ConvexCombinationLayer.cpp 4.8 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

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"
#include "paddle/math/Matrix.h"
Y
Yu Yang 已提交
17
#include "paddle/utils/Logging.h"
18 19 20 21 22
#include "paddle/utils/Stat.h"

namespace paddle {

/**
23
 * @brief A layer for weighted sum of vectors,
24 25
 * which is used in NEURAL MACHINE TRANSLATION BY JOINTLY LEARNING TO ALIGN AND
 * TRANSLATE
26 27 28
 * - Input: the the size of the first input is weightDim,
 *          and the size of the second input is weightdim * dataDim.
 * - Output: the sizeof the output is dataDim
29
 * \f[
30 31
 *   out(j) = \sum_{i}(in0(i) * in1(i,j + i * dataDim)),
 *               i = 0,1,...,(weightDim-1); j = 0, 1,...,(dataDim-1)
32
 * \f]
33 34
 * Note that the above computation is for one sample. Multiple samples are
 * processed in one batch.
35
 *
36
 * The config file api is linear_comb_layer.
37 38 39 40 41 42 43 44 45 46 47 48 49 50 51
 */
class ConvexCombinationLayer : public Layer {
protected:
  /// A matrix pointer pointing to second input.
  MatrixPtr tmpMtx0;
  /// A matrix pointer pointing to first input.
  MatrixPtr tmpRow0;
  /// A matrix pointer pointing to output.
  MatrixPtr tmpRow1;

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

  ~ConvexCombinationLayer() {}

52 53
  bool init(const LayerMap& layerMap,
            const ParameterMap& parameterMap) override;
54

55 56
  void forward(PassType passType) override;
  void backward(const UpdateCallback& callback = nullptr) override;
57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72
};

REGISTER_LAYER(convex_comb, ConvexCombinationLayer);

bool ConvexCombinationLayer::init(const LayerMap& layerMap,
                                  const ParameterMap& parameterMap) {
  /* Initialize the basic parent class */
  Layer::init(layerMap, parameterMap);

  CHECK_EQ(2U, inputLayers_.size());
  size_t dataDim = getSize();
  size_t weightDim = inputLayers_[0]->getSize();

  CHECK_EQ(weightDim * dataDim, inputLayers_[1]->getSize())
      << "Dimension mismatch";

73 74 75 76 77 78 79 80 81 82 83 84 85 86 87
  tmpRow0 = Matrix::create(nullptr,
                           /* height= */ 1,
                           weightDim,
                           /* trans= */ false,
                           useGpu_);
  tmpRow1 = Matrix::create(nullptr,
                           /* height= */ 1,
                           dataDim,
                           /* trans= */ false,
                           useGpu_);
  tmpMtx0 = Matrix::create(nullptr,
                           /* height= */ weightDim,
                           dataDim,
                           /* trans= */ false,
                           useGpu_);
88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116

  return true;
}

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

  MatrixPtr inV0 = getInputValue(0);
  MatrixPtr inV1 = getInputValue(1);

  size_t batchSize = inV0->getHeight();
  size_t weightDim = inV0->getWidth();
  size_t dataDim = getSize();

  CHECK_EQ(batchSize, inV1->getHeight());

  {
    REGISTER_TIMER_INFO("FwResetTimer", getName().c_str());
    reserveOutput(batchSize, dataDim);
  }

  MatrixPtr outV = getOutputValue();

  REGISTER_TIMER_INFO("FwCvxCombTimer", getName().c_str());
  for (size_t i = 0; i < batchSize; i++) {
    tmpMtx0->setData(inV1->getData() + i * weightDim * dataDim);
    tmpRow0->setData(inV0->getData() + i * weightDim);
    tmpRow1->setData(outV->getData() + i * dataDim);

117
    tmpRow1->mul(*tmpRow0, *tmpMtx0, 1, 0);
118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139
  }
}

void ConvexCombinationLayer::backward(const UpdateCallback& callback) {
  MatrixPtr outG = getOutputGrad();
  MatrixPtr inV0 = getInputValue(0);
  MatrixPtr inV1 = getInputValue(1);
  MatrixPtr inG0 = getInputGrad(0);
  MatrixPtr inG1 = getInputGrad(1);

  size_t batchSize = inV0->getHeight();
  size_t weightDim = inV0->getWidth();
  size_t dataDim = getSize();

  REGISTER_TIMER_INFO("BwCvxCombTimer", getName().c_str());

  if (inG0) {
    for (size_t i = 0; i < batchSize; i++) {
      tmpRow0->setData(inG0->getData() + i * weightDim);
      tmpRow1->setData(outG->getData() + i * dataDim);
      tmpMtx0->setData(inV1->getData() + i * weightDim * dataDim);

140
      tmpRow0->mul(*tmpRow1, *(tmpMtx0->getTranspose()), 1, 1);
141 142 143 144 145 146 147 148 149
    }
  }

  if (inG1) {
    for (size_t i = 0; i < batchSize; i++) {
      tmpRow0->setData(inV0->getData() + i * weightDim);
      tmpRow1->setData(outG->getData() + i * dataDim);
      tmpMtx0->setData(inG1->getData() + i * weightDim * dataDim);

150
      tmpMtx0->mul(*(tmpRow0->getTranspose()), *tmpRow1, 1, 1);
151 152 153 154 155
    }
  }
}

}  // namespace paddle
新手
引导
客服 返回
顶部