examples/graph_mobilenet_qasymm8.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 using namespace arm_compute;
  31 using namespace arm_compute::graph;
  32 using namespace arm_compute::graph_utils;
  33
  34 /** Example demonstrating how to implement QASYMM8 MobileNet's network using the Compute Library's graph API
  35  *
  36  * @param[in] argc Number of arguments
  37  * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Path to the weights folder, [optional] npy_input, [optional] labels )
  38  */
  39 class GraphMobileNetQASYMM8Example : public utils::Example
  40 {
  41 public:
  42     void do_setup(int argc, char **argv) override
  43     {
  44         std::string data_path; /* Path to the trainable data */
  45         std::string input;     /* Image data */
  46         std::string label;     /* Label data */
  47
  48         // Quantization info taken from the AndroidNN QASYMM8 MobileNet example
  49         const QuantizationInfo in_quant_info  = QuantizationInfo(0.0078125f, 128);
  50         const QuantizationInfo mid_quant_info = QuantizationInfo(0.0784313753247f, 128);
  51
  52         const std::vector<QuantizationInfo> conv_weights_quant_info =
  53         {
  54             QuantizationInfo(0.031778190285f, 156), // conv0
  55             QuantizationInfo(0.00604454148561f, 66) // conv14
  56         };
  57
  58         const std::vector<QuantizationInfo> depth_weights_quant_info =
  59         {
  60             QuantizationInfo(0.254282623529f, 129),  // dwsc1
  61             QuantizationInfo(0.12828284502f, 172),   // dwsc2
  62             QuantizationInfo(0.265911251307f, 83),   // dwsc3
  63             QuantizationInfo(0.0985597148538f, 30),  // dwsc4
  64             QuantizationInfo(0.0631204470992f, 54),  // dwsc5
  65             QuantizationInfo(0.0137207424268f, 141), // dwsc6
  66             QuantizationInfo(0.0817828401923f, 125), // dwsc7
  67             QuantizationInfo(0.0393880493939f, 164), // dwsc8
  68             QuantizationInfo(0.211694166064f, 129),  // dwsc9
  69             QuantizationInfo(0.158015936613f, 103),  // dwsc10
  70             QuantizationInfo(0.0182712618262f, 137), // dwsc11
  71             QuantizationInfo(0.0127998134121f, 134), // dwsc12
  72             QuantizationInfo(0.299285322428f, 161)   // dwsc13
  73         };
  74
  75         const std::vector<QuantizationInfo> point_weights_quant_info =
  76         {
  77             QuantizationInfo(0.0425766184926f, 129),  // dwsc1
  78             QuantizationInfo(0.0250773020089f, 94),   // dwsc2
  79             QuantizationInfo(0.015851572156f, 93),    // dwsc3
  80             QuantizationInfo(0.0167811904103f, 98),   // dwsc4
  81             QuantizationInfo(0.00951790809631f, 135), // dwsc5
  82             QuantizationInfo(0.00999817531556f, 128), // dwsc6
  83             QuantizationInfo(0.00590536883101f, 126), // dwsc7
  84             QuantizationInfo(0.00576109671965f, 133), // dwsc8
  85             QuantizationInfo(0.00830461271107f, 142), // dwsc9
  86             QuantizationInfo(0.0152327232063f, 72),   // dwsc10
  87             QuantizationInfo(0.00741417845711f, 125), // dwsc11
  88             QuantizationInfo(0.0135628981516f, 142),  // dwsc12
  89             QuantizationInfo(0.0338749065995f, 140)   // dwsc13
  90         };
  91
  92         // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
  93         const int  int_target_hint = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
  94         TargetHint target_hint     = set_target_hint(int_target_hint);
  95
  96         // Parse arguments
  97         if(argc < 2)
  98         {
  99             // Print help
 100             std::cout << "Usage: " << argv[0] << " [target] [path_to_data] [npy_input] [labels]\n\n";
 101             std::cout << "No data folder provided: using random values\n\n";
 102         }
 103         else if(argc == 2)
 104         {
 105             std::cout << "Usage: " << argv[0] << " " << argv[1] << " [path_to_data] [npy_input] [labels]\n\n";
 106             std::cout << "No input provided: using random values\n\n";
 107         }
 108         else if(argc == 4)
 109         {
 110             data_path = argv[2];
 111             input     = argv[3];
 112             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels]\n\n";
 113             std::cout << "No text file with labels provided: skipping output accessor\n\n";
 114         }
 115         else
 116         {
 117             data_path = argv[2];
 118             input     = argv[3];
 119             label     = argv[4];
 120         }
 121
 122         graph << target_hint
 123               << arm_compute::graph::Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::QASYMM8, in_quant_info),
 124                                             get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/" + input))
 125               << ConvolutionLayer(
 126                   3U, 3U, 32U,
 127                   get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Conv2d_0_weights.npy"),
 128                   get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Conv2d_0_bias.npy"),
 129                   PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR),
 130                   1, WeightsInfo(),
 131                   conv_weights_quant_info.at(0),
 132                   mid_quant_info)
 133               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f))
 134               << get_dwsc_node(data_path, "Conv2d_1", 64U, PadStrideInfo(1U, 1U, 1U, 1U), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(0), point_weights_quant_info.at(0))
 135               << get_dwsc_node(data_path, "Conv2d_2", 128U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(1),
