Refactor ExpandConvTransLayer to share codes with ExpandConvLayer

paddle/gserver/layers/ConvBaseLayer.cpp

@@ -21,6 +21,12 @@ bool ConvBaseLayer::init(const LayerMap& layerMap,
   /* Initialize the basic parent class */
   Layer::init(layerMap, parameterMap);
 
+  if (config_.type() == "exconv" || config_.type() == "cudnn_conv") {
+    isConv_ = true;
+  } else {
+    isConv_ = false;
+  }
+
   /* Initialize the convolutional layer parameter */
   numFilters_ = config_.num_filters();
   sharedBiases_ = config_.shared_biases();
@@ -88,6 +94,7 @@ size_t ConvBaseLayer::calOutputSize() {
   getOutput().setFrameWidth(outputW_[0]);
   layerSize = outputH_[0] * outputW_[0] * size_t(numFilters_);
   return layerSize;
+
 }
 
 }  // namespace paddle

paddle/gserver/layers/ConvBaseLayer.h

@@ -28,6 +28,9 @@ class ConvBaseLayer : public Layer {
 protected:
   typedef std::vector<int> IntV;
 
+  /// True if it's convolution layer, false if it's deconv layer
+  bool isConv_;
+
   /// The number of filters.
   int numFilters_;
   /// The x dimension of the padding.
@@ -75,6 +78,13 @@ protected:
   /// of output size.
   bool caffeMode_;
 
+  /*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_;
+
 public:
   explicit ConvBaseLayer(const LayerConfig& config) : Layer(config) {}
 
@@ -88,6 +98,63 @@ public:
   virtual size_t calOutputSize();
 
   Weight& getWeight(int idx) { return *weights_[idx]; }
+
+  /**
+   * Calculate output size based on caffeMode_.
+   * - input(+padding): 0123456789
+   * - imageSize(+padding) = 10;
+   * - filterSize = 3;
+   * - stride = 2;
+   * - caffeMode_ is true:
+       - output: (012), (234), (456), (678)
+       - outputSize = 4;
+   * - caffeMode_ is false:
+   *   - output: (012), (234), (456), (678), (9)
+   *   - outputSize = 5;
+   */
+  int outputSize(int imageSize, int filterSize, int padding, int stride) {
+    int outputSize;
+    if (!caffeMode_) {
+     outputSize =
+          (imageSize - filterSize + 2 * padding + stride - 1) / stride + 1;
+    } else {
+      outputSize = (imageSize - filterSize + 2 * padding) / stride + 1;
+    }
+    CHECK_GE(outputSize, 1);
+    return outputSize;
+  }
+
+  int imageSize(int outputSize, int filterSize, int padding, int stride) {
+    int imageSize;
+    if (!caffeMode_) {
+     imageSize =
+         (outputSize - 1) * stride + filterSize - 2 * padding - stride + 1;
+    } else {
+     imageSize = (outputSize - 1) * stride + filterSize - 2 * padding;
+    }
+    CHECK_GE(imageSize, 1);
+    return imageSize;
+  }
+
+  /**
+   * Create or resize expandInput_.
+   */
+  void resetExpandInput(size_t height, size_t width);
+
+  /**
+   * Create or resize transOutValue_.
+   */
+  void resetConvOutput(size_t batchSize, int inIdx);
+
+  /**
+   * Add shared bias.
+   */
+  void addSharedBias();
+
+  /**
+   * Add unshared bias.
+   */
+  void addUnsharedBias();
 };
 
 }  // namespace paddle

paddle/gserver/layers/ExpandConvLayer.cpp

@@ -63,14 +63,6 @@ size_t ExpandConvLayer::getOutputSize() {
   return layerSize;
 }
 
-void ExpandConvLayer::resetExpandInput(size_t height, size_t width) {
-  Matrix::resizeOrCreate(expandInput_, height, width, false, useGpu_);
-}
-
-void ExpandConvLayer::resetConvOutput(size_t batchSize, int inIdx) {
-  Matrix::resizeOrCreate(transOutValue_, batchSize * numFilters_, subN_[inIdx],
-                         false, useGpu_);
-}
 
 void ExpandConvLayer::expandOneFrame(MatrixPtr image, size_t startIdx,
                                      int inIdx) {
@@ -135,17 +127,7 @@ void ExpandConvLayer::addSharedBias() {
   transOutValue_->reshape(mapW, mapH);
   transOutValue_->transpose(out, false);  // false means no memory allocation
 
-  out->clear();
-  bias->clear();
-}
 
-void ExpandConvLayer::addUnsharedBias() {
-  MatrixPtr outValue = getOutputValue();
-  MatrixPtr bias =
-      Matrix::create(biases_->getW()->getData(), 1,
-                     biases_->getW()->getElementCnt(), false, useGpu_);
-  outValue->addBias(*bias, 1.0f);
-}
 
 void ExpandConvLayer::forward(PassType passType) {
   Layer::forward(passType);

paddle/gserver/layers/ExpandConvLayer.h

@@ -43,6 +43,7 @@ protected:
   /// The transpose of output, which is an auxiliary matrix.
   MatrixPtr transOutValue_;
 
+
 public:
   explicit ExpandConvLayer(const LayerConfig& config) : ConvBaseLayer(config) {}
 
