ConcatenateLayer.cpp 5.7 KB
Newer Older
Z
zhangjinchao01 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 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
/* 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. */

#include "paddle/utils/Stat.h"
#include "Layer.h"
#include "Projection.h"

namespace paddle {

/**
 * A concatenate layer has multiple input layers. It concatenates rows of
 * each input as one row for the output of this layer and apply activation.
 */
class ConcatenateLayer : public Layer {
public:
  explicit ConcatenateLayer(const LayerConfig& config) : Layer(config) {}

  ~ConcatenateLayer() {}

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

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

REGISTER_LAYER(concat, ConcatenateLayer);

bool ConcatenateLayer::init(const LayerMap& layerMap,
                            const ParameterMap& parameterMap) {
  /* Initialize the basic parent class */
  if (!Layer::init(layerMap, parameterMap)) return false;

  CHECK(!biasParameter_);

  return true;
}

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

  int batchSize = getInput(0).getBatchSize();
  int size = getSize();
  reserveOutput(batchSize, size);

  const MatrixPtr& out = getOutputValue();
  int offset = 0;

  for (size_t i = 0; i != inputLayers_.size(); ++i) {
    const MatrixPtr& in = getInputValue(i);
    size_t inSize = in->getWidth();
    out->assignAtOffset(*in, offset);
    offset += inSize;
  }
  CHECK_EQ(size, offset);

  /* activation */ {
    REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str());
    forwardActivation();
  }
}

void ConcatenateLayer::backward(const UpdateCallback& callback) {
  (void)callback;

  /* Do activation */ {
    REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str());
    backwardActivation();
  }

  const MatrixPtr& out = getOutputGrad();
  int offset = 0;

  for (size_t i = 0; i != inputLayers_.size(); ++i) {
    const MatrixPtr& in = getInputGrad(i);
    size_t inSize = getInputValue(i)->getWidth();
    if (in) {
      in->addAtOffset(*out, offset);
    }
    offset += inSize;
  }
}

/**
 * concat2 layer is like concat layer, but each input layer was
 * processed by a Projection.
 */
class ConcatenateLayer2 : public Layer {
public:
100
  explicit ConcatenateLayer2(const LayerConfig& config) : Layer(config) {}
Z
zhangjinchao01 已提交
101 102 103 104 105 106 107 108 109 110 111 112

  ~ConcatenateLayer2() {}

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

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

protected:
  std::vector<std::unique_ptr<Projection>> projections_;
  std::vector<Argument> projOutput_;
  std::vector<std::pair<size_t, size_t>> projCol_;
113 114
  bool sharedBias_;
  std::unique_ptr<Weight> biases_;
Z
zhangjinchao01 已提交
115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131
};

REGISTER_LAYER(concat2, ConcatenateLayer2);

bool ConcatenateLayer2::init(const LayerMap& layerMap,
                             const ParameterMap& parameterMap) {
  /* Initialize the basic parent class */
  if (!Layer::init(layerMap, parameterMap)) return false;

  CHECK_EQ(inputLayers_.size(), parameters_.size());
  projections_.reserve(inputLayers_.size());
  projCol_.reserve(inputLayers_.size());
  projOutput_.resize(inputLayers_.size());

  size_t startCol = 0;
  size_t endCol = 0;
  for (size_t i = 0; i < inputLayers_.size(); i++) {
132 133
    projections_.emplace_back(Projection::create(
        config_.inputs(i).proj_conf(), parameters_[i], useGpu_));
Z
zhangjinchao01 已提交
134 135 136 137 138 139 140

    endCol += projections_[i]->getOutputSize();
    projCol_.push_back(std::make_pair(startCol, endCol));
    startCol = endCol;
  }
  CHECK_EQ(getSize(), endCol);

141 142 143 144 145 146 147
  /* initialize biases_ */
  if (biasParameter_.get() != NULL) {
    sharedBias_ = config_.shared_biases();
    size_t psize = config_.bias_size();
    biases_ = std::unique_ptr<Weight>(new Weight(1, psize, biasParameter_));
  }

Z
zhangjinchao01 已提交
148 149 150 151 152 153 154 155 156 157 158 159 160 161
  return true;
}

void ConcatenateLayer2::forward(PassType passType) {
  Layer::forward(passType);

  int batchSize = getInput(0).getBatchSize();
  int size = getSize();
  resetOutput(batchSize, size);

  for (size_t i = 0; i < projections_.size(); i++) {
    size_t startCol = projCol_[i].first;
    size_t endCol = projCol_[i].second;
    projOutput_[i].value = output_.value->subColMatrix(startCol, endCol);
D
dangqingqing 已提交
162 163 164
    if (output_.grad) {
      projOutput_[i].grad = output_.grad->subColMatrix(startCol, endCol);
    }
Z
zhangjinchao01 已提交
165 166
  }

167 168 169 170 171 172 173 174 175 176 177
  {
    AsyncGpuBlock block;
    for (size_t i = 0; i != inputLayers_.size(); ++i) {
      projections_[i]->forward(&getInput(i), &projOutput_[i], passType);
    }
  }

  /* add the bias-vector */
  if (biases_) {
    REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str());
    output_.value->addBias(*(biases_->getW()), 1, sharedBias_);
Z
zhangjinchao01 已提交
178 179 180 181 182 183 184 185 186 187 188 189 190 191
  }

  /* activation */ {
    REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str());
    forwardActivation();
  }
}

void ConcatenateLayer2::backward(const UpdateCallback& callback) {
  /* Do activation */ {
    REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str());
    backwardActivation();
  }

192 193 194 195 196 197 198
  AsyncGpuBlock block;
  if (biases_ && biases_->getWGrad()) {
    REGISTER_TIMER_INFO("Concat2BpBiasTimer", getName().c_str());
    biases_->getWGrad()->collectBias(*getOutputGrad(), 1, sharedBias_);
    biases_->getParameterPtr()->incUpdate(callback);
  }

Z
zhangjinchao01 已提交
199 200 201 202 203 204 205 206
  for (size_t i = 0; i != inputLayers_.size(); ++i) {
    if (projections_[i]) {
      projections_[i]->backward(callback);
    }
  }
}

}  // namespace paddle