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.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 namespace
  37 {
  38 } // namespace
  39
  40 /** Example demonstrating how to implement Squeezenet's v1.1 network using the Compute Library's graph API
  41  *
  42  * @param[in] argc Number of arguments
  43  * @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) )
  44  */
  45 class GraphSqueezenet_v1_1Example : public Example
  46 {
  47 public:
  48     void do_setup(int argc, char **argv) override
  49     {
  50         std::string data_path; /* Path to the trainable data */
  51         std::string image;     /* Image data */
  52         std::string label;     /* Label data */
  53
  54         // Create a preprocessor object
  55         const std::array<float, 3> mean_rgb{ { 122.68f, 116.67f, 104.01f } };
  56         std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<CaffePreproccessor>(mean_rgb);
  57
  58         // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
  59         const int    target         = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
  60         Target       target_hint    = set_target_hint(target);
  61         FastMathHint fast_math_hint = FastMathHint::DISABLED;
  62
  63         // Parse arguments
  64         if(argc < 2)
  65         {
  66             // Print help
  67             std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
  68             std::cout << "No data folder provided: using random values\n\n";
  69         }
  70         else if(argc == 2)
  71         {
  72             std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
  73             std::cout << "No data folder provided: using random values\n\n";
  74         }
  75         else if(argc == 3)
  76         {
  77             data_path = argv[2];
  78             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels] [fast_math_hint]\n\n";
  79             std::cout << "No image provided: using random values\n\n";
  80         }
  81         else if(argc == 4)
  82         {
  83             data_path = argv[2];
  84             image     = argv[3];
  85             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
  86             std::cout << "No text file with labels provided: skipping output accessor\n\n";
  87         }
  88         else if(argc == 5)
  89         {
  90             data_path = argv[2];
  91             image     = argv[3];
  92             label     = argv[4];
  93             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
  94             std::cout << "No fast math info provided: disabling fast math\n\n";
  95         }
  96         else
  97         {
  98             data_path      = argv[2];
  99             image          = argv[3];
 100             label          = argv[4];
 101             fast_math_hint = (std::strtol(argv[5], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
 102         }
 103
 104         graph << target_hint
 105               << fast_math_hint
 106               << InputLayer(TensorDescriptor(TensorShape(227U, 227U, 3U, 1U), DataType::F32),
 107                             get_input_accessor(image, std::move(preprocessor)))
 108               << ConvolutionMethod::DIRECT
 109               << ConvolutionLayer(
 110                   3U, 3U, 64U,
 111                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv1_w.npy"),
 112                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv1_b.npy"),
 113                   PadStrideInfo(2, 2, 0, 0))
 114               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 115               << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 116               << ConvolutionMethod::DEFAULT
 117               << ConvolutionLayer(
 118                   1U, 1U, 16U,
 119                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire2_squeeze1x1_w.npy"),
 120                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire2_squeeze1x1_b.npy"),
 121                   PadStrideInfo(1, 1, 0, 0))
 122               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 123         graph << get_expand_fire_node(data_path, "fire2", 64U, 64U);
 124         graph << ConvolutionLayer(
 125                   1U, 1U, 16U,
 126                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire3_squeeze1x1_w.npy"),
 127                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire3_squeeze1x1_b.npy"),
 128                   PadStrideInfo(1, 1, 0, 0))
 129               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 130         graph << get_expand_fire_node(data_path, "fire3", 64U, 64U);
 131         graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 132               << ConvolutionLayer(
 133                   1U, 1U, 32U,
 134                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire4_squeeze1x1_w.npy"),
 135                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire4_squeeze1x1_b.npy"),
 136                   PadStrideInfo(1, 1, 0, 0))
