SubSequenceLayer.cpp 6.7 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

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"
X
Xin Pan 已提交
16 17
#include "paddle/legacy/math/Matrix.h"
#include "paddle/legacy/math/Vector.h"
Y
Yu Yang 已提交
18
#include "paddle/utils/Logging.h"
Z
zhangjinchao01 已提交
19 20 21 22 23 24 25 26 27 28 29
#include "paddle/utils/Stat.h"

namespace paddle {

/**
 * A layer for taking the subsequence according to given offset and size
 * Input: original sequence, offset, size
 * Output: subsequence
 */

class SubSequenceLayer : public Layer {
W
Wu Yi 已提交
30
 protected:
Z
zhangjinchao01 已提交
31 32 33 34
  std::unique_ptr<Weight> biases_;
  MatrixPtr tmpSrc_;
  MatrixPtr tmpDest_;

W
Wu Yi 已提交
35
 public:
Z
zhangjinchao01 已提交
36 37
  explicit SubSequenceLayer(const LayerConfig& config) : Layer(config) {}

Y
Yu Yang 已提交
38 39
  bool init(const LayerMap& layerMap,
            const ParameterMap& parameterMap) override;
Z
zhangjinchao01 已提交
40

Y
Yu Yang 已提交
41 42
  void forward(PassType passType) override;
  void backward(const UpdateCallback& callback = nullptr) override;
Z
zhangjinchao01 已提交
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
};

REGISTER_LAYER(subseq, SubSequenceLayer);

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

  // sequene concatenation layer should have exactly 2 inputs
  CHECK_EQ(3U, inputLayers_.size());

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

  tmpSrc_ =
      Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_);
  tmpDest_ =
      Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_);

  setNeedSequenceInfo(false);
  return true;
}

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

  size_t dim = getSize();

  const Argument& input = getInput(0);
  size_t numSequences1 = input.getNumSequences();
76
  auto startPositions1 = input.sequenceStartPositions->getVector(false);
Z
zhangjinchao01 已提交
77 78 79

  const Argument& offsetSeq = getInput(1);
  size_t numSequences2 = offsetSeq.getNumSequences();
80
  auto startPositions2 = offsetSeq.sequenceStartPositions->getVector(false);
Z
zhangjinchao01 已提交
81 82 83

  const Argument& sizeSeq = getInput(2);
  size_t numSequences3 = sizeSeq.getNumSequences();
84
  auto startPositions3 = sizeSeq.sequenceStartPositions->getVector(false);
Z
zhangjinchao01 已提交
85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100

  CHECK_EQ(dim, input.value->getWidth());

  CHECK_EQ(startPositions1->getData()[numSequences1], input.getBatchSize());
  CHECK_EQ(numSequences1, startPositions1->getSize() - 1);

  CHECK_EQ(startPositions2->getData()[numSequences2], offsetSeq.getBatchSize());
  CHECK_EQ(numSequences2, startPositions2->getSize() - 1);

  CHECK_EQ(startPositions3->getData()[numSequences3], sizeSeq.getBatchSize());
  CHECK_EQ(numSequences3, startPositions3->getSize() - 1);

  CHECK_EQ(numSequences1, numSequences2);
  CHECK_EQ(numSequences2, numSequences3);

  MatrixPtr inputValue = input.value;
Y
yangyaming 已提交
101 102 103 104 105 106 107 108 109 110 111 112 113
  IVectorPtr offsetValue;
  IVectorPtr sizeValue;

  if (useGpu_) {
    // copy to cpu
    IVector::resizeOrCreate(offsetValue, offsetSeq.ids->getSize(), false);
    IVector::resizeOrCreate(sizeValue, sizeSeq.ids->getSize(), false);
    offsetValue->copyFrom(*offsetSeq.ids);
    sizeValue->copyFrom(*sizeSeq.ids);
  } else {
    offsetValue = offsetSeq.ids;
    sizeValue = sizeSeq.ids;
  }
Z
zhangjinchao01 已提交
114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151

  CHECK_EQ(offsetValue->getSize(), numSequences1);
  CHECK_EQ(sizeValue->getSize(), numSequences1);

  int* offsets = offsetValue->getData();
  int* sizes = sizeValue->getData();

  // get total height of output
  size_t height = 0;
  for (size_t seqId = 0; seqId < numSequences1; seqId++) {
    height += sizes[seqId];
  }

  // reset output
  resetOutput(height, dim);

  MatrixPtr outputValue = getOutputValue();

  const int* starts1 = startPositions1->getData();

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

    size_t offsetIn = 0;
    size_t offsetOut = 0;
    size_t size = 0;
    for (size_t seqId = 0; seqId < numSequences1; ++seqId) {
      offsetIn = starts1[seqId] + offsets[seqId];
      size = sizes[seqId];

      outputValue->subMatrix(offsetOut, size, tmpDest_)
          ->assign(*(inputValue->subMatrix(offsetIn, size, tmpSrc_)));

      offsetOut += size;
    }

    // modify the sequenceStartPositions
152 153
    ICpuGpuVector::resizeOrCreate(
        output_.sequenceStartPositions, numSequences1 + 1, false);
Z
zhangjinchao01 已提交
154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185

    int* tgtBuf = output_.sequenceStartPositions->getMutableData(false);
    int offset = 0;
    for (size_t seqId = 0; seqId < numSequences1; ++seqId) {
      tgtBuf[seqId] = offset;
      offset += sizes[seqId];
    }
    tgtBuf[numSequences1] = offset;
  }

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

  /* activation */
  forwardActivation();
}

void SubSequenceLayer::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 inputGrad1 = getInputGrad(0);
  MatrixPtr outputGrad = getOutputGrad();
186
  auto startPositions1 = getInput(0).sequenceStartPositions->getVector(false);
Z
zhangjinchao01 已提交
187 188 189
  size_t numSequences1 = startPositions1->getSize() - 1;
  const int* starts1 = startPositions1->getData();

Y
yangyaming 已提交
190 191 192 193 194 195 196 197 198 199 200 201 202 203 204
  const Argument& offsetSeq = getInput(1);
  const Argument& sizeSeq = getInput(2);
  IVectorPtr offsetValue;
  IVectorPtr sizeValue;

  if (useGpu_) {
    // copy to cpu
    IVector::resizeOrCreate(offsetValue, offsetSeq.ids->getSize(), false);
    IVector::resizeOrCreate(sizeValue, sizeSeq.ids->getSize(), false);
    offsetValue->copyFrom(*offsetSeq.ids);
    sizeValue->copyFrom(*sizeSeq.ids);
  } else {
    offsetValue = offsetSeq.ids;
    sizeValue = sizeSeq.ids;
  }
Z
zhangjinchao01 已提交
205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226

  int* offsets = offsetValue->getData();
  int* sizes = sizeValue->getData();
  {
    AsyncGpuBlock asyncGpuBlock;
    REGISTER_TIMER_INFO("SubSequenceLayerBackward", getName().c_str());

    int offsetIn = 0;
    int offsetOut = 0;
    int size = 0;
    for (size_t seqId = 0; seqId < numSequences1; ++seqId) {
      offsetIn = starts1[seqId] + offsets[seqId];
      size = sizes[seqId];

      inputGrad1->subMatrix(offsetIn, size, tmpDest_)
          ->add(*(outputGrad->subMatrix(offsetOut, size, tmpSrc_)));
      offsetOut += size;
    }
  }
}

}  // namespace paddle