ScaleShiftLayer.cpp 3.4 KB
Newer Older
G
guosheng 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#include "Layer.h"

namespace paddle {

/**
20 21 22
 * A layer applies a slope and an intercept to the input element-wise for
 * scaling and shifting. Noting that this layer is trainable which differs
 * from the SlopeInterceptLayer.
G
guosheng 已提交
23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107
 *
 * \f[
 *    y = wx + b
 * \f]
 *
 * Here, w is scale and b is offset, which are scalars and trainable.
 *
 */

class ScaleShiftLayer : public Layer {
protected:
  std::unique_ptr<Weight> scale_;
  std::unique_ptr<Weight> offset_;

public:
  explicit ScaleShiftLayer(const LayerConfig& config) : Layer(config) {}

  bool init(const LayerMap& layerMap,
            const ParameterMap& parameterMap) override;

  void forward(PassType passType) override;
  void backward(const UpdateCallback& callback = nullptr) override;
};

REGISTER_LAYER(scale_shift, ScaleShiftLayer);

bool ScaleShiftLayer::init(const LayerMap& layerMap,
                           const ParameterMap& parameterMap) {
  Layer::init(layerMap, parameterMap);
  CHECK_EQ(inputLayers_.size(), 1U);
  scale_.reset(new Weight(1, 1, parameters_[0]));
  if (biasParameter_.get() != NULL) {
    offset_ = std::unique_ptr<Weight>(new Weight(1, 1, biasParameter_));
  }
  return true;
}

void ScaleShiftLayer::forward(PassType passType) {
  Layer::forward(passType);

  MatrixPtr inV = getInputValue(0);
  resetOutput(inV->getHeight(), inV->getWidth());
  MatrixPtr outV = getOutputValue();
  real scaleValue = scale_->getW()->getElement(0, 0);
  outV->mulScalar(*inV, scaleValue);
  if (offset_) {
    real offsetValue = offset_->getW()->getElement(0, 0);
    outV->add(offsetValue);
  }
}

void ScaleShiftLayer::backward(const UpdateCallback& callback) {
  MatrixPtr inV = getInputValue(0);
  MatrixPtr inG = getInputGrad(0);
  MatrixPtr outV = getOutputValue();
  MatrixPtr outG = getOutputGrad();

  /* Calculate the parameter gradient for the current layer */
  if (scale_->getWGrad()) {
    MatrixPtr rowSumMtx;
    Matrix::resizeOrCreate(rowSumMtx, outG->getHeight(), 1, false, useGpu_);
    // this_i = scaleDest * this_i + scaleSum * \sum_j b_{ij} * c_{ij}
    rowSumMtx->sumOfProducts(
        /* b= */ *inV, /* c= */ *outG, /* scaleSum= */ 1, /* scaleDest= */ 0.);
    // this_i = scaleDest * this_i + scaleSum * \sum_j b_{ji}
    scale_->getWGrad()->sumCols(
        /* b= */ *rowSumMtx, /* scaleSum= */ 1., /* scaleDest= */ 1.);
    scale_->getParameterPtr()->incUpdate(callback);
  }
  if (offset_ && offset_->getWGrad()) {
    MatrixPtr rowSumMtx;
    Matrix::resizeOrCreate(rowSumMtx, outG->getHeight(), 1, false, useGpu_);
    rowSumMtx->sumRows(*outG, 1., 0.);
    offset_->getWGrad()->sumCols(*rowSumMtx, 1., 1.);
    offset_->getParameterPtr()->incUpdate(callback);
  }

  /* Calculate the input layers error */
  if (inG) {
    real scaleValue = scale_->getW()->getElement(0, 0);
    inG->add(*outG, scaleValue);
  }
}

}  // namespace paddle