Skip to content

P
Paddle

wzpei9 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle

体验新版 GitCode，发现更多精彩内容 >>

未验证 提交 d7319c22 编写于 作者: Z Zhaolong Xing 提交者： GitHub
浏览文件
操作

Merge pull request #5165 from NHZlX/add_dilation 

Add dilation for exconv layer
上级 3e6f7684 f3818bd3
隐藏空白更改
内联 并排
Showing with 299 addition and 157 deletion
paddle/function/ConvOp.h
浏览文件 @ d7319c22
...@@ -61,6 +61,7 @@ public: ...@@ -61,6 +61,7 @@ public:
// function arguments // function arguments
strides_ = config.get<std::vector<size_t>>("strides"); strides_ = config.get<std::vector<size_t>>("strides");
paddings_ = config.get<std::vector<size_t>>("paddings"); paddings_ = config.get<std::vector<size_t>>("paddings");
dilations_ = config.get<std::vector<size_t>>("dilations");
groups_ = config.get<size_t>("groups"); groups_ = config.get<size_t>("groups");
// number of inputs and outputs // number of inputs and outputs
...@@ -118,6 +119,7 @@ protected: ...@@ -118,6 +119,7 @@ protected:
std::vector<size_t> strides_; std::vector<size_t> strides_;
std::vector<size_t> paddings_; std::vector<size_t> paddings_;
std::vector<size_t> dilations_;
/// Group size, refer to grouped convolution in /// Group size, refer to grouped convolution in
/// Alex Krizhevsky's paper: when group=2, the first half of the /// Alex Krizhevsky's paper: when group=2, the first half of the
...@@ -133,6 +135,10 @@ protected: ...@@ -133,6 +135,10 @@ protected:
inline int paddingW() const { return paddings_[1]; } inline int paddingW() const { return paddings_[1]; }
inline int dilationH() const { return dilations_[0]; }
inline int dilationW() const { return dilations_[1]; }
// A temporary memory in convolution calculation. // A temporary memory in convolution calculation.
MemoryHandlePtr memory_; MemoryHandlePtr memory_;
......
paddle/function/ConvOpTest.h
浏览文件 @ d7319c22
...@@ -79,45 +79,59 @@ void Convolution(const std::string& conv1, ...@@ -79,45 +79,59 @@ void Convolution(const std::string& conv1,
if (outputChannels < inputChannels) continue; if (outputChannels < inputChannels) continue;
for (size_t stride : {1, 2}) { for (size_t stride : {1, 2}) {
for (size_t padding : {0, 1}) { for (size_t padding : {0, 1}) {
if (padding >= filterSize) break; for (size_t dilation : {1, 3}) {
if (padding >= filterSize) break;
size_t filterS = (filterSize - 1) * dilation + 1;
// NNPACK only supports stride = 1 if batchSize > 1 if (inputSize + 2 * padding < filterS) break;
if ((conv1 == "NNPACKConv-CPU" || conv2 == "NNPACKConv-CPU") &&
batchSize > 1 && stride > 1)
break;
size_t outputSize = if ((conv1 == "NaiveConv-CPU" || conv2 == "NaiveConv-CPU" ||
(inputSize - filterSize + 2 * padding + stride) / stride; conv1 == "NNPACKConv-CPU" ||
VLOG(3) << " batchSize=" << batchSize conv2 == "NNPACKConv-CPU") &&
<< " inputChannels=" << inputChannels dilation > 1)
<< " inputHeight=" << inputSize break;
<< " inputWidth=" << inputSize
<< " outputChannels=" << outputChannels
<< " filterHeight=" << filterSize
<< " filterWidth=" << filterSize
<< " outputHeight=" << outputSize
<< " outputWidth=" << outputSize << " stride=" << stride
<< " padding=" << padding;
std::vector<size_t> paddings = {padding, padding}; // NNPACK only supports stride = 1 if batchSize > 1
std::vector<size_t> strides = {stride, stride}; if ((conv1 == "NNPACKConv-CPU" ||
Compare2Function<DType1, DType2> test( conv2 == "NNPACKConv-CPU") &&
conv1, batchSize > 1 && stride > 1)
conv2, break;
FuncConfig()
.set("paddings", paddings)
.set("strides", strides)
.set("groups", (size_t)1)
.set("algo", (std::string) "auto"));
TensorShape input{ size_t outputSize =
batchSize, inputChannels, inputSize, inputSize}; (inputSize - filterS + 2 * padding + stride) / stride;
TensorShape filter{ VLOG(3) << " batchSize=" << batchSize
outputChannels, inputChannels, filterSize, filterSize}; << " inputChannels=" << inputChannels
TensorShape output{ << " inputHeight=" << inputSize
batchSize, outputChannels, outputSize, outputSize}; << " inputWidth=" << inputSize
<< " outputChannels=" << outputChannels
<< " filterHeight=" << filterSize
<< " filterWidth=" << filterSize
<< " outputHeight=" << outputSize
<< " outputWidth=" << outputSize
<< " stride=" << stride << " padding=" << padding;
function(test, input, filter, output); std::vector<size_t> paddings = {padding, padding};
std::vector<size_t> strides = {stride, stride};
std::vector<size_t> dilations = {dilation, dilation};
Compare2Function<DType1, DType2> test(
conv1,
conv2,
FuncConfig()
.set("paddings", paddings)
.set("strides", strides)
.set("dilations", dilations)
.set("groups", (size_t)1)
.set("algo", (std::string) "auto"));
TensorShape input{