 136                                point_weights_quant_info.at(1))
 137               << get_dwsc_node(data_path, "Conv2d_3", 128U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(2),
 138                                point_weights_quant_info.at(2))
 139               << get_dwsc_node(data_path, "Conv2d_4", 256U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(3),
 140                                point_weights_quant_info.at(3))
 141               << get_dwsc_node(data_path, "Conv2d_5", 256U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(4),
 142                                point_weights_quant_info.at(4))
 143               << get_dwsc_node(data_path, "Conv2d_6", 512U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(5),
 144                                point_weights_quant_info.at(5))
 145               << get_dwsc_node(data_path, "Conv2d_7", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(6),
 146                                point_weights_quant_info.at(6))
 147               << get_dwsc_node(data_path, "Conv2d_8", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(7),
 148                                point_weights_quant_info.at(7))
 149               << get_dwsc_node(data_path, "Conv2d_9", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(8),
 150                                point_weights_quant_info.at(8))
 151               << get_dwsc_node(data_path, "Conv2d_10", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(9),
 152                                point_weights_quant_info.at(9))
 153               << get_dwsc_node(data_path, "Conv2d_11", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(10),
 154                                point_weights_quant_info.at(10))
 155               << get_dwsc_node(data_path, "Conv2d_12", 1024U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(11),
 156                                point_weights_quant_info.at(11))
 157               << get_dwsc_node(data_path, "Conv2d_13", 1024U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(12),
 158                                point_weights_quant_info.at(12))
 159               << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
 160               << ConvolutionLayer(
 161                   1U, 1U, 1001U,
 162                   get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Logits_Conv2d_1c_1x1_weights.npy"),
 163                   get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Logits_Conv2d_1c_1x1_bias.npy"),
 164                   PadStrideInfo(1U, 1U, 0U, 0U), 1, WeightsInfo(), conv_weights_quant_info.at(1))
 165               << ReshapeLayer(TensorShape(1001U))
 166               << SoftmaxLayer()
 167               << arm_compute::graph::Tensor(get_output_accessor(label, 5));
 168
 169         // In order to enable the OpenCL tuner, graph_init() has to be called only when all nodes have been instantiated
 170         graph.graph_init(int_target_hint == 2);
 171     }
 172     void do_run() override
 173     {
 174         // Run graph
 175         graph.run();
 176     }
 177
 178 private:
 179     Graph graph{};
 180
 181     /** This function produces a depthwise separable convolution node (i.e. depthwise + pointwise layers) with ReLU6 activation after each layer.
 182      *
 183      * @param[in] data_path                Path to trainable data folder
 184      * @param[in] param_path               Prefix of specific set of weights/biases data
 185      * @param[in] conv_filt                Filters depths for pointwise convolution
 186      * @param[in] dwc_pad_stride_info      PadStrideInfo for depthwise convolution
 187      * @param[in] conv_pad_stride_info     PadStrideInfo for pointwise convolution
 188      * @param[in] depth_weights_quant_info QuantizationInfo for depthwise convolution's weights
 189      * @param[in] point_weights_quant_info QuantizationInfo for pointwise convolution's weights
 190      *
 191      * @return The complete dwsc node
 192      */
 193     BranchLayer get_dwsc_node(const std::string &data_path, std::string &&param_path,
 194                               const unsigned int conv_filt,
 195                               PadStrideInfo dwc_pad_stride_info, PadStrideInfo conv_pad_stride_info,
 196                               QuantizationInfo depth_weights_quant_info, QuantizationInfo point_weights_quant_info)
 197     {
 198         std::string total_path = "/cnn_data/mobilenet_qasymm8_model/" + param_path + "_";
 199         SubGraph    sg;
 200
 201         sg << DepthwiseConvolutionLayer(
 202                3U, 3U,
 203                get_weights_accessor(data_path, total_path + "depthwise_weights.npy"),
 204                get_weights_accessor(data_path, total_path + "depthwise_bias.npy"),
 205                dwc_pad_stride_info,
 206                true,
 207                depth_weights_quant_info)
 208            << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f))
 209            << ConvolutionLayer(
 210                1U, 1U, conv_filt,
 211                get_weights_accessor(data_path, total_path + "pointwise_weights.npy"),
 212                get_weights_accessor(data_path, total_path + "pointwise_bias.npy"),
 213                conv_pad_stride_info,
 214                1, WeightsInfo(),
 215                point_weights_quant_info)
 216            << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f));
 217
 218         return BranchLayer(std::move(sg));
 219     }
 220 };
 221 /** Main program for MobileNetQASYMM8
 222  *
 223  * @param[in] argc Number of arguments
 224  * @param[in] argv Arguments ( [optional] Path to the weights folder, [optional] npy_input, [optional] labels )
 225  */
 226 int main(int argc, char **argv)
 227 {
 228     return utils::run_example<GraphMobileNetQASYMM8Example>(argc, argv);
 229 }