SequenceReshapeLayer.cpp 4.6 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Z
zhangjinchao01 已提交
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"
Z
zhangjinchao01 已提交
18 19 20 21 22
#include "paddle/utils/Stat.h"

namespace paddle {

/**
23 24 25 26 27 28
 *  A layer for reshaping the sequence. Assume the input sequence has
 *  T instances, the dimension of each instance is M, and the input
 *  reshape_dim is N, then the output sequence has T*M/N instances,
 *  the dimension of each instance is N.
 *
 *  Note that T*M/N must be an integer.
Z
zhangjinchao01 已提交
29 30 31
 */

class SequenceReshapeLayer : public Layer {
W
Wu Yi 已提交
32
 protected:
Z
zhangjinchao01 已提交
33 34 35 36
  std::unique_ptr<Weight> biases_;

  MatrixPtr reshapedOutputGrad;

W
Wu Yi 已提交
37
 public:
Z
zhangjinchao01 已提交
38 39
  explicit SequenceReshapeLayer(const LayerConfig& config) : Layer(config) {}

Y
Yu Yang 已提交
40 41
  bool init(const LayerMap& layerMap,
            const ParameterMap& parameterMap) override;
Z
zhangjinchao01 已提交
42

Y
Yu Yang 已提交
43 44
  void forward(PassType passType) override;
  void backward(const UpdateCallback& callback = nullptr) override;
Z
zhangjinchao01 已提交
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
};

REGISTER_LAYER(seqreshape, SequenceReshapeLayer);

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

  CHECK_EQ(1U, inputLayers_.size());

  /* initialize biases_ */
  if (biasParameter_.get() != NULL) {
    biases_ = std::unique_ptr<Weight>(new Weight(1, getSize(), biasParameter_));
  }
  setNeedSequenceInfo(false);
  return true;
}

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

  const Argument& input = getInput(0);

  size_t inDim = input.value->getWidth();
  size_t outDim = getSize();

  size_t numSequences = input.getNumSequences();

74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89
  // by default, we assume each instance as a sequence
  IVectorPtr seqStarts;
  IVector::resizeOrCreate(seqStarts, input.getBatchSize() + 1, false);
  int* startsData = seqStarts->getData();
  for (int i = 0; i < input.getBatchSize() + 1; i++) {
    startsData[i] = i;
  }
  const int* starts = startsData;

  // if there is sequence, then use start positions
  if (input.sequenceStartPositions) {
    auto startPositions = input.sequenceStartPositions->getVector(false);
    starts = startPositions->getData();
    CHECK_EQ(starts[numSequences], input.getBatchSize());
    CHECK_EQ(numSequences, startPositions->getSize() - 1);
  }
Z
zhangjinchao01 已提交
90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110

  for (size_t seqID = 0; seqID < numSequences; seqID++) {
    size_t inNumIns = starts[seqID + 1] - starts[seqID];
    size_t outNumIns = inNumIns * inDim / outDim;
    CHECK_EQ(outNumIns * outDim, inNumIns * inDim);
  }

  MatrixPtr inputValue = getInputValue(0);

  // reset output
  reserveOutput(inputValue->getHeight() * inDim / outDim, outDim);
  MatrixPtr outputValue = getOutputValue();

  {
    AsyncGpuBlock asyncGpuBlock;
    REGISTER_TIMER_INFO("SequenceReshapeLayerForward", getName().c_str());

    outputValue->copyFrom(*inputValue);

    // modify the sequenceStartPositions
    ICpuGpuVector::resizeOrCreate(
111
        output_.sequenceStartPositions, numSequences + 1, false);
Z
zhangjinchao01 已提交
112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146

    int* tgtBuf = output_.sequenceStartPositions->getMutableData(false);

    for (size_t seqId = 0; seqId < numSequences + 1; ++seqId) {
      tgtBuf[seqId] = starts[seqId] * inDim / outDim;
    }
  }

  if (biases_.get() != NULL) {
    MatrixPtr outV = getOutputValue();
    outV->addBias(*(biases_->getW()), 1);
  }

  /* activation */
  forwardActivation();
}

void SequenceReshapeLayer::backward(const UpdateCallback& callback) {
  /* activation */
  backwardActivation();

  if (biases_ && biases_->getWGrad()) {
    biases_->getWGrad()->collectBias(*getOutputGrad(), 1);

    // Increasing the number of gradient
    biases_->getParameterPtr()->incUpdate(callback);
  }

  MatrixPtr inputGrad = getInputGrad(0);
  MatrixPtr outputGrad = getOutputGrad();

  AsyncGpuBlock asyncGpuBlock;
  REGISTER_TIMER_INFO("SequenceReshapeLayerBackward", getName().c_str());

  if (inputGrad) {
147 148 149 150 151
    Matrix::resizeOrCreate(reshapedOutputGrad,
                           inputGrad->getHeight(),
                           inputGrad->getWidth(),
                           false,
                           useGpu_);
Z
zhangjinchao01 已提交
152 153 154 155 156 157
    reshapedOutputGrad->copyFrom(*outputGrad);
    inputGrad->add(*reshapedOutputGrad);
  }
}

}  // namespace paddle