Line data    Source code

       1              : // SPDX-License-Identifier: Apache-2.0
       2              : /**
       3              :  * Copyright (C) 2020 Parichay Kapoor <pk.kapoor@samsung.com>
       4              :  *
       5              :  * @file   preprocess_flip_layer.cpp
       6              :  * @date   20 January 2020
       7              :  * @see    https://github.com/nnstreamer/nntrainer
       8              :  * @author Parichay Kapoor <pk.kapoor@samsung.com>
       9              :  * @bug    No known bugs except for NYI items
      10              :  * @brief  This is Preprocess Random Flip Layer Class for Neural Network
      11              :  *
      12              :  */
      13              : 
      14              : #include <random>
      15              : 
      16              : #include <common_properties.h>
      17              : #include <layer_context.h>
      18              : #include <nntrainer_error.h>
      19              : #include <nntrainer_log.h>
      20              : #include <node_exporter.h>
      21              : #include <preprocess_flip_layer.h>
      22              : #include <util_func.h>
      23              : 
      24              : namespace nntrainer {
      25              : 
      26           33 : PreprocessFlipLayer::PreprocessFlipLayer() :
      27              :   Layer(),
      28           66 :   preprocess_flip_props(props::FlipDirection()) {}
      29              : 
      30            4 : void PreprocessFlipLayer::finalize(InitLayerContext &context) {
      31            4 :   context.setOutputDimensions(context.getInputDimensions());
      32              : 
      33              :   rng.seed(0);
      34            4 :   flip_dist = std::uniform_real_distribution<float>(0.0, 1.0);
      35            4 : }
      36              : 
      37           63 : void PreprocessFlipLayer::setProperty(const std::vector<std::string> &values) {
      38           63 :   auto remain_props = loadProperties(values, preprocess_flip_props);
      39           52 :   NNTR_THROW_IF(!remain_props.empty(), std::invalid_argument)
      40            2 :     << "[PreprocessFilpLayer] Unknown Layer Properties count " +
      41            4 :          std::to_string(values.size());
      42           52 : }
      43              : 
      44            0 : void PreprocessFlipLayer::forwarding(RunLayerContext &context, bool training) {
      45              :   props::FlipDirectionInfo::Enum flipdirection =
      46            0 :     std::get<props::FlipDirection>(preprocess_flip_props).get();
      47              : 
      48            0 :   if (!training) {
      49            0 :     for (unsigned int idx = 0; idx < context.getNumInputs(); idx++) {
      50              :       /** TODO: tell the graph to not include this when not training */
      51            0 :       context.getOutput(idx) = context.getInput(idx);
      52              :     }
      53              : 
      54              :     return;
      55              :   }
      56              : 
      57              :   using std::swap;
      58              :   bool fliph, flipw;
      59              : 
      60            0 :   for (unsigned int idx = 0; idx < context.getNumInputs(); idx++) {
      61            0 :     Tensor &hidden_ = context.getOutput(idx);
      62            0 :     Tensor &input_ = context.getInput(idx);
      63            0 :     const TensorDim input_dim = input_.getDim();
      64            0 :     unsigned int width = input_dim.width();
      65            0 :     unsigned int height = input_dim.height();
      66              : 
      67            0 :     for (unsigned int b = 0; b < input_dim.batch(); b++) {
      68              :       fliph = flipw = false;
      69            0 :       if (flip_dist(rng) < 0.5 &&
      70              :           flipdirection != props::FlipDirectionInfo::Enum::vertical)
      71              :         flipw = true;
      72              : 
      73            0 :       if (flip_dist(rng) < 0.5 &&
      74              :           flipdirection != props::FlipDirectionInfo::Enum::horizontal)
      75              :         fliph = true;
      76              : 
      77            0 :       if (!flipw && !fliph)
      78            0 :         continue;
      79              : 
      80            0 :       if (flipw) {
      81            0 :         for (unsigned int c = 0; c < input_dim.channel(); c++)
      82            0 :           for (unsigned int h = 0; h < input_dim.height(); h++)
      83            0 :             for (unsigned int w = 0; w < input_dim.width() / 2; w++)
      84            0 :               swap(*input_.getAddress<float>(b, c, h, w),
      85            0 :                    *input_.getAddress<float>(b, c, h, width - w - 1));
      86              :       }
      87            0 :       if (fliph) {
      88            0 :         for (unsigned int c = 0; c < input_dim.channel(); c++)
      89            0 :           for (unsigned int h = 0; h < input_dim.height() / 2; h++)
      90            0 :             for (unsigned int w = 0; w < input_dim.width(); w++)
      91            0 :               swap(*input_.getAddress<float>(b, c, h, w),
      92            0 :                    *input_.getAddress<float>(b, c, height - h - 1, w));
      93              :       }
      94              :     }
      95              :     /** @todo enable inPlace support for this layer */
      96            0 :     hidden_ = input_;
      97              :   }
      98              : }
      99              : 
     100            0 : void PreprocessFlipLayer::calcDerivative(RunLayerContext &context) {
     101              :   throw exception::not_supported(
     102            0 :     "calcDerivative for preprocess layer is not supported");
     103              : }
     104              : 
     105            0 : void PreprocessFlipLayer::exportTo(
     106              :   Exporter &exporter, const ml::train::ExportMethods &method) const {
     107            0 :   exporter.saveResult(preprocess_flip_props, method, this);
     108            0 : }
     109              : 
     110              : } /* namespace nntrainer */