examples/graph_squeezenet.cpp

   1 /*
   2  * Copyright (c) 2017 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::graph;
  35 using namespace arm_compute::graph_utils;
  36 using namespace arm_compute::logging;
  37
  38 namespace
  39 {
  40 BranchLayer get_expand_fire_node(const std::string &data_path, std::string &&param_path, unsigned int expand1_filt, unsigned int expand3_filt)
  41 {
  42     std::string total_path = "/cnn_data/squeezenet_v1.0_model/" + param_path + "_";
  43     SubGraph    i_a;
  44     i_a << ConvolutionLayer(
  45             1U, 1U, expand1_filt,
  46             get_weights_accessor(data_path, total_path + "expand1x1_w.npy"),
  47             get_weights_accessor(data_path, total_path + "expand1x1_b.npy"),
  48             PadStrideInfo(1, 1, 0, 0))
  49         << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
  50
  51     SubGraph i_b;
  52     i_b << ConvolutionLayer(
  53             3U, 3U, expand3_filt,
  54             get_weights_accessor(data_path, total_path + "expand3x3_w.npy"),
  55             get_weights_accessor(data_path, total_path + "expand3x3_b.npy"),
  56             PadStrideInfo(1, 1, 1, 1))
  57         << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
  58
  59     return BranchLayer(BranchMergeMethod::DEPTH_CONCATENATE, std::move(i_a), std::move(i_b));
  60 }
  61 } // namespace
  62
  63 /** Example demonstrating how to implement Squeezenet's network using the Compute Library's graph API
  64  *
  65  * @param[in] argc Number of arguments
  66  * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
  67  */
  68 void main_graph_squeezenet(int argc, const char **argv)
  69 {
  70     std::string data_path; /* Path to the trainable data */
  71     std::string image;     /* Image data */
  72     std::string label;     /* Label data */
  73
  74     constexpr float mean_r = 122.68f; /* Mean value to subtract from red channel */
  75     constexpr float mean_g = 116.67f; /* Mean value to subtract from green channel */
  76     constexpr float mean_b = 104.01f; /* Mean value to subtract from blue channel */
  77
  78     // Set target. 0 (NEON), 1 (OpenCL). By default it is NEON
  79     TargetHint            target_hint      = set_target_hint(argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0);
  80     ConvolutionMethodHint convolution_hint = target_hint == TargetHint::NEON ? ConvolutionMethodHint::GEMM : ConvolutionMethodHint::DIRECT;
  81
  82     // Parse arguments
  83     if(argc < 2)
  84     {
  85         // Print help
  86         std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [image] [labels]\n\n";
  87         std::cout << "No data folder provided: using random values\n\n";
  88     }
  89     else if(argc == 2)
  90     {
  91         std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [image] [labels]\n\n";
  92         std::cout << "No data folder provided: using random values\n\n";
  93     }
  94     else if(argc == 3)
  95     {
  96         data_path = argv[2];
  97         std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [image] [labels]\n\n";
  98         std::cout << "No image provided: using random values\n\n";
  99     }
 100     else if(argc == 4)
 101     {
 102         data_path = argv[2];
 103         image     = argv[3];
 104         std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
 105         std::cout << "No text file with labels provided: skipping output accessor\n\n";
 106     }
 107     else
 108     {
 109         data_path = argv[2];
 110         image     = argv[3];
 111         label     = argv[4];
 112     }
 113
 114     Graph graph;
 115
 116     graph << target_hint
 117           << Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::F32),
 118                     get_input_accessor(image, mean_r, mean_g, mean_b))
 119           << ConvolutionLayer(
 120               7U, 7U, 96U,
 121               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_w.npy"),
 122               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv1_b.npy"),
 123               PadStrideInfo(2, 2, 0, 0))
 124           << convolution_hint
 125           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 126           << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 127           << ConvolutionLayer(
 128               1U, 1U, 16U,
 129               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_w.npy"),
 130               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire2_squeeze1x1_b.npy"),
 131               PadStrideInfo(1, 1, 0, 0))
 132           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 133           << get_expand_fire_node(data_path, "fire2", 64U, 64U)
 134           << ConvolutionLayer(
 135               1U, 1U, 16U,
 136               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_w.npy"),
 137               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire3_squeeze1x1_b.npy"),
 138               PadStrideInfo(1, 1, 0, 0))
 139           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 140           << get_expand_fire_node(data_path, "fire3", 64U, 64U)
 141           << ConvolutionLayer(
 142               1U, 1U, 32U,
 143               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_w.npy"),
 144               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire4_squeeze1x1_b.npy"),
 145               PadStrideInfo(1, 1, 0, 0))
 146           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 147           << get_expand_fire_node(data_path, "fire4", 128U, 128U)
 148           << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 149           << ConvolutionLayer(
 150               1U, 1U, 32U,
 151               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_w.npy"),
 152               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire5_squeeze1x1_b.npy"),
 153               PadStrideInfo(1, 1, 0, 0))
 154           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 155           << get_expand_fire_node(data_path, "fire5", 128U, 128U)
 156           << ConvolutionLayer(
 157               1U, 1U, 48U,
 158               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_w.npy"),
 159               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire6_squeeze1x1_b.npy"),
 160               PadStrideInfo(1, 1, 0, 0))
 161           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 162           << get_expand_fire_node(data_path, "fire6", 192U, 192U)
 163           << ConvolutionLayer(
 164               1U, 1U, 48U,
 165               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_w.npy"),
 166               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire7_squeeze1x1_b.npy"),
 167               PadStrideInfo(1, 1, 0, 0))
 168           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 169           << get_expand_fire_node(data_path, "fire7", 192U, 192U)
 170           << ConvolutionLayer(
 171               1U, 1U, 64U,
 172               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_w.npy"),
 173               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire8_squeeze1x1_b.npy"),
 174               PadStrideInfo(1, 1, 0, 0))
 175           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 176           << get_expand_fire_node(data_path, "fire8", 256U, 256U)
 177           << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)))
 178           << ConvolutionLayer(
 179               1U, 1U, 64U,
 180               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_w.npy"),
 181               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/fire9_squeeze1x1_b.npy"),
 182               PadStrideInfo(1, 1, 0, 0))
 183           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 184           << get_expand_fire_node(data_path, "fire9", 256U, 256U)
 185           << ConvolutionLayer(
 186               1U, 1U, 1000U,
 187               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_w.npy"),
 188               get_weights_accessor(data_path, "/cnn_data/squeezenet_v1.0_model/conv10_b.npy"),
 189               PadStrideInfo(1, 1, 0, 0))
 190           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))
 191           << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
 192           << FlattenLayer()
 193           << SoftmaxLayer()
 194           << Tensor(get_output_accessor(label, 5));
 195
 196     graph.run();
 197 }
 198
 199 /** Main program for Squeezenet v1.0
 200  *
 201  * @param[in] argc Number of arguments
 202  * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL), [optional] Path to the weights folder, [optional] image, [optional] labels )
 203  */
 204 int main(int argc, const char **argv)
 205 {
 206     return arm_compute::utils::run_example(argc, argv, main_graph_squeezenet);
 207 }