batchSize, inputChannels, inputSize, inputSize};
TensorShape filter{
outputChannels, inputChannels, filterSize, filterSize};
TensorShape output{
batchSize, outputChannels, outputSize, outputSize};
function(test, input, filter, output);
}
} }
} }
} }
...@@ -144,6 +158,7 @@ void Convolution2(const std::string& conv1, ...@@ -144,6 +158,7 @@ void Convolution2(const std::string& conv1,
for (size_t outputChannels : {7}) { for (size_t outputChannels : {7}) {
size_t stride = 1; size_t stride = 1;
size_t padding = 0; size_t padding = 0;
size_t dilation = 1;
size_t outputHeight = size_t outputHeight =
(inputHeight - filterHeight + 2 * padding + stride) / (inputHeight - filterHeight + 2 * padding + stride) /
stride; stride;
...@@ -162,6 +177,7 @@ void Convolution2(const std::string& conv1, ...@@ -162,6 +177,7 @@ void Convolution2(const std::string& conv1,
std::vector<size_t> paddings = {padding, padding}; std::vector<size_t> paddings = {padding, padding};
std::vector<size_t> strides = {stride, stride}; std::vector<size_t> strides = {stride, stride};
std::vector<size_t> dilations = {dilation, dilation};
Compare2Function<DType1, DType2> test( Compare2Function<DType1, DType2> test(
conv1, conv1,
conv2, conv2,
...@@ -169,6 +185,7 @@ void Convolution2(const std::string& conv1, ...@@ -169,6 +185,7 @@ void Convolution2(const std::string& conv1,
.set("paddings", paddings) .set("paddings", paddings)
.set("strides", strides) .set("strides", strides)
.set("groups", (size_t)1) .set("groups", (size_t)1)
.set("dilations", dilations)
.set("algo", (std::string) "auto")); .set("algo", (std::string) "auto"));
TensorShape input{ TensorShape input{
...@@ -223,6 +240,7 @@ void DepthwiseConvolution(const std::string& conv1, ...@@ -223,6 +240,7 @@ void DepthwiseConvolution(const std::string& conv1,
std::vector<size_t> paddings = {padding, padding}; std::vector<size_t> paddings = {padding, padding};
std::vector<size_t> strides = {stride, stride}; std::vector<size_t> strides = {stride, stride};
std::vector<size_t> dilations = {1, 1};
size_t groups = inputChannels; size_t groups = inputChannels;
Compare2Function<DType1, DType2> test( Compare2Function<DType1, DType2> test(
conv1, conv1,
...@@ -231,6 +249,7 @@ void DepthwiseConvolution(const std::string& conv1, ...@@ -231,6 +249,7 @@ void DepthwiseConvolution(const std::string& conv1,
.set("paddings", paddings) .set("paddings", paddings)
.set("strides", strides) .set("strides", strides)
.set("groups", groups) .set("groups", groups)
.set("dilations", dilations)
.set("algo", (std::string) "auto")); .set("algo", (std::string) "auto"));
TensorShape input{ TensorShape input{
......
paddle/function/GemmConvOp.cpp
浏览文件 @ d7319c22
...@@ -100,7 +100,9 @@ public: ...@@ -100,7 +100,9 @@ public:
strideH(), strideH(),
strideW(), strideW(),
paddingH(), paddingH(),
paddingW()); paddingW(),
dilationH(),
dilationW());
} else { } else {
colData = inputData + g * inputOffset; colData = inputData + g * inputOffset;
} }
...@@ -223,7 +225,9 @@ public: ...@@ -223,7 +225,9 @@ public:
strideH(), strideH(),
strideW(), strideW(),
paddingH(), paddingH(),
paddingW()); paddingW(),
dilationH(),
dilationW());
} }
} }
inputGrad += inputChannels * inputHeight * inputWidth; inputGrad += inputChannels * inputHeight * inputWidth;
...@@ -310,7 +314,9 @@ public: ...@@ -310,7 +314,9 @@ public:
strideH(), strideH(),
strideW(), strideW(),
paddingH(), paddingH(),
paddingW()); paddingW(),
dilationH(),
dilationW());
} else { } else {
colData = inputData + g * inputOffset; colData = inputData + g * inputOffset;
} }
......
paddle/function/Im2Col.h
浏览文件 @ d7319c22
...@@ -78,7 +78,9 @@ public: ...@@ -78,7 +78,9 @@ public:
int strideHeight, int strideHeight,
int strideWidth, int strideWidth,
int paddingHeight, int paddingHeight,
int paddingWidth); int paddingWidth,
int dilationHeight = 1,
int dilationWidth = 1);
}; };
template <ColFormat Format, DeviceType Device, class T> template <ColFormat Format, DeviceType Device, class T>
...@@ -91,7 +93,9 @@ public: ...@@ -91,7 +93,9 @@ public:
int strideHeight, int strideHeight,
int strideWidth, int strideWidth,
int paddingHeight, int paddingHeight,
int paddingWidth); int paddingWidth,
int dilationHeight = 1,
int dilationWidth = 1);
}; };
} // namespace paddle } // namespace paddle
paddle/function/Im2ColOp.cpp
浏览文件 @ d7319c22
...@@ -31,7 +31,9 @@ public: ...@@ -31,7 +31,9 @@ public:
int strideHeight, int strideHeight,
int strideWidth, int strideWidth,
int paddingHeight, int paddingHeight,
int paddingWidth) { int paddingWidth,
int dilationHeight,
int dilationWidth) {
int inputChannels = imShape[0]; int inputChannels = imShape[0];
int inputHeight = imShape[1]; int inputHeight = imShape[1];
