ExpandConvBaseLayer.cpp 8.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2016 Baidu, Inc. 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. */


16 17
#include "ExpandConvBaseLayer.h"

18 19 20
#include "paddle/utils/Logging.h"
namespace paddle {

21
bool ExpandConvBaseLayer::init(const LayerMap &layerMap,
22 23 24 25
                           const ParameterMap &parameterMap) {
  /* Initialize the basic convolutional parent class */
  ConvBaseLayer::init(layerMap, parameterMap);

26 27 28 29 30 31 32
  /* The class fields channels_ and numFilters_ are the same as in the config
   * i.e., channels_ is the for the input and numFilters_ is for the output
   *
   * But in order for the variables in convTrans having the same semantic
   * meaning as in conv, we need to swap channels_ and numFilters here for
   * convTrans, and in other functions too.
   * */
33
  int channel;
34
  int numFilters;
35 36 37
  /* Initialize the projection */
  for (auto &inputConfig : config_.inputs()) {
    const ConvConfig &conf = inputConfig.conv_conf();
38 39
    numFilters = isDeconv_ ? conf.channels() : numFilters_;
    subM_.push_back(numFilters / conf.groups());
40
    subN_.push_back(conf.output_x() * conf.output_x());
41
    channel = isDeconv_ ? numFilters_ : conf.channels();
42 43 44 45 46 47
    subK_.push_back(channel * conf.filter_size() * conf.filter_size() /
                    conf.groups());
    /* Consistent caffe mode for multiple input */
    caffeMode_ = conf.caffe_mode();
  }

48 49
  getOutputSize();

50 51 52
  return true;
}

53 54 55 56 57 58 59 60 61 62
size_t ExpandConvBaseLayer::getOutputSize() {
  CHECK_NE(inputLayers_.size(), 0UL);
  size_t layerSize = ConvBaseLayer::calOutputSize();
  subN_.clear();
  for (size_t i = 0; i < inputLayers_.size(); i++) {
    subN_.push_back(outputH_[i] * outputW_[i]);
  }
  return layerSize;
}

63
void ExpandConvBaseLayer::resetExpandInput(size_t height, size_t width) {
64 65 66
  Matrix::resizeOrCreate(expandInput_, height, width, false, useGpu_);
}

67
void ExpandConvBaseLayer::addSharedBias() {
68
  size_t mapW = getOutputSize() / numFilters_;
69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90
  size_t mapH = getOutputValue()->getElementCnt() / mapW;
  MatrixPtr out =
      Matrix::create(getOutputValue()->getData(), mapH, mapW, false, useGpu_);

  Matrix::resizeOrCreate(transOutValue_, mapW, mapH, false, useGpu_);

  out->transpose(transOutValue_, false);  // false means no memory allocation
  transOutValue_->reshape(transOutValue_->getElementCnt() / numFilters_,
                          numFilters_);

  MatrixPtr bias =
      Matrix::create(biases_->getW()->getData(), 1,
                     biases_->getW()->getElementCnt(), false, useGpu_);
  transOutValue_->addBias(*bias, 1.0f);

  transOutValue_->reshape(mapW, mapH);
  transOutValue_->transpose(out, false);  // false means no memory allocation

  out->clear();
  bias->clear();
}

91
void ExpandConvBaseLayer::addUnsharedBias() {
92 93 94 95 96 97 98 99
  MatrixPtr outValue = getOutputValue();
  MatrixPtr bias =
      Matrix::create(biases_->getW()->getData(), 1,
                     biases_->getW()->getElementCnt(), false, useGpu_);
  outValue->addBias(*bias, 1.0f);
}


100
void ExpandConvBaseLayer::expandOneFrame(MatrixPtr image, size_t startIdx,
101
                                     int inIdx) {
102
  int channel = isDeconv_ ? numFilters_ : channels_[inIdx];
103 104 105 106 107 108 109 110 111 112 113 114 115 116

  resetExpandInput(subK_[inIdx] * groups_[inIdx], subN_[inIdx]);
  real *imgData = image->getData() + startIdx * image->getWidth();
  MatrixPtr imageTmp = Matrix::create(
      imgData, 1, imgSizeH_[inIdx] * imgSizeW_[inIdx] * channel, false,
      useGpu_);
  expandInput_->convExpand(*imageTmp, imgSizeH_[inIdx], imgSizeW_[inIdx],
                           channel, filterSize_[inIdx],
                           filterSize_[inIdx], stride_[inIdx], stride_[inIdx],
                           padding_[inIdx], padding_[inIdx],
                           outputH_[inIdx], outputW_[inIdx]);
  imageTmp->clear();
}

117
void ExpandConvBaseLayer::expandFwdOnce(MatrixPtr image, MatrixPtr out,
118 119 120 121 122 123 124
                                     int inIdx, int startIdx) {
  int subM = subM_[inIdx];
  int subN = subN_[inIdx];
  int subK = subK_[inIdx];

  expandOneFrame(image, startIdx, inIdx);

125
  int numFilters = isDeconv_ ? channels_[inIdx] : numFilters_;
126 127

  real *outData =
128
      out->getData() + startIdx * subN * numFilters;
129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147

  real *wgtData = weights_[inIdx]->getW()->getData();
  real *expInData = expandInput_->getData();
  for (int g = 0; g < groups_[inIdx]; ++g) {
    MatrixPtr A =
        Matrix::create(wgtData, subK, subM, true, useGpu_);  // mark transpose
    MatrixPtr B = Matrix::create(expInData, subK, subN, false, useGpu_);
    MatrixPtr C = Matrix::create(outData, subM, subN, false, useGpu_);
    C->mul(A, B, 1, 1);

