TensorLayer.cpp 4.4 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Z
zhangjinchao01 已提交
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

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 "TensorLayer.h"

#include "paddle/utils/Logging.h"
#include "paddle/utils/Stat.h"

namespace paddle {

REGISTER_LAYER(tensor, TensorLayer);

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

  /* initialize the weightList */
  CHECK_EQ(inputLayers_.size(), 2LU);
  CHECK(parameters_[0]);
  CHECK(!parameters_[1]);

  // Option the parameters
  size_t height = inputLayers_[0]->getSize();
  size_t width = inputLayers_[1]->getSize();
  CHECK_EQ(width * height * getSize(), parameters_[0]->getSize());

  for (size_t i = 0; i < getSize(); ++i) {
    // create a new weight
    Weight* w = new Weight(height, width, parameters_[0], i * width * height);

    // append the new weight to the list
    weights_.emplace_back(w);
  }

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

  return true;
}

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

  /* malloc memory for the output_ if necessary */
  int batchSize = getInputValue(0)->getHeight();
  int size = getSize();

  { resetOutput(batchSize, size); }

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

  /* e1 * W * trans(e2) */ {
    MatrixPtr input1 = getInputValue(0);
    MatrixPtr input2 = getInputValue(1);
    MatrixPtr tmpMat = Matrix::create(input2->getHeight(),
74 75 76
                                      input2->getWidth(),
                                      /* trans= */ false,
                                      input2->useGpu());
Z
zhangjinchao01 已提交
77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104
    REGISTER_TIMER_INFO("TensorFwMulTimer", getName().c_str());
    for (size_t i = 0; i < getSize(); ++i) {
      MatrixPtr weights = weights_[i]->getW();
      tmpMat->mul(input1, weights, 1, 0);
      outV->rowDotMul(i, *tmpMat, *input2);
    }
  }

  /* activation */ { forwardActivation(); }
}

void TensorLayer::backward(const UpdateCallback& callback) {
  /* Do derivation */ { backwardActivation(); }

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

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

  bool syncFlag = hl_get_sync_flag();

  /* Calculate the W-gradient for the current layer */
  MatrixPtr input1 = getInputValue(0);
  MatrixPtr input2 = getInputValue(1);
  MatrixPtr oGrad = getOutputGrad();
  MatrixPtr tmpMat = Matrix::create(input1->getHeight(),
105 106 107
                                    input1->getWidth(),
                                    /* trans= */ false,
                                    input1->useGpu());
Z
zhangjinchao01 已提交
108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145

  /* trans(grad * e1) * e2 */ {
    REGISTER_TIMER_INFO("TensorGradMulTimer", getName().c_str());
    for (size_t i = 0; i < getSize(); ++i) {
      if (weights_[i]->getWGrad()) {
        tmpMat->rowScale(i, *input1, *oGrad);
        MatrixPtr input1_T = tmpMat->getTranspose();
        weights_[i]->getWGrad()->mul(input1_T, input2, 1, 1);
      }
    }
  }

  hl_set_sync_flag(false);

  /* Calculate the input layers error */ {
    MatrixPtr preGrad1 = getInputGrad(0);
    MatrixPtr preGrad2 = getInputGrad(1);

    REGISTER_TIMER_INFO("TensorBpMulTimer", getName().c_str());
    for (size_t i = 0; i < getSize(); ++i) {
      MatrixPtr weights = weights_[i]->getW();

      if (NULL != preGrad1) { /* (grad * e2) * trans(W) */
        tmpMat->rowScale(i, *input2, *oGrad);
        MatrixPtr weights_T = weights->getTranspose();
        preGrad1->mul(tmpMat, weights_T, 1, 1);
      }
      if (NULL != preGrad2) { /* (grad * e1) * W */
        tmpMat->rowScale(i, *input1, *oGrad);
        preGrad2->mul(tmpMat, weights, 1, 1);
      }
    }
  }
  hl_set_sync_flag(syncFlag);
  parameters_[0]->incUpdate(callback);
}

}  // namespace paddle