abstract outputSize function in CNN-related layers (#314)

PaddlePaddle · Nov 7, 2016 · e802471 · e802471
1 parent f9849ac
commit e802471
Show file tree

Hide file tree

Showing 11 changed files with 139 additions and 184 deletions.
diff --git a/paddle/gserver/layers/ConvBaseLayer.cpp b/paddle/gserver/layers/ConvBaseLayer.cpp
@@ -12,7 +12,6 @@ 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/Logging.h"
 #include "ConvBaseLayer.h"
 namespace paddle {
@@ -78,10 +77,10 @@ size_t ConvBaseLayer::calOutputSize() {
       imgSizeH_[i] = config_.inputs(i).conv_conf().img_size();
     if (imgSizeW_[i] == 0)
       imgSizeW_[i] = config_.inputs(i).conv_conf().img_size();
-    outputH_.push_back(
-        outputSize(imgSizeH_[i], filterSizeY_[i], paddingY_[i], strideY_[i]));
-    outputW_.push_back(
-        outputSize(imgSizeW_[i], filterSize_[i], padding_[i], stride_[i]));
+    outputH_.push_back(outputSize(imgSizeH_[i], filterSizeY_[i], paddingY_[i],
+                                  strideY_[i], caffeMode_));
+    outputW_.push_back(outputSize(imgSizeW_[i], filterSize_[i], padding_[i],
+                                  stride_[i], caffeMode_));
     CHECK_EQ(outputH_[i], outputH_[0]);
     CHECK_EQ(outputW_[i], outputW_[0]);
   }

diff --git a/paddle/gserver/layers/ConvBaseLayer.h b/paddle/gserver/layers/ConvBaseLayer.h
@@ -16,6 +16,7 @@ limitations under the License. */
 #pragma once
 
 #include "Layer.h"
+#include "paddle/math/MathUtils.h"
 namespace paddle {
 
 /**
@@ -87,31 +88,6 @@ class ConvBaseLayer : public Layer {
   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;
-  }
 };
 
 }  // namespace paddle
diff --git a/paddle/gserver/layers/ConvOperator.cpp b/paddle/gserver/layers/ConvOperator.cpp
@@ -12,8 +12,8 @@ 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/math/Matrix.h"
+#include "paddle/math/MathUtils.h"
 #include "Operator.h"
 
 namespace paddle {
@@ -35,8 +35,8 @@ class ConvOperator : public Operator {
    */
   virtual ~ConvOperator() {
     if (workSpaceInBytes_ != 0) {
-        hl_free_mem_device(workSpace_);
-        workSpaceInBytes_ = 0;
+      hl_free_mem_device(workSpace_);
+      workSpaceInBytes_ = 0;
     }
 
     hl_destroy_tensor_descriptor(inputDesc_);
@@ -83,33 +83,6 @@ class ConvOperator : public Operator {
              filterSize_ * filterSizeY_ * channels_ * numFilters_);
   }
 
-  /**
-   * Calculate output size.
-   */
-  int outputSize(int imageSize, int filterSize, int padding, int stride) {
-    int outputSize;
-    if (!caffeMode_) {
-      /* input(+padding): 0123456789
-       * imageSize(+padding) = 10;
-       * filterSize = 3;
-       * stride = 2;
-       * output: (012), (234), (456), (678), (9)
-       * outputSize = 5;
-       */
-      outputSize =
-          (imageSize - filterSize + 2 * padding + stride - 1) / stride + 1;
-    } else {
-      /* input(+padding): 0123456789
-       * imageSize(+padding) = 10;
-       * filterSize = 3;
-       * stride = 2;
-       * output: (012), (234), (456), (678)
-       * outputSize = 4;
-       */
-      outputSize = (imageSize - filterSize + 2 * padding) / stride + 1;
-    }
-    return outputSize;
-  }
   /// Most of member variables are same with CudnnConvLayer.
   /// There is no explanation here.
   int imageH_, imageW_, outputH_, outputW_;
@@ -129,7 +102,7 @@ class ConvOperator : public Operator {
   int fwdAlgo_, bwdFilterAlgo_, bwdDataAlgo_;
   size_t fwdLimitBytes_, bwdDataLimitBytes_, bwdFilterLimitBytes_;
   size_t workSpaceInBytes_;
-  void* workSpace_;
+  void *workSpace_;
   bool isSelectAlgo_;
 };
 
@@ -160,33 +133,32 @@ ConvOperator::ConvOperator(const OperatorConfig &config, bool useGpu)
 void ConvOperator::allocConvWorkSpace(size_t maxWorkSpace) {
   if (maxWorkSpace > workSpaceInBytes_) {
     if (workSpaceInBytes_ != 0) {
-        hl_free_mem_device(workSpace_);
+      hl_free_mem_device(workSpace_);
     }
     // total amount of storage needed
     workSpace_ = hl_malloc_device(maxWorkSpace);
     workSpaceInBytes_ = maxWorkSpace;
   }
 }
 
