examples/graph_squeezenet.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 using namespace arm_compute::logging;
  36
  37 /** Example demonstrating how to implement Squeezenet's network using the Compute Library's graph API
  38  *
  39  * @param[in] argc Number of arguments
  40  * @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) )
  41  */
  42 class GraphSqueezenetExample : public Example
  43 {
  44 public:
  45     void do_setup(int argc, char **argv) override
  46     {
  47         std::string data_path; /* Path to the trainable data */
  48         std::string image;     /* Image data */
  49         std::string label;     /* Label data */
  50
  51         // Create a preprocessor object
  52         const std::array<float, 3> mean_rgb{ { 122.68f, 116.67f, 104.01f } };
  53         std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<CaffePreproccessor>(mean_rgb);
  54
  55         // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
  56         const int    target         = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
  57         Target       target_hint    = set_target_hint(target);
  58         FastMathHint fast_math_hint = FastMathHint::DISABLED;
  59
  60         // Parse arguments
  61         if(argc < 2)
  62         {
  63             // Print help
  64             std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
  65             std::cout << "No data folder provided: using random values\n\n";
  66         }
  67         else if(argc == 2)
  68         {
  69             std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
  70             std::cout << "No data folder provided: using random values\n\n";
  71         }
  72         else if(argc == 3)
  73         {
  74             data_path = argv[2];
  75             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels] [fast_math_hint]\n\n";
  76             std::cout << "No image provided: using random values\n\n";
  77         }
  78         else if(argc == 4)
  79         {
  80             data_path = argv[2];
  81             image     = argv[3];
  82             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
  83             std::cout << "No text file with labels provided: skipping output accessor\n\n";
  84         }
  85         else if(argc == 5)
  86         {
  87             data_path = argv[2];
  88             image     = argv[3];
  89             label     = argv[4];
  90             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
  91             std::cout << "No fast math info provided: disabling fast math\n\n";
  92         }
  93         else
  94         {
  95             data_path      = argv[2];
  96             image          = argv[3];
  97             label          = argv[4];
  98             fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
  99         }
 100
 101         graph << target_hint
 102               << fast_math_hint
 103               << InputLayer(TensorDescriptor(TensorShape(224U, 224U, 3U, 1U), DataType::F32),
 104                             get_input_accessor(image, std::move(preprocessor)))
 105               << ConvolutionLayer(
 106                   7U, 7U, 96U,
 107                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_w.npy"),
 108                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_b.npy"),
 109                   PadStrideInfo(2, 2, 0, 0))
 110               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 111               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 112               << ConvolutionLayer(
 113                   1U, 1U, 16U,
 114                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_w.npy"),
 115                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_b.npy"),
 116                   PadStrideInfo(1, 1, 0, 0))
 117               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 118         graph << get_expand_fire_node(data_path, "fire2", 64U, 64U);
 119         graph << ConvolutionLayer(
 120                   1U, 1U, 16U,
 121                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_w.npy"),
 122                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_b.npy"),
 123                   PadStrideInfo(1, 1, 0, 0))
 124               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 125         graph << get_expand_fire_node(data_path, "fire3", 64U, 64U);
 126         graph << ConvolutionLayer(
 127                   1U, 1U, 32U,
 128                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_w.npy"),
 129                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_b.npy"),
 130                   PadStrideInfo(1, 1, 0, 0))
 131               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 132         graph << get_expand_fire_node(data_path, "fire4", 128U, 128U);
 133         graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 134               << ConvolutionLayer(
 135                   1U, 1U, 32U,
 136                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_w.npy"),
 137                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_b.npy"),
 138                   PadStrideInfo(1, 1, 0, 0))
 139               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 140         graph << get_expand_fire_node(data_path, "fire5", 128U, 128U);
 141         graph << ConvolutionLayer(
 142                   1U, 1U, 48U,
 143                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_w.npy"),
 144                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_b.npy"),
 145                   PadStrideInfo(1, 1, 0, 0))
 146               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 147         graph << get_expand_fire_node(data_path, "fire6", 192U, 192U);
 148         graph << ConvolutionLayer(
 149                   1U, 1U, 48U,
 150                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_w.npy"),
 151                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_b.npy"),
 152                   PadStrideInfo(1, 1, 0, 0))
 153               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 154         graph << get_expand_fire_node(data_path, "fire7", 192U, 192U);
 155         graph << ConvolutionLayer(
 156                   1U, 1U, 64U,
 157                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_w.npy"),
 158                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_b.npy"),
 159                   PadStrideInfo(1, 1, 0, 0))
 160               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 161         graph << get_expand_fire_node(data_path, "fire8", 256U, 256U);
 162         graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 163               << ConvolutionLayer(
 164                   1U, 1U, 64U,
 165                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_w.npy"),
 166                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_b.npy"),
 167                   PadStrideInfo(1, 1, 0, 0))
 168               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 169         graph << get_expand_fire_node(data_path, "fire9", 256U, 256U);
 170         graph << ConvolutionLayer(
 171                   1U, 1U, 1000U,
 172                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_w.npy"),
 173                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_b.npy"),
 174                   PadStrideInfo(1, 1, 0, 0))
 175               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 176               << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
 177               << FlattenLayer()
 178               << SoftmaxLayer()
 179               << OutputLayer(get_output_accessor(label, 5));
 180
 181         // Finalize graph
 182         GraphConfig config;
 183         config.use_tuner = (target == 2);
 184         graph.finalize(target_hint, config);
 185     }
 186     void do_run() override
 187     {
 188         // Run graph
 189         graph.run();
 190     }
 191
 192 private:
 193     Stream graph{ 0, "SqueezeNetV1" };
 194
 195     BranchLayer get_expand_fire_node(const std::string &data_path, std::string &&param_path, unsigned int expand1_filt, unsigned int expand3_filt)
 196     {
 197         std::string total_path = "/cnn_data/squeezenet_v1.0_model/" + param_path + "_";
 198         SubStream   i_a(graph);
 199         i_a << ConvolutionLayer(
 200                 1U, 1U, expand1_filt,
 201                 get_weights_accessor(data_path, total_path + "expand1x1_w.npy"),
 202                 get_weights_accessor(data_path, total_path + "expand1x1_b.npy"),
 203                 PadStrideInfo(1, 1, 0, 0))
 204             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 205
 206         SubStream i_b(graph);
 207         i_b << ConvolutionLayer(
 208                 3U, 3U, expand3_filt,
 209                 get_weights_accessor(data_path, total_path + "expand3x3_w.npy"),
 210                 get_weights_accessor(data_path, total_path + "expand3x3_b.npy"),
 211                 PadStrideInfo(1, 1, 1, 1))
 212             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 213
 214         return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
 215     }
 216 };
 217
 218 /** Main program for Squeezenet v1.0
 219  *
 220  * @param[in] argc Number of arguments
 221  * @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) )
 222  */
 223 int main(int argc, char **argv)
 224 {
 225     return arm_compute::utils::run_example<GraphSqueezenetExample>(argc, argv);
 226 }