Remove the code of ExpandConvTransLayer.

paddle/gserver/layers/ExpandConvBaseLayer.cpp

@@ -22,26 +22,8 @@ bool ExpandConvBaseLayer::init(const LayerMap &layerMap,
   /* Initialize the basic convolutional parent class */
   ConvBaseLayer::init(layerMap, parameterMap);
 
-  /* 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.
-   * */
-
-  /* Initialize the projection */
   for (auto &inputConfig : config_.inputs()) {
     const ConvConfig &conf = inputConfig.conv_conf();
-    int numFilters = isDeconv_ ? conf.channels() : numFilters_;
-    subM_.push_back(numFilters / conf.groups());
-    subN_.push_back(conf.output_x() *
-                    (conf.has_output_y() ? conf.output_y() : conf.output_x()));
-    int channel = isDeconv_ ? numFilters_ : conf.channels();
-    subK_.push_back(
-        channel * conf.filter_size() *
-        (conf.has_filter_size_y() ? conf.filter_size_y() : conf.filter_size()) /
-        conf.groups());
     /* Consistent caffe mode for multiple input */
     caffeMode_ = conf.caffe_mode();
   }
@@ -54,17 +36,9 @@ bool ExpandConvBaseLayer::init(const LayerMap &layerMap,
 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;
 }
 
-void ExpandConvBaseLayer::resetExpandInput(size_t height, size_t width) {
-  Matrix::resizeOrCreate(expandInput_, height, width, false, useGpu_);
-}
-
 void ExpandConvBaseLayer::addSharedBias() {
   size_t mapW = getOutputSize() / numFilters_;
   size_t mapH = getOutputValue()->getElementCnt() / mapW;
@@ -101,173 +75,6 @@ void ExpandConvBaseLayer::addUnsharedBias() {
   outValue->addBias(*bias, 1.0f);
 }
 
-void ExpandConvBaseLayer::expandOneFrame(MatrixPtr image,
-                                         size_t startIdx,
-                                         int inIdx) {
-  int channel = isDeconv_ ? numFilters_ : channels_[inIdx];
-
-  resetExpandInput(subK_[inIdx] * groups_[inIdx], subN_[inIdx]);
-
-  CHECK_EQ(image->getWidth(),
-           static_cast<size_t>(imgSizeH_[inIdx] * imgSizeW_[inIdx] * channel));
-
-  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,
-                           filterSizeY_[inIdx],
-                           filterSize_[inIdx],
-                           strideY_[inIdx],
-                           stride_[inIdx],
-                           paddingY_[inIdx],
-                           padding_[inIdx],
-                           outputH_[inIdx],
-                           outputW_[inIdx]);
-  imageTmp->clear();
-}
-
-void ExpandConvBaseLayer::expandFwdOnce(MatrixPtr image,
-                                        MatrixPtr out,
-                                        int inIdx,
-                                        int startIdx) {
-  int subM = subM_[inIdx];
-  int subN = subN_[inIdx];
-  int subK = subK_[inIdx];
-
-  expandOneFrame(image, startIdx, inIdx);
-
-  int numFilters = isDeconv_ ? channels_[inIdx] : numFilters_;
-
-  real *outData = out->getData() + startIdx * subN * numFilters;
-
-  real *wgtData = weights_[inIdx]->getW()->getData();
-  real *expInData = expandInput_->getData();
-  for (int g = 0; g < groups_[inIdx]; ++g) {
-    MatrixPtr A =
-        Matrix::create(wgtData, subM, subK, false, 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;
-  }
-}
-
-void ExpandConvBaseLayer::bpropActs(MatrixPtr out,
-                                    MatrixPtr image,
-                                    int inpIdx) {
-  int channel = isDeconv_ ? numFilters_ : channels_[inpIdx];
-
-  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);
-
-  real *localGradData = out->getData();
-  real *tgtGradData = image->getData();
-  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, subM, subK, true, 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,
-                     filterSizeY_[inpIdx],
-                     filterSize_[inpIdx],
-                     strideY_[inpIdx],
-                     stride_[inpIdx],
-                     paddingY_[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;
-  }
-}
-
-void ExpandConvBaseLayer::bpropWeights(MatrixPtr image,
-                                       MatrixPtr out,
-                                       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, true, useGpu_);
-      MatrixPtr B = Matrix::create(gradData, subM, subN, false, useGpu_);
-      MatrixPtr C = Matrix::create(wGradData, subM, subK, false, useGpu_);
-      C->mul(*B, *A, 1, 1);
-
-      A->clear();
-      B->clear();
-      C->clear();
-      gradData += subM * subN;
-      wGradData += subK * subM;
-      expandInData += subK * subN;
-    }
-  }
-}
-
 void ExpandConvBaseLayer::bpropSharedBias(MatrixPtr biases, MatrixPtr v) {
   size_t mapW = getOutputSize() / numFilters_;
   size_t mapH = v->getElementCnt() / mapW;