-
 void ConvOperator::reshape(int batchSize) {
   imageH_ = ins_[0]->getFrameHeight();
   imageW_ = ins_[0]->getFrameWidth();
   if (imageH_ == 0) imageH_ = imgSize_;
   if (imageW_ == 0) imageW_ = imgSize_;
-  outputH_ = outputSize(imageH_, filterSizeY_, paddingY_, strideY_);
-  outputW_ = outputSize(imageW_, filterSize_, padding_, stride_);
+  outputH_ = outputSize(imageH_, filterSizeY_, paddingY_, strideY_, caffeMode_);
+  outputW_ = outputSize(imageW_, filterSize_, padding_, stride_, caffeMode_);
 
   out_->setFrameHeight(outputH_);
   out_->setFrameWidth(outputW_);
 
   reshapeImageDescriptors();
 
   if (!isSelectAlgo_) {
-    hl_conv_workspace(inputDesc_, outputDesc_, filterDesc_,
-               convDesc_, &fwdAlgo_, &fwdLimitBytes_,
-               &bwdDataAlgo_, &bwdDataLimitBytes_,
-               &bwdFilterAlgo_, &bwdFilterLimitBytes_);
+    hl_conv_workspace(inputDesc_, outputDesc_, filterDesc_, convDesc_,
+                      &fwdAlgo_, &fwdLimitBytes_, &bwdDataAlgo_,
+                      &bwdDataLimitBytes_, &bwdFilterAlgo_,
+                      &bwdFilterLimitBytes_);
 
     size_t maxWorkSpace = 0;
     maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_);
@@ -202,7 +174,8 @@ void ConvOperator::computeConvSizes() {
   hl_create_filter_descriptor(&filterDesc_, channels_, numFilters_,
                               filterSizeY_, filterSize_);
   hl_create_tensor_descriptor(&inputDesc_);
-  int outputX = outputSize(imgSize_, filterSize_, padding_, stride_);
+  int outputX =
+      outputSize(imgSize_, filterSize_, padding_, stride_, caffeMode_);
   CHECK_EQ(outputX, outputX_);
   hl_create_tensor_descriptor(&outputDesc_);
   hl_create_convolution_descriptor(&convDesc_, inputDesc_, filterDesc_,
@@ -211,13 +184,13 @@ void ConvOperator::computeConvSizes() {
 
 void ConvOperator::reshapeImageDescriptors() {
   hl_tensor_reshape(inputDesc_, 1, channels_, imageH_, imageW_,
-                    channels_ * imageH_ * imageW_, imageH_ * imageW_,
-                    imageW_, 1);
+                    channels_ * imageH_ * imageW_, imageH_ * imageW_, imageW_,
+                    1);
   hl_tensor_reshape(outputDesc_, 1, numFilters_, outputH_, outputW_,
                     numFilters_ * outputH_ * outputW_, outputH_ * outputW_,
                     outputW_, 1);
-  hl_reset_convolution_descriptor(convDesc_, inputDesc_, filterDesc_,
-                                  paddingY_, padding_, strideY_, stride_);
+  hl_reset_convolution_descriptor(convDesc_, inputDesc_, filterDesc_, paddingY_,
+                                  padding_, strideY_, stride_);
   inputOffset_ = channels_ * imageH_ * imageW_;
   outputOffset_ = numFilters_ * outputH_ * outputW_;
   weightOffset_ = numFilters_ * channels_ * filterSize_ * filterSize_;
@@ -273,18 +246,17 @@ void ConvOperator::backward() {
         real *weightGrad = ins_[1]->grad->getData() + weightOffset_ * batchId;
         hl_convolution_backward_filter(inputDesc_, inputData, outputDesc_,
                                        outGrad, filterDesc_, weightGrad,
-                                       convDesc_, workSpace_,
-                                       workSpaceInBytes_, bwdFilterAlgo_);
+                                       convDesc_, workSpace_, workSpaceInBytes_,
+                                       bwdFilterAlgo_);
       }
 
       MatrixPtr preGrad = ins_[0]->grad;
       if (NULL != preGrad) {
         real *inputGrad = preGrad->getData() + inputOffset_ * batchId;
         real *wgtData = ins_[1]->value->getData() + weightOffset_ * batchId;
-        hl_convolution_backward_data(inputDesc_, inputGrad, outputDesc_,
-                                     outGrad, filterDesc_, wgtData,
-                                     convDesc_, workSpace_,
-                                     workSpaceInBytes_, bwdDataAlgo_);
+        hl_convolution_backward_data(
+            inputDesc_, inputGrad, outputDesc_, outGrad, filterDesc_, wgtData,
+            convDesc_, workSpace_, workSpaceInBytes_, bwdDataAlgo_);
       }
     }
   }

diff --git a/paddle/gserver/layers/ConvProjection.h b/paddle/gserver/layers/ConvProjection.h
@@ -12,10 +12,10 @@ 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 "Projection.h"
+#include "paddle/math/MathUtils.h"
 
 namespace paddle {
 
@@ -42,17 +42,15 @@ class ConvProjection : public Projection {
   void reshapeTensorDesc(int batchSize);
   void reshape(int batchSize);
 