int inputWidth = imShape[2]; int inputWidth = imShape[2];
...@@ -47,8 +49,8 @@ public: ...@@ -47,8 +49,8 @@ public:
int c_im = c / filterWidth / filterHeight; int c_im = c / filterWidth / filterHeight;
for (int h = 0; h < outputHeight; ++h) { for (int h = 0; h < outputHeight; ++h) {
for (int w = 0; w < outputWidth; ++w) { for (int w = 0; w < outputWidth; ++w) {
int imRowIdx = h * strideHeight + hOffset; int imRowIdx = h * strideHeight + hOffset * dilationHeight;
int imColIdx = w * strideWidth + wOffset; int imColIdx = w * strideWidth + wOffset * dilationWidth;
if ((imRowIdx - paddingHeight) < 0 || if ((imRowIdx - paddingHeight) < 0 ||
(imRowIdx - paddingHeight) >= inputHeight || (imRowIdx - paddingHeight) >= inputHeight ||
(imColIdx - paddingWidth) < 0 || (imColIdx - paddingWidth) < 0 ||
...@@ -81,7 +83,9 @@ public: ...@@ -81,7 +83,9 @@ public:
int strideHeight, int strideHeight,
int strideWidth, int strideWidth,
int paddingHeight, int paddingHeight,
int paddingWidth) { int paddingWidth,
int dilationHeight,
int dilationWidth) {
int inputChannels = imShape[0]; int inputChannels = imShape[0];
int inputHeight = imShape[1]; int inputHeight = imShape[1];
int inputWidth = imShape[2]; int inputWidth = imShape[2];
...@@ -97,8 +101,8 @@ public: ...@@ -97,8 +101,8 @@ public:
int c_im = c / filterWidth / filterHeight; int c_im = c / filterWidth / filterHeight;
for (int h = 0; h < outputHeight; ++h) { for (int h = 0; h < outputHeight; ++h) {
for (int w = 0; w < outputWidth; ++w) { for (int w = 0; w < outputWidth; ++w) {
int imRowIdx = h * strideHeight + hOffset; int imRowIdx = h * strideHeight + hOffset * dilationHeight;
int imColIdx = w * strideWidth + wOffset; int imColIdx = w * strideWidth + wOffset * dilationWidth;
if ((imRowIdx - paddingHeight) >= 0 && if ((imRowIdx - paddingHeight) >= 0 &&
(imRowIdx - paddingHeight) < inputHeight && (imRowIdx - paddingHeight) < inputHeight &&
(imColIdx - paddingWidth) >= 0 && (imColIdx - paddingWidth) >= 0 &&
...@@ -134,7 +138,9 @@ public: ...@@ -134,7 +138,9 @@ public:
int strideHeight, int strideHeight,
int strideWidth, int strideWidth,
int paddingHeight, int paddingHeight,
int paddingWidth) { int paddingWidth,
int dilationHeight = 1,
int dilationWidth = 1) {
int inputChannels = imShape[0]; int inputChannels = imShape[0];
int inputHeight = imShape[1]; int inputHeight = imShape[1];
int inputWidth = imShape[2]; int inputWidth = imShape[2];
...@@ -147,9 +153,10 @@ public: ...@@ -147,9 +153,10 @@ public:
for (int channel = 0; channel < inputChannels; ++channel) { for (int channel = 0; channel < inputChannels; ++channel) {
for (int filterH = 0; filterH < filterHeight; ++filterH) { for (int filterH = 0; filterH < filterHeight; ++filterH) {
for (int filterW = 0; filterW < filterWidth; ++filterW) { for (int filterW = 0; filterW < filterWidth; ++filterW) {
int imRowOffset = int imRowOffset = outputH * strideHeight +
outputH * strideHeight + filterH - paddingHeight; filterH * dilationHeight - paddingHeight;
int imColOffset = outputW * strideWidth + filterW - paddingWidth; int imColOffset = outputW * strideWidth +
filterW * dilationWidth - paddingWidth;
int colDataOffset = int colDataOffset =
(((outputH * outputWidth + outputW) * inputChannels + (((outputH * outputWidth + outputW) * inputChannels +
channel) * channel) *
...@@ -189,7 +196,9 @@ public: ...@@ -189,7 +196,9 @@ public:
int strideHeight, int strideHeight,
int strideWidth, int strideWidth,
int paddingHeight, int paddingHeight,
int paddingWidth) { int paddingWidth,
int dilationHeight = 1,
int dilationWidth = 1) {
int inputChannels = imShape[0]; int inputChannels = imShape[0];
int inputHeight = imShape[1]; int inputHeight = imShape[1];
int inputWidth = imShape[2]; int inputWidth = imShape[2];
...@@ -202,9 +211,10 @@ public: ...@@ -202,9 +211,10 @@ public:
for (int channel = 0; channel < inputChannels; ++channel) { for (int channel = 0; channel < inputChannels; ++channel) {
for (int filterH = 0; filterH < filterHeight; ++filterH) { for (int filterH = 0; filterH < filterHeight; ++filterH) {
for (int filterW = 0; filterW < filterWidth; ++filterW) { for (int filterW = 0; filterW < filterWidth; ++filterW) {
int imRowOffset = int imRowOffset = outputH * strideHeight +
outputH * strideHeight + filterH - paddingHeight; filterH * dilationHeight - paddingHeight;
int imColOffset = outputW * strideWidth + filterW - paddingWidth; int imColOffset = outputW * strideWidth +
filterW * dilationWidth - paddingWidth;
int colDataOffset = int colDataOffset =
(((outputH * outputWidth + outputW) * inputChannels + (((outputH * outputWidth + outputW) * inputChannels +
channel) * channel) *
......