paddle/gserver/layers/ExpandConvBaseLayer.h

@@ -26,19 +26,6 @@ namespace paddle {
  */
 class ExpandConvBaseLayer : public ConvBaseLayer {
 protected:
-  /// For expand convolution.
-  /// subM_ = numFilters_ / groups_.
-  IntV subM_;
-  /// subN_ = outputH_ * outputW_.
-  IntV subN_;
-  /// subK_ = channels_ * filterPixels_ * groups_.
-  IntV subK_;
-
-  /*The expandInput_ and transOutValue_ are used for CPU expand conv calc
-   * Expand one sample at a time. shape:
-   * (numChannels * filterPixels_, outputSizeH * outputSizeW)
-   * */
-  MatrixPtr expandInput_;
   /// The transpose of output, which is an auxiliary matrix.
   MatrixPtr transOutValue_;
 
@@ -52,10 +39,6 @@ public:
             const ParameterMap& parameterMap) override;
 
   size_t getOutputSize();
-  /**
-   * Create or resize expandInput_.
-   */
-  void resetExpandInput(size_t height, size_t width);
 
   /**
    * Add shared bias.
@@ -66,20 +49,9 @@ public:
    * Add unshared bias.
    */
   void addUnsharedBias();
-  /**
-   * Expand one input sample.
-   */
-  void expandOneFrame(MatrixPtr image, size_t startIdx, int inIdx);
-
-  /**
-   * Expand one input sample and perform matrix multiplication.
-   */
-  void expandFwdOnce(MatrixPtr image, MatrixPtr out, int inIdx, int startIdx);
 
   void bpropSharedBias(MatrixPtr biases, MatrixPtr v);
   void bpropBiases(MatrixPtr v);
-  void bpropWeights(MatrixPtr image, MatrixPtr out, int inpIdx);
-  void bpropActs(MatrixPtr image, MatrixPtr out, int inpIdx);
 };
 
