examples/graph_mobilenet.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
  31 using namespace arm_compute::utils;
  32 using namespace arm_compute::graph::frontend;
  33 using namespace arm_compute::graph_utils;
  34
  35 /** Example demonstrating how to implement MobileNet's network using the Compute Library's graph API
  36  *
  37  * @param[in] argc Number of arguments
  38  * @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] data layout, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
  39  */
  40 class GraphMobilenetExample : public Example
  41 {
  42 public:
  43     void do_setup(int argc, char **argv) override
  44     {
  45         std::string data_path; /* Path to the trainable data */
  46         std::string image;     /* Image data */
  47         std::string label;     /* Label data */
  48
  49         // Create a preprocessor object
  50         std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<TFPreproccessor>();
  51
  52         // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
  53         const int                  target                     = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
  54         Target                     target_hint                = set_target_hint(target);
  55         ConvolutionMethod          convolution_hint           = ConvolutionMethod::GEMM;
  56         DepthwiseConvolutionMethod depthwise_convolution_hint = DepthwiseConvolutionMethod::OPTIMIZED_3x3;
  57         FastMathHint               fast_math_hint             = FastMathHint::DISABLED;
  58
  59         // Set model to execute. 0 (MobileNetV1_1.0_224), 1 (MobileNetV1_0.75_160)
  60         int model_id = (argc > 2) ? std::strtol(argv[2], nullptr, 10) : 0;
  61         ARM_COMPUTE_ERROR_ON_MSG(model_id > 1, "Invalid model ID. Model must be 0 (MobileNetV1_1.0_224) or 1 (MobileNetV1_0.75_160)");
  62         int layout_id = (argc > 3) ? std::strtol(argv[3], nullptr, 10) : 0;
  63         ARM_COMPUTE_ERROR_ON_MSG(layout_id > 1, "Invalid layout ID. Layout must be 0 (NCHW) or 1 (NHWC)");
  64
  65         float            depth_scale           = (model_id == 0) ? 1.f : 0.75;
  66         unsigned int     spatial_size          = (model_id == 0) ? 224 : 160;
  67         std::string      model_path            = (model_id == 0) ? "/cnn_data/mobilenet_v1_1_224_model/" : "/cnn_data/mobilenet_v1_075_160_model/";
  68         TensorDescriptor input_descriptor_nchw = TensorDescriptor(TensorShape(spatial_size, spatial_size, 3U, 1U), DataType::F32);
  69         TensorDescriptor input_descriptor_nhwc = TensorDescriptor(TensorShape(3U, spatial_size, spatial_size, 1U), DataType::F32).set_layout(DataLayout::NHWC);
  70         TensorDescriptor input_descriptor      = (layout_id == 0) ? input_descriptor_nchw : input_descriptor_nhwc;
  71
  72         // Parse arguments
  73         if(argc < 2)
  74         {
  75             // Print help
  76             std::cout << "Usage: " << argv[0] << " [target] [model] [layout] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
  77             std::cout << "No model ID provided: using MobileNetV1_1.0_224\n\n";
  78             std::cout << "No data layout provided: using NCHW\n\n";
  79             std::cout << "No data folder provided: using random values\n\n";
  80         }
  81         else if(argc == 2)
  82         {
  83             std::cout << "Usage: " << argv[0] << " " << argv[1] << " [model] [layout] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
  84             std::cout << "No model ID provided: using MobileNetV1_1.0_224\n\n";
  85             std::cout << "No data layout provided: using NCHW\n\n";
  86             std::cout << "No data folder provided: using random values\n\n";
  87         }
  88         else if(argc == 3)
  89         {
  90             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [layout] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
  91             std::cout << "No data layout provided: using NCHW\n\n";
  92             std::cout << "No data folder provided: using random values\n\n";
  93         }
  94         else if(argc == 4)
  95         {
  96             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
  97             std::cout << "No data folder provided: using random values\n\n";
  98         }
  99         else if(argc == 5)
 100         {
 101             data_path = argv[4];
 102             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [image] [labels] [fast_math_hint]\n\n";
 103             std::cout << "No image provided: using random values\n\n";
 104             std::cout << "No text file with labels provided: skipping output accessor\n\n";
 105         }
 106         else if(argc == 6)
 107         {
 108             data_path = argv[4];
 109             image     = argv[5];
 110             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
 111             std::cout << "No text file with labels provided: skipping output accessor\n\n";
 112         }
 113         else if(argc == 7)
 114         {
 115             data_path = argv[4];
 116             image     = argv[5];
 117             label     = argv[6];
 118             std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
 119             std::cout << "No fast math info provided: disabling fast math\n\n";
 120         }
 121         else
 122         {
 123             data_path      = argv[4];
 124             image          = argv[5];
 125             label          = argv[6];
 126             fast_math_hint = (std::strtol(argv[7], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
 127         }
 128
 129         // Add model path to data path
 130         if(!data_path.empty())
 131         {
 132             data_path += model_path;
 133         }
 134
 135         graph << target_hint
 136               << convolution_hint
 137               << depthwise_convolution_hint
 138               << fast_math_hint
 139               << InputLayer(input_descriptor,
 140                             get_input_accessor(image, std::move(preprocessor), false))
 141               << ConvolutionLayer(
 142                   3U, 3U, 32U * depth_scale,
 143                   get_weights_accessor(data_path, "Conv2d_0_weights.npy", DataLayout::NCHW),
 144                   std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
