From 8067a42befef3eec593546c8fba11d3b802f2a62 Mon Sep 17 00:00:00 2001 From: Luo Tao Date: Mon, 18 Sep 2017 15:31:49 +0800 Subject: [PATCH] refine avg-pooling, which is exclusive. refine related code. --- paddle/cuda/src/hl_cuda_cnn.cu | 31 ++-- paddle/math/Matrix.cpp | 283 +++++++++++++++------------------ 2 files changed, 142 insertions(+), 172 deletions(-) diff --git a/paddle/cuda/src/hl_cuda_cnn.cu b/paddle/cuda/src/hl_cuda_cnn.cu index 9ba3d1426..596dd6d1b 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; @@ -299,11 +297,11 @@ __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 + padH); 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 poolsize = (hend - hstart) * (wend - wstart); gradient += outGrad[ph * pooledW + pw] / poolsize; @@ -600,16 +598,13 @@ __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; @@ -712,14 +707,14 @@ __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 + padD); for (int ph = phstart; ph < phend; ++ph) { + int hstart = ph * strideH - padH; + int hend = min(hstart + sizeY, height + padH); 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 poolsize = (dend - dstart) * (hend - hstart) * (wend - wstart); gradient += outGrad[(pd * pooledH + ph) * pooledW + pw] / poolsize; diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp index 4a2132c8d..83c06a04c 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 +1078,9 @@ 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(width_ == imgSizeW * imgSizeH * channels); CHECK(outGrad.getHeight() == outV.getHeight() && outGrad.getWidth() == outV.getWidth()); @@ -1093,8 +1089,8 @@ void GpuMatrix::maxPoolBackward(Matrix& inputMat, outData, outDiff, channels, - height, - width, + imgSizeH, + imgSizeW, outputH, outputW, sizeX, @@ -1125,17 +1121,15 @@ 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 +1160,15 @@ 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 +1206,16 @@ 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 +1258,17 @@ 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(width_ == imgSizeD * imgSizeH * imgSizeW * channels); 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 +1309,16 @@ 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 +1358,16 @@ 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 +1979,11 @@ 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 +2000,24 @@ 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 +2038,10 @@ 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 +2062,12 @@ 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 +2080,10 @@ 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 +2102,10 @@ 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 +2115,27 @@ 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 +2155,9 @@ 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 +2167,14 @@ 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 +2185,8 @@ void CpuMatrix::avgPoolBackward(Matrix& input, } } // offset - outData += imgSizeH * imgSizeW; - inData += outputH * outputW; + outData += inLength; + inData += outLength; } } } @@ -2234,12 +2213,11 @@ 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 +2236,16 @@ 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 +2253,9 @@ 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 +2266,9 @@ 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 +2293,9 @@ 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 +2321,9 @@ 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 +2347,10 @@ 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 +2360,36 @@ 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 +2413,10 @@ 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 +2426,18 @@ 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 +2451,8 @@ void CpuMatrix::avgPool3DBackward(Matrix& input, } } // offset - outData += imgSizeD * imgSizeH * imgSizeW; - inData += outputD * outputH * outputW; + outData += inLength; + inData += outLength; } } } -- GitLab