diff --git a/paddle/cuda/include/hl_cuda_cudnn.h b/paddle/cuda/include/hl_cuda_cudnn.h index 3f68c62de6d9b3aaadc9180d86159089dc728ea9..b44b071bd1b3b6e9e5539d5dc0c2b155c524fd57 100644 --- a/paddle/cuda/include/hl_cuda_cudnn.h +++ b/paddle/cuda/include/hl_cuda_cudnn.h @@ -22,10 +22,10 @@ limitations under the License. */ */ typedef enum { HL_POOLING_MAX = 0, - // average includes padded values - HL_POOLING_AVERAGE = 1, // average does not include padded values - HL_POOLING_AVERAGE_EXCLUDE_PADDING = 2, + HL_POOLING_AVERAGE = 1, + // average includes padded values + HL_POOLING_AVERAGE_INCLUDE_PADDING = 2, HL_POOLING_END } hl_pooling_mode_t; diff --git a/paddle/cuda/src/hl_cuda_cnn.cu b/paddle/cuda/src/hl_cuda_cnn.cu index 9ba3d142617537c0160f6dccb86ddca43ada15a5..58674febdc4a094c95ff03701e4586c32729847d 100644 --- a/paddle/cuda/src/hl_cuda_cnn.cu +++ b/paddle/cuda/src/hl_cuda_cnn.cu @@ -211,13 +211,11 @@ __global__ void KeAvgPoolForward(const int nthreads, int hstart = ph * strideH - padH; int wstart = pw * strideW - padW; - int hend = min(hstart + sizeY, height + padH); - int wend = min(wstart + sizeX, width + padW); - int pool_size = (hend - hstart) * (wend - wstart); + int hend = min(hstart + sizeY, height); + int wend = min(wstart + sizeX, width); hstart = max(hstart, 0); wstart = max(wstart, 0); - hend = min(hend, height); - wend = min(wend, width); + int pool_size = (hend - hstart) * (wend - wstart); real aveval = 0; inputData += (frameNum * channels + c) * height * width; @@ -299,12 +297,14 @@ __global__ void KeAvgPoolBackward(const int nthreads, outGrad += (frameNum * outStride + offsetC * pooledH * pooledW); for (int ph = phstart; ph < phend; ++ph) { + int hstart = ph * strideH - padH; + int hend = min(hstart + sizeY, height); + hstart = max(hstart, 0); for (int pw = pwstart; pw < pwend; ++pw) { // figure out the pooling size - int hstart = ph * strideH - padH; int wstart = pw * strideW - padW; - int hend = min(hstart + sizeY, height + padH); - int wend = min(wstart + sizeX, width + padW); + int wend = min(wstart + sizeX, width); + wstart = max(wstart, 0); int poolsize = (hend - hstart) * (wend - wstart); gradient += outGrad[ph * pooledW + pw] / poolsize; } @@ -600,16 +600,13 @@ __global__ void KeAvgPool3DForward(const int nthreads, int dstart = pd * strideD - padD; int hstart = ph * strideH - padH; int wstart = pw * strideW - padW; - int dend = min(dstart + sizeZ, depth + padD); - int hend = min(hstart + sizeY, height + padH); - int wend = min(wstart + sizeX, width + padW); - int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); + int dend = min(dstart + sizeZ, depth); + int hend = min(hstart + sizeY, height); + int wend = min(wstart + sizeX, width); dstart = max(dstart, 0); hstart = max(hstart, 0); wstart = max(wstart, 0); - dend = min(dend, depth); - hend = min(hend, height); - wend = min(wend, width); + int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); real aveval = 0; inputData += (frameNum * channels + c) * depth * height * width; @@ -712,15 +709,18 @@ __global__ void KeAvgPool3DBackward(const int nthreads, outGrad += (frameNum * channels + offsetC) * pooledD * pooledH * pooledW; for (int pd = pdstart; pd < pdend; ++pd) { + int dstart = pd * strideD - padD; + int dend = min(dstart + sizeZ, depth); + dstart = max(dstart, 0); for (int ph = phstart; ph < phend; ++ph) { + int hstart = ph * strideH - padH; + int hend = min(hstart + sizeY, height); + hstart = max(hstart, 0); for (int pw = pwstart; pw < pwend; ++pw) { // figure out the pooling size - int dstart = pd * strideD - padD; - int hstart = ph * strideH - padH; int wstart = pw * strideW - padW; - int dend = min(dstart + sizeZ, depth + padD); - int hend = min(hstart + sizeY, height + padH); - int wend = min(wstart + sizeX, width + padW); + int wend = min(wstart + sizeX, width); + wstart = max(wstart, 0); int poolsize = (dend - dstart) * (hend - hstart) * (wend - wstart); gradient += outGrad[(pd * pooledH + ph) * pooledW + pw] / poolsize; } diff --git a/paddle/cuda/src/hl_cuda_cudnn.cc b/paddle/cuda/src/hl_cuda_cudnn.cc index f38ef692558b908ed65d2c84821bbb7c3b439742..b8caf48f9c06094e85765f7aa5a3f4195d0ca931 100644 --- a/paddle/cuda/src/hl_cuda_cudnn.cc +++ b/paddle/cuda/src/hl_cuda_cudnn.cc @@ -432,11 +432,11 @@ void hl_create_pooling_descriptor(hl_pooling_descriptor* pooling_desc, cudnn_mode = CUDNN_POOLING_MAX; break; case HL_POOLING_AVERAGE: - cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING; - break; - case HL_POOLING_AVERAGE_EXCLUDE_PADDING: cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING; break; + case HL_POOLING_AVERAGE_INCLUDE_PADDING: + cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING; + break; default: LOG(FATAL) << "parameter mode error"; } diff --git a/paddle/gserver/layers/CudnnPoolLayer.cpp b/paddle/gserver/layers/CudnnPoolLayer.cpp index 4adb2d4709e585a6fec052435c33714d6e3a3f0e..810a1af2d09c63c3787a1ac225c2c7de4238d609 100644 --- a/paddle/gserver/layers/CudnnPoolLayer.cpp +++ b/paddle/gserver/layers/CudnnPoolLayer.cpp @@ -29,9 +29,9 @@ bool CudnnPoolLayer::typeCheck(const std::string &poolType, if (mode) { *mode = HL_POOLING_AVERAGE; } - } else if (poolType == "cudnn-avg-excl-pad-pool") { + } else if (poolType == "cudnn-avg-incl-pad-pool") { if (mode) { - *mode = HL_POOLING_AVERAGE_EXCLUDE_PADDING; + *mode = HL_POOLING_AVERAGE_INCLUDE_PADDING; } } else { return false; diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp index 4a2132c8d1bfa329ced575f9b78052bdbfe3e4d5..0023b4d0f5da500f380ecb836b7c54e050b13d67 100644 --- a/paddle/math/Matrix.cpp +++ b/paddle/math/Matrix.cpp @@ -1033,17 +1033,15 @@ void GpuMatrix::maxPoolForward(Matrix& inputMat, real* inputData = inputMat.getData(); size_t frameNum = inputMat.getHeight(); - size_t width = imgSizeW; - size_t height = imgSizeH; - CHECK(height * width * channels == inputMat.getWidth()); + CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth()); CHECK(height_ == inputMat.getHeight()); CHECK(width_ == outputH * outputW * channels); hl_maxpool_forward(frameNum, inputData, channels, - height, - width, + imgSizeH, + imgSizeW, outputH, outputW, sizeX, @@ -1080,11 +1078,8 @@ void GpuMatrix::maxPoolBackward(Matrix& inputMat, real* outDiff = outGrad.getData(); size_t frameNum = inputMat.getHeight(); size_t channels = outV.getWidth() / outputH / outputW; - size_t width = imgSizeW; - size_t height = imgSizeH; - CHECK(height * width * channels == inputMat.getWidth()); + CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth()); CHECK(height_ == inputMat.getHeight()); - CHECK(width_ == width * height * channels); CHECK(outGrad.getHeight() == outV.getHeight() && outGrad.getWidth() == outV.getWidth()); @@ -1093,8 +1088,8 @@ void GpuMatrix::maxPoolBackward(Matrix& inputMat, outData, outDiff, channels, - height, - width, + imgSizeH, + imgSizeW, outputH, outputW, sizeX, @@ -1125,17 +1120,15 @@ void GpuMatrix::avgPoolForward(Matrix& inputMat, real* inputData = inputMat.getData(); size_t frameNum = inputMat.getHeight(); - size_t height = imgSizeH; - size_t width = imgSizeW; - CHECK(height * width * channels == inputMat.getWidth()); + CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth()); CHECK(height_ == inputMat.getHeight()); CHECK(width_ == outputH * outputW * channels); hl_avgpool_forward(frameNum, inputData, channels, - height, - width, + imgSizeH, + imgSizeW, outputH, outputW, sizeX, @@ -1166,17 +1159,15 @@ void GpuMatrix::avgPoolBackward(Matrix& outGrad, real* outDiff = outGrad.getData(); size_t frameNum = outGrad.getHeight(); size_t channels = outGrad.getWidth() / outputH / outputW; - size_t height = imgSizeH; - size_t width = imgSizeW; - CHECK(height * width * channels == width_); + CHECK(imgSizeH * imgSizeW * channels == width_); CHECK(height_ == outGrad.getHeight()); CHECK(outGrad.getWidth() == outputH * outputW * channels); hl_avgpool_backward(frameNum, outDiff, channels, - height, - width, + imgSizeH, + imgSizeW, outputH, outputW, sizeX, @@ -1214,19 +1205,16 @@ void GpuMatrix::maxPool3DForward(Matrix& inputMat, real* inputData = inputMat.getData(); real* maxPoolIdxData = maxPoolIdx.getData(); size_t num = inputMat.getHeight(); - size_t width = imgSizeW; - size_t height = imgSizeH; - size_t depth = imgSizeD; - CHECK(depth * height * width * channels == inputMat.getWidth()); + CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth()); CHECK(height_ == inputMat.getHeight()); CHECK(width_ == outputD * outputH * outputW * channels); hl_maxpool3D_forward(num, inputData, channels, - depth, - height, - width, + imgSizeD, + imgSizeH, + imgSizeW, outputD, outputH, outputW, @@ -1269,20 +1257,16 @@ void GpuMatrix::maxPool3DBackward(Matrix& outGrad, real* maxPoolIdxData = maxPoolIdx.getData(); size_t frameNum = getHeight(); size_t channels = outGrad.getWidth() / outputD / outputH / outputW; - size_t width = imgSizeW; - size_t height = imgSizeH; - size_t depth = imgSizeD; - CHECK(depth * height * width * channels == getWidth()); - CHECK(width_ == depth * width * height * channels); + CHECK(imgSizeD * imgSizeH * imgSizeW * channels == getWidth()); CHECK(outGrad.getHeight() == maxPoolIdx.getHeight() && outGrad.getWidth() == maxPoolIdx.getWidth()); hl_maxpool3D_backward(frameNum, outDiff, channels, - depth, - height, - width, + imgSizeD, + imgSizeH, + imgSizeW, outputD, outputH, outputW, @@ -1323,19 +1307,16 @@ void GpuMatrix::avgPool3DForward(Matrix& inputMat, real* inputData = inputMat.getData(); size_t frameNum = inputMat.getHeight(); - size_t height = imgSizeH; - size_t width = imgSizeW; - size_t depth = imgSizeD; - CHECK(depth * height * width * channels == inputMat.getWidth()); + CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth()); CHECK(height_ == inputMat.getHeight()); CHECK(width_ == outputD * outputH * outputW * channels); hl_avgpool3D_forward(frameNum, inputData, channels, - depth, - height, - width, + imgSizeD, + imgSizeH, + imgSizeW, outputD, outputH, outputW, @@ -1375,19 +1356,16 @@ void GpuMatrix::avgPool3DBackward(Matrix& outGrad, real* outDiff = outGrad.getData(); size_t frameNum = outGrad.getHeight(); size_t channels = outGrad.getWidth() / outputD / outputH / outputW; - size_t height = imgSizeH; - size_t width = imgSizeW; - size_t depth = imgSizeD; - CHECK(depth * height * width * channels == width_); + CHECK(imgSizeD * imgSizeH * imgSizeW * channels == width_); CHECK(height_ == outGrad.getHeight()); CHECK(outGrad.getWidth() == outputD * outputH * outputW * channels); hl_avgpool3D_backward(frameNum, outDiff, channels, - depth, - height, - width, + imgSizeD, + imgSizeH, + imgSizeW, outputD, outputH, outputW, @@ -1999,11 +1977,11 @@ void CpuMatrix::maxPoolForward(Matrix& inputMat, real* inputData = inputMat.getData(); real* outData = data_; size_t num = inputMat.getHeight(); - size_t inWidth = imgSizeW; - size_t inHeight = imgSizeH; - CHECK(inHeight * inWidth == inputMat.getWidth() / channels); + size_t inLength = imgSizeH * imgSizeW; + size_t outLength = outputH * outputW; + CHECK(inLength == inputMat.getWidth() / channels); CHECK_EQ(num, this->getHeight()); - CHECK_EQ(channels * outputH * outputW, this->getWidth()); + CHECK_EQ(channels * outLength, this->getWidth()); size_t outStride = getStride(); /* initialize the data_ */ @@ -2020,24 +1998,24 @@ void CpuMatrix::maxPoolForward(Matrix& inputMat, } for (size_t c = 0; c < channels; ++c) { // channel by channel for (size_t ph = 0; ph < outputH; ++ph) { + int hstart = ph * strideH - paddingH; + int hend = std::min(hstart + sizeY, imgSizeH); + hstart = std::max(hstart, 0); for (size_t pw = 0; pw < outputW; ++pw) { - int hstart = ph * strideH - paddingH; int wstart = pw * strideW - paddingW; - int hend = std::min(hstart + sizeY, inHeight); - int wend = std::min(wstart + sizeX, inWidth); - hstart = std::max(hstart, 0); + int wend = std::min(wstart + sizeX, imgSizeW); wstart = std::max(wstart, 0); for (int h = hstart; h < hend; ++h) { for (int w = wstart; w < wend; ++w) { - outData[ph * outputW + pw] = std::max(outData[ph * outputW + pw], - inputData[h * inWidth + w]); + outData[ph * outputW + pw] = std::max( + outData[ph * outputW + pw], inputData[h * imgSizeW + w]); } } } } // compute offset - inputData += inHeight * inWidth; - outData += outputH * outputW; + inputData += inLength; + outData += outLength; } } } @@ -2058,8 +2036,10 @@ void CpuMatrix::maxPoolBackward(Matrix& image, size_t paddingH, size_t paddingW) { size_t num = image.getHeight(); - size_t channels = size_t(width_ / imgSizeH / imgSizeW); - CHECK(image.getWidth() == imgSizeH * imgSizeW * channels); + size_t inLength = imgSizeH * imgSizeW; + size_t outLength = outputH * outputW; + size_t channels = size_t(width_ / inLength); + CHECK(image.getWidth() == inLength * channels); CHECK(image.getHeight() == height_ && image.getWidth() == width_); CHECK(outV.getHeight() == outGrad.getHeight() && outV.getWidth() == outGrad.getWidth()); @@ -2080,12 +2060,12 @@ void CpuMatrix::maxPoolBackward(Matrix& image, } for (size_t c = 0; c < channels; ++c) { for (size_t ph = 0; ph < outputH; ++ph) { + int hstart = ph * strideH - paddingH; + int hend = std::min(hstart + sizeY, imgSizeH); + hstart = std::max(hstart, 0); for (size_t pw = 0; pw < outputW; ++pw) { - int hstart = ph * strideH - paddingH; int wstart = pw * strideW - paddingW; - int hend = std::min(hstart + sizeY, imgSizeH); int wend = std::min(wstart + sizeX, imgSizeW); - hstart = std::max(hstart, 0); wstart = std::max(wstart, 0); for (int h = hstart; h < hend; ++h) { for (int w = wstart; w < wend; ++w) { @@ -2098,10 +2078,10 @@ void CpuMatrix::maxPoolBackward(Matrix& image, } } // offset - inData += imgSizeH * imgSizeW; - tgtGrad += imgSizeH * imgSizeW; - otData += outputH * outputW; - otGrad += outputH * outputW; + inData += inLength; + tgtGrad += inLength; + otData += outLength; + otGrad += outLength; } } } @@ -2120,10 +2100,10 @@ void CpuMatrix::avgPoolForward(Matrix& input, size_t paddingW) { // The main loop size_t num = input.getHeight(); - size_t inHeight = imgSizeH; - size_t inWidth = imgSizeW; - CHECK(inHeight * inWidth * channels == input.getWidth()); - CHECK(outputH * outputW * channels * num == height_ * width_); + size_t inLength = imgSizeH * imgSizeW; + size_t outLength = outputH * outputW; + CHECK(inLength * channels == input.getWidth()); + CHECK(outLength * channels * num == height_ * width_); real* tgtData = data_; real* inData = input.getData(); @@ -2133,30 +2113,27 @@ void CpuMatrix::avgPoolForward(Matrix& input, } for (size_t c = 0; c < channels; ++c) { for (size_t ph = 0; ph < outputH; ++ph) { + int hstart = ph * strideH - paddingH; + int hend = std::min(hstart + sizeY, imgSizeH); + hstart = std::max(hstart, 0); for (size_t pw = 0; pw < outputW; ++pw) { - int hstart = ph * strideH - paddingH; int wstart = pw * strideW - paddingW; - int hend = std::min(hstart + sizeY, inHeight + paddingH); - int wend = std::min(wstart + sizeX, inWidth + paddingW); - int poolSize = (hend - hstart) * (wend - wstart); - hstart = std::max(hstart, 0); + int wend = std::min(wstart + sizeX, imgSizeW); wstart = std::max(wstart, 0); - hend = std::min(hend, static_cast(inHeight)); - wend = std::min(wend, static_cast(inWidth)); - - CHECK(poolSize); tgtData[ph * outputW + pw] = 0; // clear for (int h = hstart; h < hend; ++h) { for (int w = wstart; w < wend; ++w) { - tgtData[ph * outputW + pw] += inData[h * inWidth + w]; + tgtData[ph * outputW + pw] += inData[h * imgSizeW + w]; } } + int poolSize = (hend - hstart) * (wend - wstart); + CHECK(poolSize); tgtData[ph * outputW + pw] /= poolSize; } } // compute offset - inData += inHeight * inWidth; - tgtData += outputH * outputW; + inData += inLength; + tgtData += outLength; } } } @@ -2176,7 +2153,9 @@ void CpuMatrix::avgPoolBackward(Matrix& input, size_t paddingW) { size_t num = input.getHeight(); size_t channels = input.getWidth() / outputH / outputW; - CHECK(imgSizeH * imgSizeW * channels == getWidth()); + size_t inLength = imgSizeH * imgSizeW; + size_t outLength = outputH * outputW; + CHECK(inLength * channels == getWidth()); real* inData = input.getData(); real* outData = getData(); @@ -2186,16 +2165,14 @@ void CpuMatrix::avgPoolBackward(Matrix& input, } for (size_t c = 0; c < channels; ++c) { for (size_t ph = 0; ph < outputH; ++ph) { + int hstart = ph * strideH - paddingH; + int hend = std::min(hstart + sizeY, imgSizeH); + hstart = std::max(hstart, 0); for (size_t pw = 0; pw < outputW; ++pw) { - int hstart = ph * strideH - paddingH; int wstart = pw * strideW - paddingW; - int hend = std::min(hstart + sizeY, imgSizeH + paddingH); - int wend = std::min(wstart + sizeX, imgSizeW + paddingW); - int poolSize = (hend - hstart) * (wend - wstart); - hstart = std::max(hstart, 0); + int wend = std::min(wstart + sizeX, imgSizeW); wstart = std::max(wstart, 0); - hend = std::min(hend, static_cast(imgSizeH)); - wend = std::min(wend, static_cast(imgSizeW)); + int poolSize = (hend - hstart) * (wend - wstart); CHECK(poolSize); for (int h = hstart; h < hend; ++h) { @@ -2206,8 +2183,8 @@ void CpuMatrix::avgPoolBackward(Matrix& input, } } // offset - outData += imgSizeH * imgSizeW; - inData += outputH * outputW; + outData += inLength; + inData += outLength; } } } @@ -2234,12 +2211,11 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat, real* outData = getData(); real* maxPoolIdxData = maxPoolIdx.getData(); size_t num = inputMat.getHeight(); - size_t inWidth = imgSizeW; - size_t inHeight = imgSizeH; - size_t inDepth = imgSizeD; - CHECK(inHeight * inWidth * inDepth == inputMat.getWidth() / channels); + size_t inLength = imgSizeH * imgSizeW * imgSizeD; + size_t outLength = outputH * outputW * outputD; + CHECK(inLength == inputMat.getWidth() / channels); CHECK_EQ(num, this->getHeight()); - CHECK_EQ(channels * outputH * outputW * outputD, this->getWidth()); + CHECK_EQ(channels * outLength, this->getWidth()); size_t outStride = getStride(); /* initialize the data_ */ @@ -2258,16 +2234,16 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat, } for (size_t c = 0; c < channels; ++c) { // channel by channel for (size_t pd = 0; pd < outputD; ++pd) { + int dstart = pd * strideD - paddingD; + int dend = std::min(dstart + sizeZ, imgSizeD); + dstart = std::max(dstart, 0); for (size_t ph = 0; ph < outputH; ++ph) { + int hstart = ph * strideH - paddingH; + int hend = std::min(hstart + sizeY, imgSizeH); + hstart = std::max(hstart, 0); for (size_t pw = 0; pw < outputW; ++pw) { - int dstart = pd * strideD - paddingD; - int hstart = ph * strideH - paddingH; int wstart = pw * strideW - paddingW; - int dend = std::min(dstart + sizeZ, inDepth); - int hend = std::min(hstart + sizeY, inHeight); - int wend = std::min(wstart + sizeX, inWidth); - dstart = std::max(dstart, 0); - hstart = std::max(hstart, 0); + int wend = std::min(wstart + sizeX, imgSizeW); wstart = std::max(wstart, 0); int maxIdx = -1; real maxOutData = outData[(pd * outputH + ph) * outputW + pw]; @@ -2275,9 +2251,9 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat, for (int h = hstart; h < hend; ++h) { for (int w = wstart; w < wend; ++w) { if (maxOutData < - inputData[(d * inHeight + h) * inWidth + w]) { - maxOutData = inputData[(d * inHeight + h) * inWidth + w]; - maxIdx = (d * inHeight + h) * inWidth + w; + inputData[(d * imgSizeH + h) * imgSizeW + w]) { + maxOutData = inputData[(d * imgSizeH + h) * imgSizeW + w]; + maxIdx = (d * imgSizeH + h) * imgSizeW + w; } } } @@ -2288,9 +2264,9 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat, } } // compute offset - inputData += inDepth * inHeight * inWidth; - outData += outputD * outputH * outputW; - maxPoolIdxData += outputD * outputH * outputW; + inputData += inLength; + outData += outLength; + maxPoolIdxData += outLength; } } } @@ -2315,7 +2291,9 @@ void CpuMatrix::maxPool3DBackward(Matrix& outGrad, real scaleTargets, real scaleOutput) { size_t num = getHeight(); - size_t channels = size_t(width_ / imgSizeD / imgSizeH / imgSizeW); + size_t inLength = imgSizeH * imgSizeW * imgSizeD; + size_t outLength = outputH * outputW * outputD; + size_t channels = size_t(width_ / inLength); CHECK(maxPoolIdx.getHeight() == outGrad.getHeight() && maxPoolIdx.getWidth() == outGrad.getWidth()); @@ -2341,9 +2319,9 @@ void CpuMatrix::maxPool3DBackward(Matrix& outGrad, } } // offset - tgtGrad += imgSizeD * imgSizeH * imgSizeW; - otGrad += outputD * outputH * outputW; - maxPoolIdxData += outputD * outputH * outputW; + tgtGrad += inLength; + otGrad += outLength; + maxPoolIdxData += outLength; } } } @@ -2367,11 +2345,10 @@ void CpuMatrix::avgPool3DForward(Matrix& input, size_t paddingW) { // The main loop size_t num = input.getHeight(); - size_t inDepth = imgSizeD; - size_t inHeight = imgSizeH; - size_t inWidth = imgSizeW; - CHECK(inDepth * inHeight * inWidth * channels == input.getWidth()); - CHECK(outputD * outputH * outputW * channels * num == height_ * width_); + size_t inLength = imgSizeH * imgSizeW * imgSizeD; + size_t outLength = outputH * outputW * outputD; + CHECK(inLength * channels == input.getWidth()); + CHECK(outLength * channels * num == height_ * width_); real* tgtData = getData(); real* inData = input.getData(); @@ -2381,39 +2358,36 @@ void CpuMatrix::avgPool3DForward(Matrix& input, } for (size_t c = 0; c < channels; ++c) { for (size_t pd = 0; pd < outputD; ++pd) { + int dstart = pd * strideD - paddingD; + int dend = std::min(dstart + sizeZ, imgSizeD); + dstart = std::max(dstart, 0); for (size_t ph = 0; ph < outputH; ++ph) { + int hstart = ph * strideH - paddingH; + int hend = std::min(hstart + sizeY, imgSizeH); + hstart = std::max(hstart, 0); for (size_t pw = 0; pw < outputW; ++pw) { - int dstart = pd * strideD - paddingD; - int hstart = ph * strideH - paddingH; int wstart = pw * strideW - paddingW; - int dend = std::min(dstart + sizeZ, inDepth + paddingD); - int hend = std::min(hstart + sizeY, inHeight + paddingH); - int wend = std::min(wstart + sizeX, inWidth + paddingW); - int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart); - dstart = std::max(dstart, 0); - hstart = std::max(hstart, 0); + int wend = std::min(wstart + sizeX, imgSizeW); wstart = std::max(wstart, 0); - dend = std::min(dend, static_cast(inDepth)); - hend = std::min(hend, static_cast(inHeight)); - wend = std::min(wend, static_cast(inWidth)); - CHECK(poolSize); tgtData[(pd * outputH + ph) * outputW + pw] = 0; // clear for (int d = dstart; d < dend; ++d) { for (int h = hstart; h < hend; ++h) { for (int w = wstart; w < wend; ++w) { tgtData[(pd * outputH + ph) * outputW + pw] += - inData[(d * inHeight + h) * inWidth + w]; + inData[(d * imgSizeH + h) * imgSizeW + w]; } } } + int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart); + CHECK(poolSize); tgtData[(pd * outputH + ph) * outputW + pw] /= poolSize; } } } // compute offset - inData += inDepth * inHeight * inWidth; - tgtData += outputD * outputH * outputW; + inData += inLength; + tgtData += outLength; } } } @@ -2437,8 +2411,10 @@ void