-  int outputSize(int imageSize, int filterSize, int padding, int stride) {
-    return (imageSize - filterSize + 2 * padding) / stride + 1;
-  }
-
   size_t calOutputSize() {
     imageH_ = in_->getFrameHeight();
     imageW_ = in_->getFrameWidth();
     if (imageH_ == 0) imageH_ = configImgH_;
     if (imageW_ == 0) imageW_ = configImgW_;
-    outputH_ = outputSize(imageH_, filterH_, paddingH_, strideH_);
-    outputW_ = outputSize(imageW_, filterW_, paddingW_, strideW_);
+    outputH_ = outputSize(imageH_, filterH_, paddingH_, strideH_,
+                          /* caffeMode */ true);
+    outputW_ = outputSize(imageW_, filterW_, paddingW_, strideW_,
+                          /* caffeMode */ true);
 
     const_cast<Argument*>(out_)->setFrameHeight(outputH_);
     const_cast<Argument*>(out_)->setFrameWidth(outputW_);

diff --git a/paddle/gserver/layers/CudnnPoolLayer.cpp b/paddle/gserver/layers/CudnnPoolLayer.cpp
@@ -12,7 +12,6 @@ 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/Logging.h"
 #include "paddle/utils/Stat.h"
 #include "paddle/math/Matrix.h"
@@ -62,9 +61,9 @@ bool CudnnPoolLayer::init(const LayerMap &layerMap,
   strideHeight = strideY_;
   strideWidth = stride_;
 
-  hl_create_pooling_descriptor(&poolingDesc_, mode_, windowHeight,
-                               windowWidth, heightPadding, widthPadding,
-                               strideHeight, strideWidth);
+  hl_create_pooling_descriptor(&poolingDesc_, mode_, windowHeight, windowWidth,
+                               heightPadding, widthPadding, strideHeight,
+                               strideWidth);
 
   return true;
 }
@@ -80,8 +79,10 @@ void CudnnPoolLayer::reshape(int batchSize) {
   }
   CHECK_EQ(inputLayers_[0]->getOutput().value->getWidth(),
            channels_ * imageH_ * imageW_);
-  outputH_ = outputSize(imageH_, sizeY_, confPaddingY_, strideY_);
-  outputW_ = outputSize(imageW_, sizeX_, confPadding_, stride_);
+  outputH_ = outputSize(imageH_, sizeY_, confPaddingY_, strideY_,
+                        /* caffeMode */ false);
+  outputW_ =
+      outputSize(imageW_, sizeX_, confPadding_, stride_, /* caffeMode */ false);
   getOutput().setFrameHeight(outputH_);
   getOutput().setFrameWidth(outputW_);
 
@@ -99,8 +100,7 @@ void CudnnPoolLayer::forward(PassType passType) {
 
   real *inputData = getInputValue(0)->getData();
   real *outData = getOutputValue()->getData();
-  hl_pooling_forward(inputDesc_, inputData, outputDesc_, outData,
-                     poolingDesc_);
+  hl_pooling_forward(inputDesc_, inputData, outputDesc_, outData, poolingDesc_);
 }
 
 void CudnnPoolLayer::backward(const UpdateCallback &callback) {
@@ -113,8 +113,8 @@ void CudnnPoolLayer::backward(const UpdateCallback &callback) {
   real *inputGrad = getInputGrad(0)->getData();
   real *outData = getOutputValue()->getData();
   real *outGrad = getOutputGrad()->getData();
-  hl_pooling_backward(inputDesc_, inputData, inputGrad, outputDesc_,
-                      outData, outGrad, poolingDesc_);
+  hl_pooling_backward(inputDesc_, inputData, inputGrad, outputDesc_, outData,
+                      outGrad, poolingDesc_);
 }
 
 CudnnPoolLayer::~CudnnPoolLayer() {

diff --git a/paddle/gserver/layers/PoolLayer.h b/paddle/gserver/layers/PoolLayer.h
@@ -17,6 +17,7 @@ limitations under the License. */
 
 #include "Layer.h"
 #include "paddle/math/Matrix.h"
+#include "paddle/math/MathUtils.h"
 #include <vector>
 
 namespace paddle {
@@ -47,16 +48,6 @@ class PoolLayer : public Layer {
   static Layer* create(const LayerConfig& config);
 
   virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-
-  /**
-   * Calculate output size according window size and padding size.
-   */
-  int outputSize(int imageSize, int windowSize, int padding, int stride) {
-    int outputSize;
-    outputSize =
-        (imageSize - windowSize + 2 * padding + stride - 1) / stride + 1;
-    return outputSize;
-  }
 };
 
 }  // namespace paddle