_path, total_path + "1x1_b.npy"),
 162                 PadStrideInfo(1, 1, 0, 0))
 163             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 164
 165         SubGraph i_b;
 166         i_b << ConvolutionLayer(
 167                 1U, 1U, std::get<0>(b_filters),
 168                 get_weights_accessor(data_path, total_path + "3x3_reduce_w.npy"),
 169                 get_weights_accessor(data_path, total_path + "3x3_reduce_b.npy"),
 170                 PadStrideInfo(1, 1, 0, 0))
 171             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 172             << ConvolutionLayer(
 173                 3U, 3U, std::get<1>(b_filters),
 174                 get_weights_accessor(data_path, total_path + "3x3_w.npy"),
 175                 get_weights_accessor(data_path, total_path + "3x3_b.npy"),
 176                 PadStrideInfo(1, 1, 1, 1))
 177             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 178
 179         SubGraph i_c;
 180         i_c << ConvolutionLayer(
 181                 1U, 1U, std::get<0>(c_filters),
 182                 get_weights_accessor(data_path, total_path + "5x5_reduce_w.npy"),
 183                 get_weights_accessor(data_path, total_path + "5x5_reduce_b.npy"),
 184                 PadStrideInfo(1, 1, 0, 0))
 185             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 186             << ConvolutionLayer(
 187                 5U, 5U, std::get<1>(c_filters),
 188                 get_weights_accessor(data_path, total_path + "5x5_w.npy"),
 189                 get_weights_accessor(data_path, total_path + "5x5_b.npy"),
 190                 PadStrideInfo(1, 1, 2, 2))
 191             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 192
 193         SubGraph i_d;
 194         i_d << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL)))
 195             << ConvolutionLayer(
 196                 1U, 1U, d_filt,
 197                 get_weights_accessor(data_path, total_path + "pool_proj_w.npy"),
 198                 get_weights_accessor(data_path, total_path + "pool_proj_b.npy"),
 199                 PadStrideInfo(1, 1, 0, 0))
 200             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 201
 202         return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b), std::move(i_c), std::move(i_d));
 203     }
 204 };
 205
 206 /** Main program for Googlenet
 207  *
 208  * @param[in] argc Number of arguments
 209  * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
 210  */
 211 int main(int argc, char **argv)
 212 {
 213     return arm_compute::utils::run_example<GraphGooglenetExample>(argc, argv);
 214 }