paddle/function/Im2ColOpGpu.cu
浏览文件 @ d7319c22
...@@ -28,6 +28,8 @@ __global__ void im2col(const T* data_im, ...@@ -28,6 +28,8 @@ __global__ void im2col(const T* data_im,
int strideW, int strideW,
int paddingH, int paddingH,
int paddingW, int paddingW,
int dilationH,
int dilationW,
int height_col, int height_col,
int width_col, int width_col,
T* data_col) { T* data_col) {
...@@ -44,8 +46,8 @@ __global__ void im2col(const T* data_im, ...@@ -44,8 +46,8 @@ __global__ void im2col(const T* data_im,
data_col += (channel_out * height_col + h_out) * width_col + w_out; data_col += (channel_out * height_col + h_out) * width_col + w_out;
for (int i = 0; i < blockH; ++i) { for (int i = 0; i < blockH; ++i) {
for (int j = 0; j < blockW; ++j) { for (int j = 0; j < blockW; ++j) {
int rIdx = int(h_in + i); int rIdx = int(h_in + i * dilationH);
int cIdx = int(w_in + j); int cIdx = int(w_in + j * dilationW);
if ((rIdx - (int)paddingH) >= (int)height || if ((rIdx - (int)paddingH) >= (int)height ||
(rIdx - (int)paddingH) < 0 || (rIdx - (int)paddingH) < 0 ||
(cIdx - (int)paddingW) >= (int)width || (cIdx - (int)paddingW) >= (int)width ||
...@@ -77,7 +79,9 @@ public: ...@@ -77,7 +79,9 @@ public:
int strideHeight, int strideHeight,
int strideWidth, int strideWidth,
int paddingHeight, int paddingHeight,
int paddingWidth) { int paddingWidth,
int dilationHeight,
int dilationWidth) {
int inputChannels = imShape[0]; int inputChannels = imShape[0];
int inputHeight = imShape[1]; int inputHeight = imShape[1];
int inputWidth = imShape[2]; int inputWidth = imShape[2];
...@@ -102,6 +106,8 @@ public: ...@@ -102,6 +106,8 @@ public:
strideWidth, strideWidth,
paddingHeight, paddingHeight,
paddingWidth, paddingWidth,
dilationHeight,
dilationWidth,
outputHeight, outputHeight,
outputWidth, outputWidth,
colData); colData);
...@@ -121,6 +127,8 @@ __global__ void col2im(size_t n, ...@@ -121,6 +127,8 @@ __global__ void col2im(size_t n,
size_t strideW, size_t strideW,
size_t paddingH, size_t paddingH,
size_t paddingW, size_t paddingW,
size_t dilationH,
size_t dilationW,
size_t height_col, size_t height_col,
size_t width_col, size_t width_col,
T* data_im) { T* data_im) {
...@@ -131,23 +139,34 @@ __global__ void col2im(size_t n, ...@@ -131,23 +139,34 @@ __global__ void col2im(size_t n,
int w = int(index % width); int w = int(index % width);
int h = int((index / width) % height); int h = int((index / width) % height);
int c = int(index / (width * height)); int c = int(index / (width * height));
int filterH = (blockH - 1) * dilationH + 1;
int filterW = (blockW - 1) * dilationW + 1;
if ((w - (int)paddingW) >= 0 && if ((w - (int)paddingW) >= 0 &&
(w - (int)paddingW) < (width - 2 * paddingW) && (w - (int)paddingW) < (width - 2 * paddingW) &&
(h - (int)paddingH) >= 0 && (h - paddingH) < (height - 2 * paddingH)) { (h - (int)paddingH) >= 0 && (h - paddingH) < (height - 2 * paddingH)) {
// compute the start and end of the output // compute the start and end of the output
int w_col_start = int w_col_start =
(w < (int)blockW) ? 0 : (w - int(blockW)) / (int)strideW + 1; (w < (int)filterW) ? 0 : (w - int(filterW)) / (int)strideW + 1;
int w_col_end = min((int)(w / (int)strideW + 1), (int)(width_col)); int w_col_end = min((int)(w / (int)strideW + 1), (int)(width_col));
int h_col_start = int h_col_start =
(h < (int)blockH) ? 0 : (h - (int)blockH) / (int)strideH + 1; (h < (int)filterH) ? 0 : (h - (int)filterH) / (int)strideH + 1;
int h_col_end = min(int(h / strideH + 1), int(height_col)); int h_col_end = min(int(h / strideH + 1), int(height_col));
for (int h_col = h_col_start; h_col < h_col_end; ++h_col) { for (int h_col = h_col_start; h_col < h_col_end; ++h_col) {
for (int w_col = w_col_start; w_col < w_col_end; ++w_col) { for (int w_col = w_col_start; w_col < w_col_end; ++w_col) {
// the col location: [c * width * height + h_out, w_out] // the col location: [c * width * height + h_out, w_out]
int c_col = int(c * blockH * blockW) + int h_k = (h - h_col * strideH);
(h - h_col * (int)strideH) * (int)blockW + int w_k = (w - w_col * strideW);
(w - w_col * (int)strideW); if (h_k % dilationH == 0 && w_k % dilationW == 0) {
val += data_col[(c_col * height_col + h_col) * width_col + w_col]; h_k /= dilationH;
w_k /= dilationW;
int c_col =
(((c * blockH + h_k) * blockW + w_k) * height_col + h_col) *
width_col +
w_col;
val += data_col[c_col];
}
} }
} }
h -= paddingH; h -= paddingH;
...@@ -173,7 +192,9 @@ public: ...@@ -173,7 +192,9 @@ public:
int strideHeight, int strideHeight,
int strideWidth, int strideWidth,
int paddingHeight, int paddingHeight,
int paddingWidth) { int paddingWidth,