 137               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 138         graph << get_expand_fire_node(data_path, "fire4", 128U, 128U);
 139         graph << ConvolutionLayer(
 140                   1U, 1U, 32U,
 141                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire5_squeeze1x1_w.npy"),
 142                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire5_squeeze1x1_b.npy"),
 143                   PadStrideInfo(1, 1, 0, 0))
 144               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 145         graph << get_expand_fire_node(data_path, "fire5", 128U, 128U);
 146         graph << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 147               << ConvolutionLayer(
 148                   1U, 1U, 48U,
 149                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire6_squeeze1x1_w.npy"),
 150                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire6_squeeze1x1_b.npy"),
 151                   PadStrideInfo(1, 1, 0, 0))
 152               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 153         graph << get_expand_fire_node(data_path, "fire6", 192U, 192U);
 154         graph << ConvolutionLayer(
 155                   1U, 1U, 48U,
 156                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire7_squeeze1x1_w.npy"),
 157                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire7_squeeze1x1_b.npy"),
 158                   PadStrideInfo(1, 1, 0, 0))
 159               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 160         graph << get_expand_fire_node(data_path, "fire7", 192U, 192U);
 161         graph << ConvolutionLayer(
 162                   1U, 1U, 64U,
 163                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire8_squeeze1x1_w.npy"),
 164                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire8_squeeze1x1_b.npy"),
 165                   PadStrideInfo(1, 1, 0, 0))
 166               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 167         graph << get_expand_fire_node(data_path, "fire8", 256U, 256U);
 168         graph << ConvolutionLayer(
 169                   1U, 1U, 64U,
 170                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire9_squeeze1x1_w.npy"),
 171                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/fire9_squeeze1x1_b.npy"),
 172                   PadStrideInfo(1, 1, 0, 0))
 173               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 174         graph << get_expand_fire_node(data_path, "fire9", 256U, 256U);
 175         graph << ConvolutionLayer(
 176                   1U, 1U, 1000U,
 177                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv10_w.npy"),
 178                   get_weights_accessor(data_path, "/cnn_data/squeezenet_v1_1_model/conv10_b.npy"),
 179                   PadStrideInfo(1, 1, 0, 0))
 180               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 181               << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
 182               << FlattenLayer()
 183               << SoftmaxLayer()
 184               << OutputLayer(get_output_accessor(label, 5));
 185
 186         // Finalize graph
 187         GraphConfig config;
 188         config.use_tuner = (target == 2);
 189         graph.finalize(target_hint, config);
 190     }
 191     void do_run() override
 192     {
 193         // Run graph
 194         graph.run();
 195     }
 196
 197 private:
 198     Stream graph{ 0, "SqueezeNetV1.1" };
 199
 200     BranchLayer get_expand_fire_node(const std::string &data_path, std::string &&param_path, unsigned int expand1_filt, unsigned int expand3_filt)
 201     {
 202         std::string total_path = "/cnn_data/squeezenet_v1_1_model/" + param_path + "_";
 203         SubStream   i_a(graph);
 204         i_a << ConvolutionLayer(
 205                 1U, 1U, expand1_filt,
 206                 get_weights_accessor(data_path, total_path + "expand1x1_w.npy"),
 207                 get_weights_accessor(data_path, total_path + "expand1x1_b.npy"),
 208                 PadStrideInfo(1, 1, 0, 0))
 209             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 210
 211         SubStream i_b(graph);
 212         i_b << ConvolutionLayer(
 213                 3U, 3U, expand3_filt,
 214                 get_weights_accessor(data_path, total_path + "expand3x3_w.npy"),
 215                 get_weights_accessor(data_path, total_path + "expand3x3_b.npy"),
 216                 PadStrideInfo(1, 1, 1, 1))
 217             << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
 218
 219         return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
 220     }
 221 };
 222
 223 /** Main program for Squeezenet v1.1
 224  *
 225  * @param[in] argc Number of arguments
 226  * @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) )
 227  */
 228 int main(int argc, char **argv)
 229 {
 230     return arm_compute::utils::run_example<GraphSqueezenet_v1_1Example>(argc, argv);
 231 }