CudnnBatchNormLayer.cpp 5.1 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

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 "CudnnBatchNormLayer.h"

namespace paddle {

REGISTER_LAYER(cudnn_batch_norm, CudnnBatchNormLayer);

const double CudnnBatchNormLayer::EPS = 1E-5;

bool CudnnBatchNormLayer::init(const LayerMap& layerMap,
                               const ParameterMap& parameterMap) {
  /* Initialize the basic parent class */
  if (!BatchNormBaseLayer::init(layerMap, parameterMap)) return false;
  CHECK(useGpu_) << "CudnnBatchNorm only support GPU";

  hl_create_tensor_descriptor(&ioDesc_);
  hl_create_tensor_descriptor(&bnParamDesc_);
  hl_tensor_reshape(bnParamDesc_, 1, channels_, 1, 1);

  return true;
}

void CudnnBatchNormLayer::reshape(int batchSize) {
  hl_tensor_reshape(ioDesc_, batchSize, channels_, imageH_, imageW_);
}

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

  int batchSize = getInputValue(0)->getHeight();
  calFeatureMapSize();
  reshape(batchSize);
  resetOutput(batchSize, getInputValue(0)->getWidth());

  // for testing in training peroid.
  useGlobalStats_ = (passType == PASS_TEST);
  if (passType == PASS_TEST && config_.has_use_global_stats()) {
    useGlobalStats_ = config_.use_global_stats();
  }

  real* input = getInputValue(0)->getData();
  real* output = getOutputValue()->getData();
  real* gamma = weight_->getW()->getData();
  real* beta = biases_->getW()->getData();
  real* movingMean = movingMean_->getW()->getData();
  real* movingVar = movingVar_->getW()->getData();

  if (!useGlobalStats_) {
    REGISTER_TIMER_INFO("CudnnBatchFwTimer", getName().c_str());
    real* savedMean = savedMean_->getData();
    real* savedInvVar = savedInvVar_->getData();
67 68 69 70
    hl_batch_norm_forward_training(ioDesc_,
                                   input,
                                   ioDesc_,
                                   output,
Z
zhangjinchao01 已提交
71
                                   bnParamDesc_,
72 73 74 75 76 77 78 79
                                   gamma,
                                   beta,
                                   1.0 - movingAvgFraction_,
                                   movingMean,
                                   movingVar,
                                   EPS,
                                   savedMean,
                                   savedInvVar);
Z
zhangjinchao01 已提交
80 81
  } else {
    // used movingMean and movingVar in testing
82 83 84 85 86 87 88 89 90 91
    hl_batch_norm_forward_inference(ioDesc_,
                                    input,
                                    ioDesc_,
                                    output,
                                    bnParamDesc_,
                                    gamma,
                                    beta,
                                    movingMean,
                                    movingVar,
                                    EPS);
Z
zhangjinchao01 已提交
92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130
  }

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

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

  real* input = getInputValue(0)->getData();
  real* outGrad = getOutputGrad()->getData();
  real* inGrad = getInputGrad(0)->getData();
  real* gamma = weight_->getW()->getData();
  real* savedMean = savedMean_->getData();
  real* savedInvVar = savedInvVar_->getData();

  auto create = [](MatrixPtr& m, size_t h, size_t w, real** p) {
    Matrix::resizeOrCreate(m, h, w, false, true);
    m->zeroMem();
    *p = m->getData();
  };

  real* gammaGrad = nullptr;
  real* betaGrad = nullptr;
  if (weight_->getWGrad()) {
    gammaGrad = weight_->getWGrad()->getData();
  } else {
    create(tmpWGrad_, 1, channels_, &gammaGrad);
  }
  if (biases_ && biases_->getWGrad()) {
    betaGrad = biases_->getWGrad()->getData();
  } else {
    create(tmpBiasGrad_, 1, channels_, &betaGrad);
  }
131

132 133 134 135 136 137 138 139 140 141 142 143 144
  hl_batch_norm_backward(ioDesc_,
                         input,
                         ioDesc_,
                         outGrad,
                         ioDesc_,
                         inGrad,
                         bnParamDesc_,
                         gamma,
                         gammaGrad,
                         betaGrad,
                         EPS,
                         savedMean,
                         savedInvVar);
Z
zhangjinchao01 已提交
145 146 147 148 149 150 151 152 153 154 155 156 157 158

  {
    REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
    biases_->getParameterPtr()->incUpdate(callback);
    weight_->getParameterPtr()->incUpdate(callback);
  }
}

CudnnBatchNormLayer::~CudnnBatchNormLayer() {
  hl_destroy_tensor_descriptor(ioDesc_);
  hl_destroy_tensor_descriptor(bnParamDesc_);
}

}  // namespace paddle