int dilationHeight,
int dilationWidth) {
int inputChannels = imShape[0]; int inputChannels = imShape[0];
int inputHeight = imShape[1]; int inputHeight = imShape[1];
int inputWidth = imShape[2]; int inputWidth = imShape[2];
...@@ -205,6 +226,8 @@ public: ...@@ -205,6 +226,8 @@ public:
strideWidth, strideWidth,
paddingHeight, paddingHeight,
paddingWidth, paddingWidth,
dilationHeight,
dilationWidth,
outputHeight, outputHeight,
outputWidth, outputWidth,
imData); imData);
...@@ -229,6 +252,8 @@ __global__ void im2colOCF(const T* imData, ...@@ -229,6 +252,8 @@ __global__ void im2colOCF(const T* imData,
int strideWidth, int strideWidth,
int paddingHeight, int paddingHeight,
int paddingWidth, int paddingWidth,
int dilationHeight,
int dilationWidth,
int outputHeight, int outputHeight,
int outputWidth) { int outputWidth) {
int swId = blockIdx.x; int swId = blockIdx.x;
...@@ -237,8 +262,10 @@ __global__ void im2colOCF(const T* imData, ...@@ -237,8 +262,10 @@ __global__ void im2colOCF(const T* imData,
channelId += blockDim.z) { channelId += blockDim.z) {
for (int idy = threadIdx.y; idy < filterHeight; idy += blockDim.y) { for (int idy = threadIdx.y; idy < filterHeight; idy += blockDim.y) {
for (int idx = threadIdx.x; idx < filterWidth; idx += blockDim.x) { for (int idx = threadIdx.x; idx < filterWidth; idx += blockDim.x) {
int widthOffset = idx + swId * strideWidth - paddingWidth; int widthOffset =
int heightOffset = idy + shId * strideHeight - paddingHeight; idx * dilationHeight + swId * strideWidth - paddingWidth;
int heightOffset =
idy * dilationWidth + shId * strideHeight - paddingHeight;
int imOffset = widthOffset + heightOffset * inputWidth + int imOffset = widthOffset + heightOffset * inputWidth +
channelId * inputHeight * inputWidth; channelId * inputHeight * inputWidth;
...@@ -273,7 +300,9 @@ public: ...@@ -273,7 +300,9 @@ public:
int strideHeight, int strideHeight,
int strideWidth, int strideWidth,
int paddingHeight, int paddingHeight,
int paddingWidth) { int paddingWidth,
int dilationHeight,
int dilationWidth) {
int inputChannels = imShape[0]; int inputChannels = imShape[0];
int inputHeight = imShape[1]; int inputHeight = imShape[1];
int inputWidth = imShape[2]; int inputWidth = imShape[2];
...@@ -312,6 +341,8 @@ public: ...@@ -312,6 +341,8 @@ public:
strideWidth, strideWidth,
paddingHeight, paddingHeight,
paddingWidth, paddingWidth,
dilationHeight,
dilationWidth,
outputHeight, outputHeight,
outputWidth); outputWidth);
CHECK_SYNC("Im2ColFunctor GPU failed"); CHECK_SYNC("Im2ColFunctor GPU failed");
...@@ -330,6 +361,8 @@ __global__ void col2imOCF(T* imData, ...@@ -330,6 +361,8 @@ __global__ void col2imOCF(T* imData,
int strideWidth, int strideWidth,
int paddingHeight, int paddingHeight,
int paddingWidth, int paddingWidth,
int dilationHeight,
int dilationWidth,
int outputHeight, int outputHeight,
int outputWidth) { int outputWidth) {
int swId = blockIdx.x; int swId = blockIdx.x;
...@@ -338,8 +371,10 @@ __global__ void col2imOCF(T* imData, ...@@ -338,8 +371,10 @@ __global__ void col2imOCF(T* imData,
channelId += blockDim.z) { channelId += blockDim.z) {
for (int idy = threadIdx.y; idy < filterHeight; idy += blockDim.y) { for (int idy = threadIdx.y; idy < filterHeight; idy += blockDim.y) {
for (int idx = threadIdx.x; idx < filterWidth; idx += blockDim.x) { for (int idx = threadIdx.x; idx < filterWidth; idx += blockDim.x) {
int widthOffset = idx + swId * strideWidth - paddingWidth; int widthOffset =
int heightOffset = idy + shId * strideHeight - paddingHeight; idx * dilationWidth + swId * strideWidth - paddingWidth;
int heightOffset =
idy * dilationHeight + shId * strideHeight - paddingHeight;
int imOffset = widthOffset + heightOffset * inputWidth + int imOffset = widthOffset + heightOffset * inputWidth +
channelId * inputHeight * inputWidth; channelId * inputHeight * inputWidth;
...@@ -372,7 +407,9 @@ public: ...@@ -372,7 +407,9 @@ public:
int strideHeight, int strideHeight,
int strideWidth, int strideWidth,
int paddingHeight, int paddingHeight,
int paddingWidth) { int paddingWidth,
int dilationHeight,
int dilationWidth) {
int inputChannels = imShape[0]; int inputChannels = imShape[0];
int inputHeight = imShape[1]; int inputHeight = imShape[1];
int inputWidth = imShape[2]; int inputWidth = imShape[2];
...@@ -411,6 +448,8 @@ public: ...@@ -411,6 +448,8 @@ public:
strideWidth, strideWidth,
paddingHeight, paddingHeight,
paddingWidth, paddingWidth,
dilationHeight,
dilationWidth,
outputHeight, outputHeight,
outputWidth); outputWidth);
CHECK_SYNC("Col2ImFunctor GPU failed"); CHECK_SYNC("Col2ImFunctor GPU failed");
......
