examples/graph_mobilenet.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 MobileNet'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 )
  40  */
  41 class GraphMobilenetExample : 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         std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<TFPreproccessor>();
  52
  53         // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
  54         const int             int_target_hint  = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
  55         TargetHint            target_hint      = set_target_hint(int_target_hint);
  56         ConvolutionMethodHint convolution_hint = ConvolutionMethodHint::GEMM;
  57
  58         // Set model to execute. 0 (MobileNetV1_1.0_224), 1 (MobileNetV1_0.75_160)
  59         int model_id = (argc > 2) ? std::strtol(argv[2], nullptr, 10) : 0;
  60         ARM_COMPUTE_ERROR_ON_MSG(model_id > 1, "Invalid model ID. Model must be 0 (MobileNetV1_1.0_224) or 1 (MobileNetV1_0.75_160)");
  61         float        depth_scale  = (model_id == 0) ? 1.f : 0.75;
  62         unsigned int spatial_size = (model_id == 0) ? 224 : 160;
  63         std::string  model_path   = (model_id == 0) ? "/cnn_data/mobilenet_v1_1_224_model/" : "/cnn_data/mobilenet_v1_075_160_model/";
  64
  65         // Parse arguments
  66         if(argc < 2)
  67         {
  68             // Print help
  69             std::cout << "Usage: " << argv[0] << " [target] [model] [path_to_data] [image] [labels]\n\n";
  70             std::cout << "No model ID provided: using MobileNetV1_1.0_224\n\n";
  71             std::cout << "No data folder provided: using random values\n\n";
  72         }
  73         else if(argc == 2)
  74         {
  75             std::cout << "Usage: " << argv[0] << " " << argv[1] << " [model] [path_to_data] [image] [labels]\n\n";
  76             std::cout << "No model ID provided: using MobileNetV1_1.0_224\n\n";
  77             std::cout << "No data folder provided: using random values\n\n";
  78         }
  79         else if(argc == 3)
  80         {
  81             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [path_to_data] [image] [labels]\n\n";
  82             std::cout << "No data folder provided: using random values\n\n";
  83         }
  84         else if(argc == 4)
  85         {
  86             data_path = argv[3];
  87             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [image] [labels]\n\n";
  88             std::cout << "No image provided: using random values\n\n";
  89         }
  90         else if(argc == 5)
  91         {
  92             data_path = argv[3];
  93             image     = argv[4];
  94             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
  95             std::cout << "No text file with labels provided: skipping output accessor\n\n";
  96         }
  97         else
  98         {
  99             data_path = argv[3];
 100             image     = argv[4];
 101             label     = argv[5];
 102         }
 103
 104         // Add model path to data path
 105         if(!data_path.empty())
 106         {
 107             data_path += model_path;
 108         }
 109
 110         graph << target_hint
 111               << convolution_hint
 112               << Tensor(TensorInfo(TensorShape(spatial_size, spatial_size, 3U, 1U), 1, DataType::F32),
 113                         get_input_accessor(image, std::move(preprocessor), false))
 114               << ConvolutionLayer(
 115                   3U, 3U, 32U * depth_scale,
 116                   get_weights_accessor(data_path, "Conv2d_0_weights.npy"),
 117                   std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
 118                   PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))
 119               << BatchNormalizationLayer(
 120                   get_weights_accessor(data_path, "Conv2d_0_BatchNorm_moving_mean.npy"),
 121                   get_weights_accessor(data_path, "Conv2d_0_BatchNorm_moving_variance.npy"),
 122                   get_weights_accessor(data_path, "Conv2d_0_BatchNorm_gamma.npy"),
 123                   get_weights_accessor(data_path, "Conv2d_0_BatchNorm_beta.npy"),
 124                   0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f))
 125               << get_dwsc_node(data_path, "Conv2d_1", 64 * depth_scale, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0))
 126               << get_dwsc_node(data_path, "Conv2d_2", 128 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
 127               << get_dwsc_node(data_path, "Conv2d_3", 128 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
 128               << get_dwsc_node(data_path, "Conv2d_4", 256 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
 129               << get_dwsc_node(data_path, "Conv2d_5", 256 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
 130               << get_dwsc_node(data_path, "Conv2d_6", 512 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
 131               << get_dwsc_node(data_path, "Conv2d_7", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
 132               << get_dwsc_node(data_path, "Conv2d_8", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
 133               << get_dwsc_node(data_path, "Conv2d_9", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
 134               << get_dwsc_node(data_path, "Conv2d_10", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
 135               << get_dwsc_node(data_path, "Conv2d_11", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
 136               << get_dwsc_node(data_path, "Conv2d_12", 1024 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
 137               << get_dwsc_node(data_path, "Conv2d_13", 1024 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0))
 138               << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
 139               << ConvolutionLayer(
 140                   1U, 1U, 1001U,
 141                   get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy"),
 142                   get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_biases.npy"),
 143                   PadStrideInfo(1, 1, 0, 0))
 144               << ReshapeLayer(TensorShape(1001U))
 145               << SoftmaxLayer()
 146               << Tensor(get_output_accessor(label, 5));
 147
 148         // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
 149         graph.graph_init(int_target_hint == 2);
 150     }
 151     void do_run() override
 152     {
 153         // Run graph
 154         graph.run();
 155     }
 156
 157 private:
 158     Graph graph{};
 159
 160     BranchLayer get_dwsc_node(const std::string &data_path, std::string &&param_path,
 161                               unsigned int  conv_filt,
 162                               PadStrideInfo dwc_pad_stride_info, PadStrideInfo conv_pad_stride_info)
 163     {
 164         std::string total_path = param_path + "_";
 165         SubGraph    sg;
 166         sg << DepthwiseConvolutionLayer(
 167                3U, 3U,
 168                get_weights_accessor(data_path, total_path + "depthwise_depthwise_weights.npy"),
 169                std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
 170                dwc_pad_stride_info,
 171                true)
 172            << BatchNormalizationLayer(
 173                get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_mean.npy"),
 174                get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_variance.npy"),
 175                get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_gamma.npy"),
 176                get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_beta.npy"),
 177                0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f))
 178            << ConvolutionLayer(
 179                1U, 1U, conv_filt,
 180                get_weights_accessor(data_path, total_path + "pointwise_weights.npy"),
 181                std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
 182                conv_pad_stride_info)
 183            << BatchNormalizationLayer(
 184                get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_moving_mean.npy"),
 185                get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_moving_variance.npy"),
 186                get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_gamma.npy"),
 187                get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_beta.npy"),
 188                0.001f, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f));
 189
 190         return BranchLayer(std::move(sg));
 191     }
 192 };
 193
 194 /** Main program for MobileNetV1
 195  *
 196  * @param[in] argc Number of arguments
 197  * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL),
 198  *                             [optional] Model ID (0 = MobileNetV1_1.0_224, 1 = MobileNetV1_0.75_160),
 199  *                             [optional] Path to the weights folder,
 200  *                             [optional] image,
 201  *                             [optional] labels )
 202  */
 203 int main(int argc, char **argv)
 204 {
 205     return arm_compute::utils::run_example<GraphMobilenetExample>(argc, argv);
 206 }