 145                   PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))
 146               .set_name("Conv2d_0")
 147               << BatchNormalizationLayer(
 148                   get_weights_accessor(data_path, "Conv2d_0_BatchNorm_moving_mean.npy"),
 149                   get_weights_accessor(data_path, "Conv2d_0_BatchNorm_moving_variance.npy"),
 150                   get_weights_accessor(data_path, "Conv2d_0_BatchNorm_gamma.npy"),
 151                   get_weights_accessor(data_path, "Conv2d_0_BatchNorm_beta.npy"),
 152                   0.001f)
 153               .set_name("Conv2d_0/BatchNorm")
 154               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name("Conv2d_0/Relu6");
 155         graph << get_dwsc_node(data_path, "Conv2d_1", 64 * depth_scale, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0));
 156         graph << get_dwsc_node(data_path, "Conv2d_2", 128 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
 157         graph << get_dwsc_node(data_path, "Conv2d_3", 128 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
 158         graph << get_dwsc_node(data_path, "Conv2d_4", 256 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
 159         graph << get_dwsc_node(data_path, "Conv2d_5", 256 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
 160         graph << get_dwsc_node(data_path, "Conv2d_6", 512 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
 161         graph << get_dwsc_node(data_path, "Conv2d_7", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
 162         graph << get_dwsc_node(data_path, "Conv2d_8", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
 163         graph << get_dwsc_node(data_path, "Conv2d_9", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
 164         graph << get_dwsc_node(data_path, "Conv2d_10", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
 165         graph << get_dwsc_node(data_path, "Conv2d_11", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
 166         graph << get_dwsc_node(data_path, "Conv2d_12", 1024 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
 167         graph << get_dwsc_node(data_path, "Conv2d_13", 1024 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
 168         graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG)).set_name("Logits/AvgPool_1a")
 169               << ConvolutionLayer(
 170                   1U, 1U, 1001U,
 171                   get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy", DataLayout::NCHW),
 172                   get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_biases.npy"),
 173                   PadStrideInfo(1, 1, 0, 0))
 174               .set_name("Logits/Conv2d_1c_1x1")
 175               << ReshapeLayer(TensorShape(1001U)).set_name("Reshape")
 176               << SoftmaxLayer().set_name("Softmax")
 177               << OutputLayer(get_output_accessor(label, 5));
 178
 179         // Finalize graph
 180         GraphConfig config;
 181         config.use_tuner = (target == 2);
 182         graph.finalize(target_hint, config);
 183     }
 184     void do_run() override
 185     {
 186         // Run graph
 187         graph.run();
 188     }
 189
 190 private:
 191     Stream graph{ 0, "MobileNetV1" };
 192
 193     BranchLayer get_dwsc_node(const std::string &data_path, std::string &&param_path,
 194                               unsigned int  conv_filt,
 195                               PadStrideInfo dwc_pad_stride_info, PadStrideInfo conv_pad_stride_info)
 196     {
 197         std::string total_path = param_path + "_";
 198         SubStream   sg(graph);
 199         sg << DepthwiseConvolutionLayer(
 200                3U, 3U,
 201                get_weights_accessor(data_path, total_path + "depthwise_depthwise_weights.npy", DataLayout::NCHW),
 202                std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
 203                dwc_pad_stride_info)
 204            .set_name(total_path + "depthwise/depthwise")
 205            << BatchNormalizationLayer(
 206                get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_mean.npy"),
 207                get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_moving_variance.npy"),
 208                get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_gamma.npy"),
 209                get_weights_accessor(data_path, total_path + "depthwise_BatchNorm_beta.npy"),
 210                0.001f)
 211            .set_name(total_path + "depthwise/BatchNorm")
 212            << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name(total_path + "depthwise/Relu6")
 213            << ConvolutionLayer(
 214                1U, 1U, conv_filt,
 215                get_weights_accessor(data_path, total_path + "pointwise_weights.npy", DataLayout::NCHW),
 216                std::unique_ptr<arm_compute::graph::ITensorAccessor>(nullptr),
 217                conv_pad_stride_info)
 218            .set_name(total_path + "pointwise/Conv2D")
 219            << BatchNormalizationLayer(
 220                get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_moving_mean.npy"),
 221                get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_moving_variance.npy"),
 222                get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_gamma.npy"),
 223                get_weights_accessor(data_path, total_path + "pointwise_BatchNorm_beta.npy"),
 224                0.001f)
 225            .set_name(total_path + "pointwise/BatchNorm")
 226            << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name(total_path + "pointwise/Relu6");
 227
 228         return BranchLayer(std::move(sg));
 229     }
 230 };
 231
 232 /** Main program for MobileNetV1
 233  *
 234  * @param[in] argc Number of arguments
 235  * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner),
 236  *                             [optional] Model ID (0 = MobileNetV1_1.0_224, 1 = MobileNetV1_0.75_160),
 237  *                             [optional] Path to the weights folder,
 238  *                             [optional] image,
 239  *                             [optional] labels,
 240  *                             [optional] data layout,
 241  *                             [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
 242  */
 243 int main(int argc, char **argv)
 244 {
 245     return arm_compute::utils::run_example<GraphMobilenetExample>(argc, argv);
 246 }