paddle/function/Im2ColTest.cpp
浏览文件 @ d7319c22
...@@ -29,82 +29,98 @@ void TestIm2ColFunctor() { ...@@ -29,82 +29,98 @@ void TestIm2ColFunctor() {
for (size_t filterWidth : {3, 7}) { for (size_t filterWidth : {3, 7}) {
for (size_t stride : {1, 2}) { for (size_t stride : {1, 2}) {
for (size_t padding : {0, 1}) { for (size_t padding : {0, 1}) {
if (inputHeight <= filterHeight || inputWidth <= filterWidth) for (size_t dilation : {1, 3}) {
break; size_t filterSizeH = (filterHeight - 1) * dilation + 1;
if (padding >= filterHeight || padding >= filterWidth) break; size_t filterSizeW = (filterWidth - 1) * dilation + 1;
size_t outputHeight = if (inputHeight + 2 * padding < filterSizeH ||
(inputHeight - filterHeight + 2 * padding + stride) / inputWidth + 2 * padding < filterSizeW)
stride; break;
size_t outputWidth = if (padding >= filterSizeH || padding >= filterSizeW) break;
(inputWidth - filterWidth + 2 * padding + stride) / stride; size_t outputHeight =
(inputHeight - filterSizeH + 2 * padding) / stride + 1;
TensorShape imShape = size_t outputWidth =
TensorShape({channels, inputHeight, inputWidth}); (inputWidth - filterSizeW + 2 * padding) / stride + 1;
TensorShape colShape1 = TensorShape({channels,
filterHeight, TensorShape imShape =
filterWidth, TensorShape({channels, inputHeight, inputWidth});
outputHeight, TensorShape colShape1 = TensorShape({channels,
outputWidth}); filterHeight,
TensorShape colShape2 = TensorShape({outputHeight, filterWidth,
outputWidth, outputHeight,
channels, outputWidth});
filterHeight, TensorShape colShape2 = TensorShape({outputHeight,
filterWidth}); outputWidth,
channels,
size_t height = channels * filterHeight * filterWidth; filterHeight,
size_t width = outputHeight * outputWidth; filterWidth});
VectorPtr input1 = Vector::create(imShape.getElements(), false);
VectorPtr input2 = Vector::create(imShape.getElements(), false); size_t height = channels * filterHeight * filterWidth;
MatrixPtr output1 = Matrix::create(height, width, false, false); size_t width = outputHeight * outputWidth;
MatrixPtr output2 = Matrix::create(width, height, false, false); VectorPtr input1 =
input1->uniform(0.001, 1); Vector::create(imShape.getElements(), false);
input2->copyFrom(*input1); VectorPtr input2 =
Vector::create(imShape.getElements(), false);
Im2ColFunctor<kCFO, Device, T> im2Col1; MatrixPtr output1 =
Im2ColFunctor<kOCF, Device, T> im2Col2; Matrix::create(height, width, false, false);
im2Col1(input1->getData(), MatrixPtr output2 =
imShape, Matrix::create(width, height, false, false);
output1->getData(), input1->uniform(0.001, 1);
colShape1, input2->copyFrom(*input1);
stride,
stride, Im2ColFunctor<kCFO, Device, T> im2Col1;
padding, Im2ColFunctor<kOCF, Device, T> im2Col2;
padding); im2Col1(input1->getData(),
im2Col2(input2->getData(), imShape,
imShape, output1->getData(),
output2->getData(), colShape1,
colShape2, stride,
stride, stride,
stride, padding,
padding, padding,
padding); dilation,
dilation);
// The transposition of the result of ColFormat == kCFO im2Col2(input2->getData(),
// is equal to the result of ColFormat == kOCF. imShape,
MatrixPtr test; output2->getData(),
output2->transpose(test, true); colShape2,
autotest::TensorCheckErr(*output1, *test); stride,
stride,
Col2ImFunctor<kCFO, Device, T> col2Im1; padding,
Col2ImFunctor<kOCF, Device, T> col2Im2; padding,
col2Im1(input1->getData(), dilation,
imShape, dilation);
output1->getData(),
colShape1, // The transposition of the result of ColFormat == kCFO
stride, // is equal to the result of ColFormat == kOCF.
stride, MatrixPtr test;
padding, output2->transpose(test, true);
padding); autotest::TensorCheckErr(*output1, *test);
col2Im2(input2->getData(),
imShape, Col2ImFunctor<kCFO, Device, T> col2Im1;
output2->getData(), Col2ImFunctor<kOCF, Device, T> col2Im2;
colShape2,
stride, col2Im1(input1->getData(),
stride, imShape,
padding, output1->getData(),
padding); colShape1,
stride,
autotest::TensorCheckErr(*input1, *input2); stride,
padding,
padding,
dilation,
dilation);
col2Im2(input2->getData(),
imShape,
output2->getData(),
colShape2,
stride,
stride,
padding,
padding,
dilation,
dilation);
autotest::TensorCheckErr(*input1, *input2);
}
} }
} }
} }
......
paddle/gserver/layers/ExpandConvLayer.cpp
浏览文件 @ d7319c22
...@@ -79,6 +79,10 @@ bool ExpandConvLayer::init(const LayerMap &layerMap, ...@@ -79,6 +79,10 @@ bool ExpandConvLayer::init(const LayerMap &layerMap,
for (int i = 0; i < config_.inputs_size(); i++) { for (int i = 0; i < config_.inputs_size(); i++) {
std::vector<size_t> paddings = {(size_t)paddingY_[i], (size_t)padding_[i]}; std::vector<size_t> paddings = {(size_t)paddingY_[i], (size_t)padding_[i]};
std::vector<size_t> strides = {(size_t)strideY_[i], (size_t)stride_[i]}; std::vector<size_t> strides = {(size_t)strideY_[i], (size_t)stride_[i]};
std::vector<size_t> dilations = {(size_t)dilationY_[i],
(size_t)dilation_[i]};
bool useDilation = ((size_t)dilationY_[i] > 1 || (size_t)dilation_[i] > 1);
// Convolution Layer uses the GemmConv function by default. // Convolution Layer uses the GemmConv function by default.
