examples/graph_squeezenet_v1_1.cpp

   1 /*
   2  * Copyright (c) 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 using namespace arm_compute::logging;
  38
  39 namespace
  40 {
  41 } // namespace
  42
  43 /** Example demonstrating how to implement Squeezenet's v1.1 network using the Compute Library's graph API
  44  *
  45  * @param[in] argc Number of arguments
  46  * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
  47  */
  48 class GraphSqueezenet_v1_1Example : public Example
  49 {
  50 public:
  51     void do_setup(int argc, char **argv) override
  52     {
  53         std::string data_path; /* Path to the trainable data */
  54         std::string image;     /* Image data */
  55         std::string label;     /* Label data */
  56
  57         // Create a preprocessor object
  58         const std::array<float, 3> mean_rgb{ { 122.68f, 116.67f, 104.01f } };
  59         std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<CaffePreproccessor>(mean_rgb);
  60
  61         // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
  62         const int  int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
  63         TargetHint target_hint     = set_target_hint(int_target_hint);
  64
  65         // Parse arguments
  66         if(argc < 2)
  67         {
  68             // Print help
  69             std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
  70             std::cout << "No data folder provided: using random values\n\n";
  71         }
  72         else if(argc == 2)
  73         {
  74             std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
  75             std::cout << "No data folder provided: using random values\n\n";
  76         }
  77         else if(argc == 3)
  78         {
  79             data_path = argv[2];
  80             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
  81             std::cout << "No image provided: using random values\n\n";
  82         }
  83         else if(argc == 4)
  84         {
  85             data_path = argv[2];
  86             image     = argv[3];
  87             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
  88             std::cout << "No text file with labels provided: skipping output accessor\n\n";
  89         }
  90         else
  91         {
  92             data_path = argv[2];
  93             image     = argv[3];
  94             label     = argv[4];
  95         }
  96
  97         graph << target_hint
  98               << Tensor(TensorInfo(TensorShape(227U, 227U, 3U, 1U), 1, DataType::F32),
  99                         get_input_accessor(image, std::move(preprocessor)))
 100               << ConvolutionLayer(
 101                   3U, 3U, 64U,
 102                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv1_w.npy"),
 103                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv1_b.npy"),
 104                   PadStrideInfo(2, 2, 0, 0))
 105               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 106               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 107               << ConvolutionLayer(
 108                   1U, 1U, 16U,
 109                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire2_squeeze1x1_w.npy"),
 110                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire2_squeeze1x1_b.npy"),
 111                   PadStrideInfo(1, 1, 0, 0))
 112               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 113               << get_expand_fire_node(data_path, "fire2", 64U, 64U)
 114               << ConvolutionLayer(
 115                   1U, 1U, 16U,
 116                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire3_squeeze1x1_w.npy"),
 117                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire3_squeeze1x1_b.npy"),
 118                   PadStrideInfo(1, 1, 0, 0))
 119               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 120               << get_expand_fire_node(data_path, "fire3", 64U, 64U)
 121               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 122               << ConvolutionLayer(
 123                   1U, 1U, 32U,
 124                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire4_squeeze1x1_w.npy"),
 125                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire4_squeeze1x1_b.npy"),
 126                   PadStrideInfo(1, 1, 0, 0))
 127               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 128               << get_expand_fire_node(data_path, "fire4", 128U, 128U)
 129               << ConvolutionLayer(
 130                   1U, 1U, 32U,
 131                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire5_squeeze1x1_w.npy"),
 132                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire5_squeeze1x1_b.npy"),
 133                   PadStrideInfo(1, 1, 0, 0))
 134               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 135               << get_expand_fire_node(data_path, "fire5", 128U, 128U)
 136               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 137               << ConvolutionLayer(
 138                   1U, 1U, 48U,
 139                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire6_squeeze1x1_w.npy"),
 140                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire6_squeeze1x1_b.npy"),
 141                   PadStrideInfo(1, 1, 0, 0))
 142               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 143               << get_expand_fire_node(data_path, "fire6", 192U, 192U)
 144               << ConvolutionLayer(
 145                   1U, 1U, 48U,
 146                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire7_squeeze1x1_w.npy"),
 147                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire7_squeeze1x1_b.npy"),
 148                   PadStrideInfo(1, 1, 0, 0))
 149               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 150               << get_expand_fire_node(data_path, "fire7", 192U, 192U)
 151               << ConvolutionLayer(
 152                   1U, 1U, 64U,
 153                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire8_squeeze1x1_w.npy"),
 154                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire8_squeeze1x1_b.npy"),
 155                   PadStrideInfo(1, 1, 0, 0))
 156               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 157               << get_expand_fire_node(data_path, "fire8", 256U, 256U)
 158               << ConvolutionLayer(
 159                   1U, 1U, 64U,
 160                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire9_squeeze1x1_w.npy"),
 161                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire9_squeeze1x1_b.npy"),
 162                   PadStrideInfo(1, 1, 0, 0))
 163               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 164               << get_expand_fire_node(data_path, "fire9", 256U, 256U)
 165               << ConvolutionLayer(
 166                   1U, 1U, 1000U,
 167                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv10_w.npy"),
 168                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv10_b.npy"),
 169                   PadStrideInfo(1, 1, 0, 0))
 170               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 171               << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
 172               << FlattenLayer()
 173               << SoftmaxLayer()
 174               << Tensor(get_output_accessor(label, 5));
 175
 176         // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
 177         graph.graph_init(int_target_hint == 2);
 178     }
 179     void do_run() override
 180     {
 181         // Run graph
 182         graph.run();
 183     }
 184
 185 private:
 186     Graph graph{};
 187
 188     BranchLayer get_expand_fire_node(const std::string &data_path, std::string &&param_path, unsigned int expand1_filt, unsigned int expand3_filt)
 189     {
 190         std::string total_path = "/cnn_data/squeezenet_v1_1_model/" + param_path + "_";
 191         SubGraph    i_a;
 192         i_a << ConvolutionLayer(
 193                 1U, 1U, expand1_filt,
 194                 get_weights_accessor(data_path, total_path + "expand1x1_w.npy"),
 195                 get_weights_accessor(data_path, total_path + "expand1x1_b.npy"),
 196                 PadStrideInfo(1, 1, 0, 0))
 197             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 198
 199         SubGraph i_b;
 200         i_b << ConvolutionLayer(
 201                 3U, 3U, expand3_filt,
 202                 get_weights_accessor(data_path, total_path + "expand3x3_w.npy"),
 203                 get_weights_accessor(data_path, total_path + "expand3x3_b.npy"),
 204                 PadStrideInfo(1, 1, 1, 1))
 205             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 206
 207         return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
 208     }
 209 };
 210
 211 /** Main program for Squeezenet v1.1
 212  *
 213  * @param[in] argc Number of arguments
 214  * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
 215  */
 216 int main(int argc, char **argv)
 217 {
 218     return arm_compute::utils::run_example<GraphSqueezenet_v1_1Example>(argc, argv);
 219 }