CpuMatrix::avgPool3DBackward(Matrix& input, real scaleTargets, real scaleOutput) { size_t num = input.getHeight(); - size_t channels = input.getWidth() / outputD / outputH / outputW; - CHECK(imgSizeD * imgSizeH * imgSizeW * channels == getWidth()); + size_t inLength = imgSizeH * imgSizeW * imgSizeD; + size_t outLength = outputH * outputW * outputD; + size_t channels = input.getWidth() / outLength; + CHECK(inLength * channels == getWidth()); real* inData = input.getData(); real* outData = getData(); @@ -2448,21 +2424,18 @@ void CpuMatrix::avgPool3DBackward(Matrix& input, } for (size_t c = 0; c < channels; ++c) { for (size_t pd = 0; pd < outputD; ++pd) { + int dstart = pd * strideD - paddingD; + int dend = std::min(dstart + sizeZ, imgSizeD); + dstart = std::max(dstart, 0); for (size_t ph = 0; ph < outputH; ++ph) { + int hstart = ph * strideH - paddingH; + int hend = std::min(hstart + sizeY, imgSizeH); + hstart = std::max(hstart, 0); for (size_t pw = 0; pw < outputW; ++pw) { - int dstart = pd * strideD - paddingD; - int hstart = ph * strideH - paddingH; int wstart = pw * strideW - paddingW; - int dend = std::min(dstart + sizeZ, imgSizeD + paddingD); - int hend = std::min(hstart + sizeY, imgSizeH + paddingH); - int wend = std::min(wstart + sizeX, imgSizeW + paddingW); - int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart); - dstart = std::max(dstart, 0); - hstart = std::max(hstart, 0); + int wend = std::min(wstart + sizeX, imgSizeW); wstart = std::max(wstart, 0); - dend = std::min(dend, static_cast(imgSizeD)); - hend = std::min(hend, static_cast(imgSizeH)); - wend = std::min(wend, static_cast(imgSizeW)); + int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart); CHECK(poolSize); for (int d = dstart; d < dend; ++d) { for (int h = hstart; h < hend; ++h) { @@ -2476,8 +2449,8 @@ void CpuMatrix::avgPool3DBackward(Matrix& input, } } // offset - outData += imgSizeD * imgSizeH * imgSizeW; - inData += outputD * outputH * outputW; + outData += inLength; + inData += outLength; } } } diff --git a/paddle/math/tests/test_matrixCompare.cpp b/paddle/math/tests/test_matrixCompare.cpp index 103f06acc57d7a23f019f5e713f6cacf2179e9e0..061fb22e3fd744d9d9895fd1008089e4a6ce6a0f 100644 --- a/paddle/math/tests/test_matrixCompare.cpp +++ b/paddle/math/tests/test_matrixCompare.cpp @@ -825,9 +825,8 @@ void testMaxPoolFwdBwd(int numSamples, int strideW, int padH, int padW) { - int outH = 0, outW = 0; - outH = (imgSizeH - ksizeH + 2 * padH + strideH - 1) / strideH + 1; - outW = (imgSizeW - ksizeW + 2 * padW + strideW - 1) / strideW + 1; + int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true); + int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true); int inWidth = imgSizeH * imgSizeW * channels; MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false); @@ -927,9 +926,8 @@ void testAvgPoolFwdBwd(int numSamples, int strideW, int padH, int padW) { - int outH = 0, outW = 0; - outH = (imgSizeH - ksizeH + 2 * padH + strideH - 1) / strideH + 1; - outW = (imgSizeW - ksizeW + 2 * padW + strideW - 1) / strideW + 1; + int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true); + int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true); int inWidth = imgSizeH * imgSizeW * channels; MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);