convType = "GemmConv"; convType = "GemmConv";
...@@ -97,13 +101,14 @@ bool ExpandConvLayer::init(const LayerMap &layerMap, ...@@ -97,13 +101,14 @@ bool ExpandConvLayer::init(const LayerMap &layerMap,
#if defined(__ARM_NEON__) || defined(__ARM_NEON) #if defined(__ARM_NEON__) || defined(__ARM_NEON)
if ((filterSize_[i] == filterSizeY_[i]) && if ((filterSize_[i] == filterSizeY_[i]) &&
(filterSize_[i] == 3 || filterSize_[i] == 4) && (filterSize_[i] == 3 || filterSize_[i] == 4) &&
(stride_[i] == strideY_[i]) && (stride_[i] == 1 || stride_[i] == 2)) { (stride_[i] == strideY_[i]) && (stride_[i] == 1 || stride_[i] == 2) &&
!useDilation) {
convType = "NeonDepthwiseConv"; convType = "NeonDepthwiseConv";
} }
#endif #endif
} }
if (FLAGS_use_nnpack && !isDeconv_) { if (FLAGS_use_nnpack && !isDeconv_ && !useDilation) {
createFunction(forward_, createFunction(forward_,
"NNPACKConv", "NNPACKConv",
FuncConfig() FuncConfig()
...@@ -117,6 +122,7 @@ bool ExpandConvLayer::init(const LayerMap &layerMap, ...@@ -117,6 +122,7 @@ bool ExpandConvLayer::init(const LayerMap &layerMap,
FuncConfig() FuncConfig()
.set("paddings", paddings) .set("paddings", paddings)
.set("strides", strides) .set("strides", strides)
.set("dilations", dilations)
.set("groups", (size_t)groups_[i])); .set("groups", (size_t)groups_[i]));
createFunction(backward_, createFunction(backward_,
...@@ -124,6 +130,7 @@ bool ExpandConvLayer::init(const LayerMap &layerMap, ...@@ -124,6 +130,7 @@ bool ExpandConvLayer::init(const LayerMap &layerMap,
FuncConfig() FuncConfig()
.set("paddings", paddings) .set("paddings", paddings)
.set("strides", strides) .set("strides", strides)
.set("dilations", dilations)
.set("groups", (size_t)groups_[i])); .set("groups", (size_t)groups_[i]));
createFunction(backward_, createFunction(backward_,
...@@ -131,6 +138,7 @@ bool ExpandConvLayer::init(const LayerMap &layerMap, ...@@ -131,6 +138,7 @@ bool ExpandConvLayer::init(const LayerMap &layerMap,
FuncConfig() FuncConfig()
.set("paddings", paddings) .set("paddings", paddings)
.set("strides", strides) .set("strides", strides)
.set("dilations", dilations)
.set("groups", (size_t)groups_[i])); .set("groups", (size_t)groups_[i]));
} }
} }
......
paddle/gserver/tests/test_LayerGrad.cpp
浏览文件 @ d7319c22
...@@ -434,7 +434,7 @@ void testConvLayer(const string& type, bool trans, bool useGpu) { ...@@ -434,7 +434,7 @@ void testConvLayer(const string& type, bool trans, bool useGpu) {
config.layerConfig.set_partial_sum(1); config.layerConfig.set_partial_sum(1);
config.layerConfig.set_shared_biases(true); config.layerConfig.set_shared_biases(true);
int dilation = 1; int dilation = 2;
if (type == "cudnn_conv") { if (type == "cudnn_conv") {
#if CUDNN_VERSION >= 6000 #if CUDNN_VERSION >= 6000
dilation = 2; dilation = 2;
......
python/paddle/trainer/config_parser.py
浏览文件 @ d7319c22
...@@ -1200,8 +1200,14 @@ def TestData(data_config, async_load_data=None): ...@@ -1200,8 +1200,14 @@ def TestData(data_config, async_load_data=None):
#caffe_mode: compute the output size using floor instead of ceil, #caffe_mode: compute the output size using floor instead of ceil,
# which is consistent of caffe and CuDNN's convention. # which is consistent of caffe and CuDNN's convention.
def cnn_output_size(img_size, filter_size, padding, stride, caffe_mode): def cnn_output_size(img_size,
output = (2 * padding + img_size - filter_size) / float(stride) filter_size,
padding,
stride,
caffe_mode,
dilation=1):
filter_s = (filter_size - 1) * dilation + 1
output = (2 * padding + img_size - filter_s) / float(stride)
if caffe_mode: if caffe_mode:
return 1 + int(math.floor(output)) return 1 + int(math.floor(output))
else: else:
...@@ -1210,8 +1216,14 @@ def cnn_output_size(img_size, filter_size, padding, stride, caffe_mode): ...@@ -1210,8 +1216,14 @@ def cnn_output_size(img_size, filter_size, padding, stride, caffe_mode):
#calcualte image_size based on output_size for de-convolution (ConvTransLayer). #calcualte image_size based on output_size for de-convolution (ConvTransLayer).
#It is the reverse function of cnn_output_size #It is the reverse function of cnn_output_size
def cnn_image_size(output_size, filter_size, padding, stride, caffe_mode): def cnn_image_size(output_size,
img_size = (output_size - 1) * stride + filter_size - 2 * padding filter_size,
padding,
stride,
caffe_mode,
dilation=1):
filter_s = (filter_size - 1) * dilation + 1
img_size = (output_size - 1) * stride + filter_s - 2 * padding
if not caffe_mode: if not caffe_mode:
img_size = img_size + 1 img_size = img_size + 1
return img_size return img_size
...@@ -1376,6 +1388,12 @@ def parse_conv(conv, input_layer_name, conv_conf, num_filters, trans=False): ...@@ -1376,6 +1388,12 @@ def parse_conv(conv, input_layer_name, conv_conf, num_filters, trans=False):
conv_conf.stride_y = conv.stride_y conv_conf.stride_y = conv.stride_y
conv_conf.groups = conv.groups conv_conf.groups = conv.groups
conv_conf.caffe_mode = conv.caffe_mode conv_conf.caffe_mode = conv.caffe_mode
if not conv.dilation:
conv.dilation = 1
conv.dilation_y = 1
else:
conv_conf.dilation = conv.dilation
conv_conf.dilation_y = conv.dilation_y
if not trans: if not trans:
conv_conf.filter_channels = conv.channels / conv.groups conv_conf.filter_channels = conv.channels / conv.groups
...@@ -1383,20 +1401,20 @@ def parse_conv(conv, input_layer_name, conv_conf, num_filters, trans=False): ...@@ -1383,20 +1401,20 @@ def parse_conv(conv, input_layer_name, conv_conf, num_filters, trans=False):
get_img_size(input_layer_name, conv.channels) get_img_size(input_layer_name, conv.channels)
conv_conf.output_x = cnn_output_size( conv_conf.output_x = cnn_output_size(
conv_conf.img_size, conv_conf.filter_size, conv_conf.padding, conv_conf.img_size, conv_conf.filter_size, conv_conf.padding,
conv_conf.stride, conv_conf.caffe_mode) conv_conf.stride, conv_conf.caffe_mode, conv.dilation)
conv_conf.output_y = cnn_output_size( conv_conf.output_y = cnn_output_size(
conv_conf.img_size_y, conv_conf.filter_size_y, conv_conf.padding_y, conv_conf.img_size_y, conv_conf.filter_size_y, conv_conf.padding_y,
conv_conf.stride_y, conv_conf.caffe_mode) conv_conf.stride_y, conv_conf.caffe_mode, conv.dilation_y)
else: else:
conv_conf.filter_channels = num_filters / conv.groups conv_conf.filter_channels = num_filters / conv.groups
conv_conf.output_x, conv_conf.output_y = \ conv_conf.output_x, conv_conf.output_y = \
get_img_size(input_layer_name, conv.channels) get_img_size(input_layer_name, conv.channels)
conv_conf.img_size = cnn_image_size( conv_conf.img_size = cnn_image_size(
conv_conf.output_x, conv_conf.filter_size, conv_conf.padding, conv_conf.output_x, conv_conf.filter_size, conv_conf.padding,
conv_conf.stride, conv_conf.caffe_mode) conv_conf.stride, conv_conf.caffe_mode, conv.dilation)
conv_conf.img_size_y = cnn_image_size( conv_conf.img_size_y = cnn_image_size(
conv_conf.output_y, conv_conf.filter_size_y, conv_conf.padding_y, conv_conf.output_y, conv_conf.filter_size_y, conv_conf.padding_y,
conv_conf.stride_y, conv_conf.caffe_mode) conv_conf.stride_y, conv_conf.caffe_mode, conv.dilation_y)
#caffe_mode: compute the output size using floor instead of ceil, #caffe_mode: compute the output size using floor instead of ceil,
......
