modify the format and delete useless comment

paddle/function/DepthwiseConvOp.cpp

@@ -15,7 +15,6 @@ limitations under the License. */
 #include "DepthwiseConvOp.h"
 #include "ConvOp.h"
 #include "GemmFunctor.h"
-//#include "paddle/math/MemoryHandle.h"
 
 namespace paddle {
 
@@ -28,6 +27,7 @@ public:
                   int outputChannels,
                   int outputHeight,
                   int outputWidth,
+                  int inputChannels,
                   int inputHeight,
                   int inputWidth,
                   int filterHeight,
@@ -114,7 +114,7 @@ public:
     const TensorShape& output = outputs[0].shape();
 
     size_t batchSize = input[0];
-    // size_t inputChannels = input[1];
+    size_t inputChannels = input[1];
     size_t inputHeight = input[2];
     size_t inputWidth = input[3];
     size_t filterHeight = getFilterHeight(filter);
@@ -134,6 +134,7 @@ public:
                   outputChannels,
                   outputHeight,
                   outputWidth,
+                  inputChannels,
                   inputHeight,
                   inputWidth,
                   filterHeight,
@@ -168,8 +169,6 @@ public:
     CHECK_EQ(numInputs_, inputs.size());
     CHECK_EQ(numOutputs_, outputs.size());
     check(inputs, outputs);
-    // Since the implementation of Col2ImFunctor is ADD_TO,
-    // this function only supports ADD_TO mode.
     CHECK_EQ(outputs[0].getArgType(), ADD_TO);
     const TensorShape& output = inputs[0].shape();
     const TensorShape& filter = inputs[1].shape();
@@ -228,12 +227,11 @@ public:
   }
 
   void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
-    // CHECK_EQ(numInputs_, inputs.size());
-    // CHECK_EQ(numOutputs_, outputs.size());
+    CHECK_EQ(numInputs_, inputs.size());
+    CHECK_EQ(numOutputs_, outputs.size());
     check(inputs, outputs);
     const TensorShape& output = inputs[0].shape();
     const TensorShape& input = inputs[1].shape();
-    // const TensorShape& multiplier = inputs[2].shape();
     const TensorShape& filter = outputs[0].shape();
 
     size_t batchSize = input[0];

paddle/function/DepthwiseConvOp.h

@@ -29,6 +29,7 @@ namespace paddle {
  * \param[in]   outputChannels    channels of outputData.
  * \param[in]   outputHeight      height of outputData.
  * \param[in]   outputWidth       width of outputData.
+ * \param[in]   inputChannels     channels of inputData.
  * \param[in]   inputHeight       height of inputData.
  * \param[in]   inputWidth        width of inputData..
  * \param[in]   filterHeight      height of filter.
@@ -49,8 +50,9 @@ public:
                   int outputChannels,
                   int outputHeight,
                   int outputWidth,
+                  int inputChannels,
                   int inputHeight,
-                  int intputWidth,
+                  int inputWidth,
                   int filterHeight,
                   int filterWidth,
                   int strideH,

paddle/function/DepthwiseConvOpGpu.cu

@@ -24,7 +24,7 @@ __global__
 void ConvolutionDepthwiseForward(const int nthreads,
     const T* const inputData, const T* const filterData,
     const int batchSize, const int outputChannels, const int outputHeight,
-    const int outputWidth, const int inputHeight, const int inputWidth,
+    const int outputWidth,const int inputChannels, const int inputHeight, const int inputWidth,
     const int filterHeight, const int filterWidth, const int strideH,
     const int strideW, const int paddingH, const int paddingW,
     T* const outputData) {
@@ -49,7 +49,7 @@ void ConvolutionDepthwiseForward(const int nthreads,
             for (int kw = 0; kw < filterWidth; ++kw) {
                 const int h_in = -paddingH + h * strideH + kh;
                 const int w_in = -paddingW + w * strideW + kw;
-			const int offset = ((n * outputChannels + c) * inputHeight + h_in)
+                const int offset = ((n * inputChannels + c) * inputHeight + h_in)
 					* inputWidth + w_in;
                 value += (*weight) * inputData[offset];
                 ++weight;
@@ -80,15 +80,15 @@ __global__
 void ConvolutionDepthwiseInputBackward(const int nthreads,
     const T* const top_diff, const T* const weight_data,
     const int num, const int outputChannels, const int outputHeight,
-    const int outputWidth, const int inputHeight, const int inputWidth,
+    const int outputWidth,const int inputChannels, const int inputHeight, const int inputWidth,
     const int filterHeight, const int filterWidth, const int strideH,
     const int strideW, const int paddingH, const int paddingW,
      T* const bottom_diff) {
   int index =
     (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x;
   if(index < nthreads) {
-    const int n = index / outputChannels / inputHeight / inputWidth;
-    const int c = (index / inputHeight / inputWidth) % outputChannels;
+    const int n = index / inputChannels / inputHeight / inputWidth;
+    const int c = (index / inputHeight / inputWidth) % inputChannels;
     const int h = (index / inputWidth) % inputHeight;
     const int w = index % inputWidth;
     const T* weight = weight_data + c * filterHeight * filterWidth;
@@ -121,7 +121,7 @@ __global__
 void ConvolutionDepthwiseFilterBackward(const int num_i, const int nthreads,
     const T* const top_diff, const T* const inputData,
     const int num, const int outputChannels, const int outputHeight,
-    const int outputWidth, const int inputHeight, const int inputWidth,
+    const int outputWidth, const int inputChannels, const int inputHeight, const int inputWidth,
     const int filterHeight, const int filterWidth, const int strideH,
     const int strideW, const int paddingH, const int paddingW,
     T* const buffer_data) {
@@ -141,7 +141,7 @@ void ConvolutionDepthwiseFilterBackward(const int num_i, const int nthreads,
       const int n = num_i;
       const int top_offset = ((n * outputChannels + c) * outputHeight + h)
             * outputWidth + w;
-      const int bottom_offset = ((n * outputChannels + c) * inputHeight + h_in)
+      const int bottom_offset = ((n * inputChannels + c) * inputHeight + h_in)
             * inputWidth + w_in;
       buffer_data[index] = top_diff[top_offset] * inputData[bottom_offset];
     } else {
@@ -159,6 +159,7 @@ public:
             int outputChannels,
             int outputHeight,
             int outputWidth,
+			int inputChannels,
             int inputHeight,
             int inputWidth,
             int filterHeight,
@@ -186,6 +187,7 @@ public:
             outputChannels,
             outputHeight,
             outputWidth,
+			inputChannels,
             inputHeight,
             inputWidth,
             filterHeight,
@@ -237,6 +239,7 @@ public:
             outputChannels,
             outputHeight,
             outputWidth,
+			inputChannels,
             inputHeight,
             inputWidth,
             filterHeight,
@@ -290,6 +293,7 @@ public:
                     outputChannels,
                     outputHeight,
                     outputWidth,
+					inputChannels,
                     inputHeight,
                     inputWidth,
                     filterHeight,