@@ -52,16 +53,6 @@ public:
 
   size_t getOutputSize();
 
-  /**
-   * Create or resize expandInput_.
-   */
-  void resetExpandInput(size_t height, size_t width);
-
-  /**
-   * Create or resize transOutValue_.
-   */
-  void resetConvOutput(size_t batchSize, int inIdx);
-
   /**
    * Expand one input sample.
    */
@@ -72,15 +63,7 @@ public:
    */
   void expandFwdOnce(MatrixPtr image, int inIdx, int startIdx);
 
-  /**
-   * Add shared bias.
-   */
-  void addSharedBias();
 
-  /**
-   * Add unshared bias.
-   */
-  void addUnsharedBias();
   void forward(PassType passType);
   void bpropSharedBias(MatrixPtr biases, MatrixPtr v);
   void bpropBiases(MatrixPtr v);

paddle/gserver/layers/ExpandConvTransLayer.cpp

@@ -29,14 +29,14 @@ REGISTER_LAYER(exconvt, ExpandConvTransLayer);
 bool ExpandConvTransLayer::init(const LayerMap &layerMap,
                            const ParameterMap &parameterMap) {
   /* Initialize the basic convolutional parent class */
-  ConvTransBaseLayer::init(layerMap, parameterMap);
+  ConvBaseLayer::init(layerMap, parameterMap);
 
   /* Initialize the projection */
   for (auto &inputConfig : config_.inputs()) {
     const ConvConfig &conf = inputConfig.conv_conf();
     subM_.push_back(conf.channels() / conf.groups());
     subN_.push_back(conf.output_x() * conf.output_x());
-    subK_.push_back(channel_ * conf.filter_size() * conf.filter_size() /
+    subK_.push_back(numFilters_ * conf.filter_size() * conf.filter_size() /
                     conf.groups());
     /* Consistent caffe mode for multiple input */
     caffeMode_ = conf.caffe_mode();
@@ -65,8 +65,8 @@ size_t ExpandConvTransLayer::getSize() {
         imageSize(outputW_[i], filterSize_[i], padding_[i], stride_[i]));
     subN_.push_back(outputH_[i] * outputW_[i]);
     CHECK(layerSize == 0 ||
-            imgSizeH_[i] * imgSizeW_[i] * (size_t)channel_ == layerSize);
-    layerSize = imgSizeH_[i] * imgSizeW_[i] * channel_;
+            imgSizeH_[i] * imgSizeW_[i] * (size_t)numFilters_ == layerSize);
+    layerSize = imgSizeH_[i] * imgSizeW_[i] * numFilters_;
   }
   getOutput().setFrameHeight(imgSizeH_[0]);
   getOutput().setFrameWidth(imgSizeW_[0]);
@@ -83,38 +83,6 @@ void ExpandConvTransLayer::resetExpandInput(size_t height, size_t width) {
 }*/
 
 
-void ExpandConvTransLayer::addSharedBias() {
-  size_t mapW = getSize() / channel_;
-  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() / channel_,
-                          channel_);
-
-  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();
-}
-
-void ExpandConvTransLayer::addUnsharedBias() {
-  MatrixPtr outValue = getOutputValue();
-  MatrixPtr bias =
-      Matrix::create(biases_->getW()->getData(), 1,
-                     biases_->getW()->getElementCnt(), false, useGpu_);
-  outValue->addBias(*bias, 1.0f);
-}
-
 
 void ExpandConvTransLayer::expandOneFrame(MatrixPtr image, size_t startIdx,
                                      int inIdx) {

paddle/gserver/layers/ExpandConvTransLayer.h

@@ -15,7 +15,7 @@ limitations under the License. */
 
 #pragma once
 
-#include "ConvTransBaseLayer.h"
+#include "ConvBaseLayer.h"
 #include "paddle/math/Matrix.h"
 #include <vector>
 
@@ -28,7 +28,7 @@ namespace paddle {
  *
  * The config file api is img_convTrans_layer.
  */
-class ExpandConvTransLayer : public ConvTransBaseLayer {
+class ExpandConvTransLayer : public ConvBaseLayer {
 protected:
   /// For expand convolution.
   /// subM_ = numFilters_ / groups_.
@@ -45,15 +45,11 @@ protected:
   IntV outputH_;
   /// The spatial dimensions of width of output feature map.
   IntV outputW_;
-  /// Expand one sample at a time. shape:
-  /// (numChannels * filterPixels_, outputSizeH * outputSizeW)
-  MatrixPtr expandInput_;
-  /// The transpose of output, which is an auxiliary matrix.
-  MatrixPtr transOutValue_;
+
 
 public:
   explicit ExpandConvTransLayer(const LayerConfig& config) :
-      ConvTransBaseLayer(config) {}
+      ConvBaseLayer(config) {}
 
   ~ExpandConvTransLayer() {}
 
@@ -86,15 +82,7 @@ public:
    */
   void shrinkFwd(MatrixPtr output, int inpIdx);
 
-  /**
-   * Add shared bias.
-   */
-  void addSharedBias();
 
-  /**
-   * Add unshared bias.
-   */
-  void addUnsharedBias();
   void forward(PassType passType);
   void bpropSharedBias(MatrixPtr biases, MatrixPtr v);
   void bpropBiases(MatrixPtr v);