 }  // namespace paddle

paddle/gserver/layers/ExpandConvTransLayer.cpp

-/* Copyright (c) 2016 PaddlePaddle Authors. 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 "ExpandConvTransLayer.h"
-#include "paddle/utils/Logging.h"
-#include "paddle/utils/Stat.h"
-
-/* The implementation of the convTransLayer is basically a swap of forward and
- * backward of the original convLayer.
- * The variable naming follows the convention of the convLayer.
- * */
-
-namespace paddle {
-
-// REGISTER_LAYER(exconvt, ExpandConvTransLayer);
-
-bool ExpandConvTransLayer::init(const LayerMap &layerMap,
-                                const ParameterMap &parameterMap) {
-  /* Initialize the basic convolutional parent class */
-  ExpandConvBaseLayer::init(layerMap, parameterMap);
-
-  return true;
-}
-
-void ExpandConvTransLayer::forward(PassType passType) {
-  Layer::forward(passType);
-
-  /* malloc memory for the output_ if necessary */
-  int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
-  resetOutput(batchSize, getOutputSize());
-
-  MatrixPtr output = nullptr;
-  for (size_t i = 0; i < inputLayers_.size(); ++i) {
-    LayerPtr prevLayer = getPrev(i);
-    output = prevLayer->getOutputValue();
-    REGISTER_TIMER_INFO("shrinkFwd", getName().c_str());
-    bpropActs(output, getOutputValue(), i);
-  }
-
-  /* add the bias-vector */
-  if (biases_.get()) {
-    if (sharedBiases_) {
-      addSharedBias();
-    } else {
-      addUnsharedBias();
-    }
-  }
-
-  /* activation */
-  forwardActivation();
-}
-
-void ExpandConvTransLayer::backward(const UpdateCallback &callback) {
-  backwardActivation();
-
-  MatrixPtr imageGrad = getOutputGrad();
-  if (biases_ && biases_->getWGrad()) {
-    bpropBiases(imageGrad);
-    /* Increasing the number of gradient */
-    biases_->getParameterPtr()->incUpdate(callback);
-  }
-
-  for (size_t i = 0; i < inputLayers_.size(); ++i) {
-    /* First, calculate the input layers error */
-    for (size_t off = 0; off < imageGrad->getHeight(); off++) {
-      if (getPrev(i)->getOutputGrad()) {
-        expandFwdOnce(imageGrad, getPrev(i)->getOutputGrad(), i, off);
-      }
-    }
-    if (weights_[i]->getWGrad()) {
-      /* Then, calculate the W-gradient for the current layer */
-      bpropWeights(imageGrad, getPrev(i)->getOutputValue(), i);
-      /* Increasing the number of gradient */
-      weights_[i]->getParameterPtr()->incUpdate(callback);
-    }
-  }
-}
-
-}  // namespace paddle

paddle/gserver/layers/ExpandConvTransLayer.h

-/* Copyright (c) 2016 PaddlePaddle Authors. 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. */
-
-#pragma once
-
-#include <vector>
-#include "ExpandConvBaseLayer.h"
-#include "paddle/math/Matrix.h"
-
-namespace paddle {
-
-/**
- * @brief A subclass of convolution layer.
- * This layer expands input and use matrix multiplication to
- * calculate convolution transpose (deconv) operation.
- *
- * The config file api is img_conv_layer with flag trans=True.
- */
-class ExpandConvTransLayer : public ExpandConvBaseLayer {
-public:
-  explicit ExpandConvTransLayer(const LayerConfig& config)
-      : ExpandConvBaseLayer(config) {}
-
-  ~ExpandConvTransLayer() {}
-
-  bool init(const LayerMap& layerMap,
-            const ParameterMap& parameterMap) override;
-
-  void forward(PassType passType) override;
-  void backward(const UpdateCallback& callback) override;
-};
-
-}  // namespace paddle

paddle/gserver/tests/test_BatchNorm.cpp

@@ -17,7 +17,6 @@ limitations under the License. */
 #include <vector>
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
-#include "paddle/gserver/layers/ExpandConvTransLayer.h"
 #include "paddle/trainer/Trainer.h"
 #include "paddle/utils/GlobalConstants.h"
 

paddle/gserver/tests/test_ConvTrans.cpp

@@ -17,7 +17,6 @@ limitations under the License. */
 #include <vector>
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
-#include "paddle/gserver/layers/ExpandConvTransLayer.h"
 #include "paddle/math/MathUtils.h"
 #include "paddle/trainer/Trainer.h"
 #include "paddle/utils/GlobalConstants.h"

paddle/gserver/tests/test_ConvUnify.cpp

@@ -17,7 +17,6 @@ limitations under the License. */
 #include <vector>
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
-#include "paddle/gserver/layers/ExpandConvTransLayer.h"
 #include "paddle/math/MathUtils.h"
 #include "paddle/trainer/Trainer.h"
 #include "paddle/utils/GlobalConstants.h"