    A->clear();
    B->clear();
    C->clear();
    wgtData += subK * subM;
    expInData += subK * subN;
    outData += subM * subN;
  }
}

148 149
void ExpandConvBaseLayer::bpropActs(MatrixPtr out, MatrixPtr image,
                                    int inpIdx) {
150
  int channel = isDeconv_ ? numFilters_ : channels_[inpIdx];
151 152 153 154 155 156 157 158 159

  int subM = subM_[inpIdx];
  int subN = subN_[inpIdx];
  int subK = subK_[inpIdx];
  size_t batchSize = image->getHeight();

  /* reset the expand-grad memory */
  resetExpandInput(subK * groups_[inpIdx], subN);

160 161
  real *localGradData = out->getData();
  real *tgtGradData = image->getData();
162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202
  for (size_t n = 0; n < batchSize; n++) {
    real *wgtData = weights_[inpIdx]->getW()->getData();
    real *expandInData = expandInput_->getData();

    for (int g = 0; g < groups_[inpIdx]; g++) {
      // create temporary matrix
      MatrixPtr C = Matrix::create(expandInData, subK, subN, false, useGpu_);
      MatrixPtr B = Matrix::create(localGradData, subM, subN, false, useGpu_);
      MatrixPtr A = Matrix::create(wgtData, subK, subM, false, useGpu_);
      C->mul(A, B);  // mul

      // clear the temporary matrix
      A->clear();
      B->clear();
      C->clear();

      expandInData += subK * subN;
      localGradData += subM * subN;
      wgtData += subK * subM;
    }

    // shrink one frame outGrad
    MatrixPtr oneGradTmp = Matrix::create(
        expandInput_->getData(), subK * groups_[inpIdx], subN, false, useGpu_);
    MatrixPtr vTmp = Matrix::create(
        tgtGradData, 1,
        imgSizeH_[inpIdx] * imgSizeW_[inpIdx] * channel, false,
        useGpu_);
    vTmp->convShrink(*oneGradTmp, imgSizeH_[inpIdx], imgSizeW_[inpIdx],
                     channel, filterSize_[inpIdx],
                     filterSize_[inpIdx], stride_[inpIdx], stride_[inpIdx],
                     padding_[inpIdx], padding_[inpIdx],
                     outputH_[inpIdx], outputW_[inpIdx], 1.0f, 1.0f);
    vTmp->clear();
    oneGradTmp->clear();

    // move the data-pointer
    tgtGradData += imgSizeH_[inpIdx] * imgSizeW_[inpIdx] * channel;
  }
}

203
void ExpandConvBaseLayer::bpropWeights(MatrixPtr image, MatrixPtr out,
204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237
                                    int inpIdx) {
  MatrixPtr weightGrad = weights_[inpIdx]->getWGrad();

  int subM = subM_[inpIdx];
  int subN = subN_[inpIdx];
  int subK = subK_[inpIdx];
  size_t batchSize = image->getHeight();
  resetExpandInput(subK * groups_[inpIdx], subN);

  real *gradData = out->getData();

  for (size_t n = 0; n < batchSize; n++) {  // frame by frame
    // expand
    expandOneFrame(image, n, inpIdx);
    real *wGradData = weightGrad->getData();
    real *expandInData = expandInput_->getData();

    // expand-mul one-group by one
    for (int g = 0; g < groups_[inpIdx]; g++) {
      MatrixPtr A = Matrix::create(expandInData, subK, subN, false, useGpu_);
      MatrixPtr B = Matrix::create(gradData, subM, subN, true, useGpu_);
      MatrixPtr C = Matrix::create(wGradData, subK, subM, false, useGpu_);
      C->mul(A, B, 1, 1);

      A->clear();
      B->clear();
      C->clear();
      gradData += subM * subN;
      wGradData += subK * subM;
      expandInData += subK * subN;
    }
  }
}

238
void ExpandConvBaseLayer::bpropSharedBias(MatrixPtr biases, MatrixPtr v) {
239
  size_t mapW = getOutputSize() / numFilters_;
240 241 242 243 244 245 246 247 248 249 250
  size_t mapH = v->getElementCnt() / mapW;
  MatrixPtr vTmp = Matrix::create(v->getData(), mapH, mapW, false, useGpu_);

  Matrix::resizeOrCreate(transOutValue_, mapW, mapH, false, useGpu_);

  vTmp->transpose(transOutValue_, false);  // false means no memory allocation
  transOutValue_->reshape(transOutValue_->getElementCnt() / numFilters_,
                          numFilters_);
  biases->collectBias(*transOutValue_, 1.0f);
}

251
void ExpandConvBaseLayer::bpropBiases(MatrixPtr v) {
252 253 254 255 256 257 258 259 260 261 262 263
  MatrixPtr biases =
      Matrix::create(biases_->getWGrad()->getData(), 1,
                     biases_->getWGrad()->getElementCnt(), false, useGpu_);
  if (sharedBiases_) {
    bpropSharedBias(biases, v);
  } else {
    biases->collectBias(*v, 1.0f);
  }
  biases->clear();
}

}  // namespace paddle