python/paddle/trainer_config_helpers/layers.py
浏览文件 @ d7319c22
...@@ -2571,7 +2571,9 @@ def img_conv_layer(input, ...@@ -2571,7 +2571,9 @@ def img_conv_layer(input,
if layer_type: if layer_type:
if dilation > 1 or dilation_y > 1: if dilation > 1 or dilation_y > 1:
assert layer_type in ["cudnn_conv", "cudnn_convt"] assert layer_type in [
"cudnn_conv", "cudnn_convt", "exconv", "exconvt"
]
if trans: if trans:
assert layer_type in ["exconvt", "cudnn_convt"] assert layer_type in ["exconvt", "cudnn_convt"]
else: else:
......
python/paddle/trainer_config_helpers/tests/configs/protostr/img_layers.protostr
浏览文件 @ d7319c22
...@@ -28,6 +28,8 @@ layers { ...@@ -28,6 +28,8 @@ layers {
stride_y: 1 stride_y: 1
output_y: 227 output_y: 227
img_size_y: 256 img_size_y: 256
dilation: 1
dilation_y: 1
} }
} }
bias_parameter_name: "___conv_0__.wbias" bias_parameter_name: "___conv_0__.wbias"
......
python/paddle/trainer_config_helpers/tests/configs/protostr/img_trans_layers.protostr
浏览文件 @ d7319c22
...@@ -28,6 +28,8 @@ layers { ...@@ -28,6 +28,8 @@ layers {
stride_y: 1 stride_y: 1
output_y: 227 output_y: 227
img_size_y: 256 img_size_y: 256
dilation: 1
dilation_y: 1
} }
} }
bias_parameter_name: "___conv_0__.wbias" bias_parameter_name: "___conv_0__.wbias"
......
python/paddle/trainer_config_helpers/tests/configs/protostr/test_bilinear_interp.protostr
浏览文件 @ d7319c22
...@@ -28,6 +28,8 @@ layers { ...@@ -28,6 +28,8 @@ layers {
stride_y: 1 stride_y: 1
output_y: 48 output_y: 48
img_size_y: 48 img_size_y: 48
dilation: 1
dilation_y: 1
} }
} }
bias_parameter_name: "___conv_0__.wbias" bias_parameter_name: "___conv_0__.wbias"
......
python/paddle/trainer_config_helpers/tests/configs/protostr/test_maxout.protostr
浏览文件 @ d7319c22
...@@ -30,6 +30,8 @@ layers { ...@@ -30,6 +30,8 @@ layers {
stride_y: 1 stride_y: 1
output_y: 48 output_y: 48
img_size_y: 48 img_size_y: 48
dilation: 1
dilation_y: 1
} }
} }
bias_parameter_name: "___conv_0__.wbias" bias_parameter_name: "___conv_0__.wbias"
...@@ -105,6 +107,8 @@ layers { ...@@ -105,6 +107,8 @@ layers {
stride_y: 1 stride_y: 1
output_y: 24 output_y: 24
img_size_y: 24 img_size_y: 24
dilation: 1
dilation_y: 1
} }
} }
bias_parameter_name: "___conv_1__.wbias" bias_parameter_name: "___conv_1__.wbias"
......
python/paddle/trainer_config_helpers/tests/configs/protostr/test_pad.protostr
浏览文件 @ d7319c22
...@@ -30,6 +30,8 @@ layers { ...@@ -30,6 +30,8 @@ layers {
stride_y: 1 stride_y: 1
output_y: 48 output_y: 48
img_size_y: 48 img_size_y: 48
dilation: 1
dilation_y: 1
} }
} }
bias_parameter_name: "___conv_0__.wbias" bias_parameter_name: "___conv_0__.wbias"
......
python/paddle/trainer_config_helpers/tests/configs/protostr/test_roi_pool_layer.protostr
浏览文件 @ d7319c22
...@@ -36,6 +36,8 @@ layers { ...@@ -36,6 +36,8 @@ layers {
stride_y: 1 stride_y: 1
output_y: 14 output_y: 14
img_size_y: 14 img_size_y: 14
dilation: 1
dilation_y: 1
} }
} }
bias_parameter_name: "___conv_0__.wbias" bias_parameter_name: "___conv_0__.wbias"
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或