m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

fix conflict

Browse Source
enforce_failed
chengduoZH 8 years ago
parent 536dbc0cb6 fcad0a3a4b
commit a55dd22649
33 changed files with 4528 additions and 34 deletions
Show Stats Download Patch File Download Diff File

92
paddle/cuda/include/hl_cnn.h
Unescape View File

@ -173,6 +173,96 @@ extern void hl_avgpool_backward(const int frameCnt,
real* backGrad,
const int outStride);
extern void hl_maxpool3D_forward(const int frameCnt,
const real* inputData,
const int channels,
const int depth,
const int height,
const int width,
const int pooledD,
const int pooledH,
const int pooledW,
const int sizeZ,
const int sizeY,
const int sizeX,
const int strideD,
const int strideH,
const int strideW,
const int paddingD,
const int paddingH,
const int paddingW,
real* tgtData,
real* maxPoolIdxData,
const int tgtStride);
extern void hl_maxpool3D_backward(const int frameCnt,
const real* outGrad,
const int channels,
const int depth,
const int height,
const int width,
const int pooledD,
const int pooledH,
const int pooledW,
const int sizeZ,
const int sizeY,
const int sizeX,
const int strideD,
const int strideH,
const int strideW,
const int paddingD,
const int paddingH,
const int paddingW,
real scaleA,
real scaleB,
real* targetGrad,
real* maxPoolIdxData,
const int outStride);
extern void hl_avgpool3D_forward(const int frameCnt,
const real* inputData,
const int channels,
const int depth,
const int height,
const int width,
const int pooledD,
const int pooledH,
const int pooledW,
const int sizeZ,
const int sizeY,
const int sizeX,
const int strideD,
const int strideH,
const int strideW,
const int paddingD,
const int paddingH,
const int paddingW,
real* tgtData,
const int tgtStride);
extern void hl_avgpool3D_backward(const int frameCnt,
const real* outGrad,
const int channels,
const int depth,
const int height,
const int width,
const int pooledD,
const int pooledH,
const int pooledW,
const int sizeZ,
const int sizeY,
const int sizeX,
const int strideD,
const int strideH,
const int strideW,
int paddingD,
int paddingH,
int paddingW,
real scaleA,
real scaleB,
real* backGrad,
const int outStride);
/**
* @brief Bilinear interpolation forward.
*
@ -275,4 +365,4 @@ extern void hl_maxout_backward(real* inGrad,
size_t featLen,
size_t groups);
#endif /* HL_CNN_H_ */
#endif // HL_CNN_H_

76
paddle/cuda/include/hl_matrix.h
Unescape View File

@ -224,4 +224,80 @@ extern void hl_matrix_collect_shared_bias(real* B_d,
extern void hl_matrix_rotate(
real* mat, real* matRot, int dimM, int dimN, bool clockWise);
/**
* @brief Matrix vol2Col: Convert 3D volume into col matrix
*
* @param[in] matSrc input matrix.
* @param[in] channel channel of matSrc.
* @param[in] depth depth of matSrc.
* @param[in] height height of matSrc.
* @param[in] width width of matSrc.
* @param[in] filterD depth of filter.
* @param[in] filterH height of filter.
* @param[in] filterW width of filter.
* @param[in] strideD stride in the depth.
* @param[in] strideH stride in the height.
* @param[in] strideW stride in the width.
* @param[in] paddingD padding in the depth.
* @param[in] paddingH padding in the height.
* @param[in] paddingW padding in the width.
* @param[out] dataDst output matrix.
*
*/
extern void hl_matrix_vol2Col(const real* dataSrc,
int channels,
int depth,
int height,
int width,
int filterD,
int filterH,
int filterW,
int strideD,
int strideH,
int strideW,
int paddingD,
int paddingH,
int paddingW,
real* dataDst);
/**
* @brief Matrix col2Vol: Convert col matrix into 3D volume
*
* @param[out] matDst output matrix.
* @param[in] channel channel of matDst.
* @param[in] depth depth of matDst.
* @param[in] height height of matDst.
* @param[in] width width of matDst.
* @param[in] filterD depth of filter.
* @param[in] filterH height of filter.
* @param[in] filterW width of filter.
* @param[in] strideD stride in the depth.
* @param[in] strideH stride in the height.
* @param[in] strideW stride in the width.
* @param[in] paddingD padding in the depth.
* @param[in] paddingH padding in the height.
* @param[in] paddingW padding in the width.
* @param[in] matSrc input matrix.
* @param[in] beta input
* @param[in] alpha input
*
*/
extern void hl_matrix_col2Vol(real* dataDst,
int channels,
int depth,
int height,
int width,
int filterD,
int filterH,
int filterW,
int strideD,
int strideH,
int strideW,
int paddingD,
int paddingH,
int paddingW,
const real* dataSrc,
real alpha,
real beta);
#endif /* HL_MATRIX_H_ */

90
paddle/cuda/include/stub/hl_cnn_stub.h
Unescape View File

@ -87,6 +87,96 @@ inline void hl_avgpool_backward(const int frameCnt,
real* backGrad,
const int outStride) {}
inline void hl_maxpool3D_forward(const int frameCnt,
const real* inputData,
const int channels,
const int depth,
const int height,
const int width,
const int pooledD,
const int pooledH,
const int pooledW,
const int sizeZ,
const int sizeY,
const int sizeX,
const int strideD,
const int strideH,
const int strideW,
const int paddingD,
const int paddingH,
const int paddingW,
real* tgtData,
real* maxPoolIdxData,
const int tgtStride) {}
inline void hl_maxpool3D_backward(const int frameCnt,
const real* outGrad,
const int channels,
const int depth,
const int height,
const int width,
const int pooledD,
const int pooledH,
const int pooledW,
const int sizeZ,
const int sizeY,
const int sizeX,
const int strideD,
const int strideH,
const int strideW,
const int paddingD,
const int paddingH,
const int paddingW,
real scaleA,
real scaleB,
real* targetGrad,
real* maxPoolIdxData,
const int outStride) {}
inline void hl_avgpool3D_forward(const int frameCnt,
const real* inputData,
const int channels,
const int depth,
const int height,
const int width,
const int pooledD,
const int pooledH,
const int pooledW,
const int sizeZ,
const int sizeY,
const int sizeX,
const int strideD,
const int strideH,
const int strideW,
const int paddingD,
const int paddingH,
const int paddingW,
real* tgtData,
const int tgtStride) {}
inline void hl_avgpool3D_backward(const int frameCnt,
const real* outGrad,
const int channels,
const int depth,
const int height,
const int width,
const int pooledD,
const int pooledH,
const int pooledW,
const int sizeZ,
const int sizeY,
const int sizeX,
const int strideD,
const int strideH,
const int strideW,
const int paddingD,
const int paddingH,
const int paddingW,
real scaleA,
real scaleB,
real* backGrad,
const int outStride) {}
inline void hl_bilinear_forward(const real* inData,
const size_t inImgH,
const size_t inImgW,

34
paddle/cuda/include/stub/hl_matrix_stub.h
Unescape View File

@ -99,4 +99,38 @@ inline void hl_matrix_collect_shared_bias(real* B_d,
inline void hl_matrix_rotate(
real* mat, real* matRot, int dimM, int dimN, bool clockWise) {}
inline void hl_matrix_vol2Col(const real* dataSrc,
int channels,
int depth,
int height,
int width,
int filterD,
int filterH,
int filterW,
int strideD,
int strideH,
int strideW,
int paddingD,
int paddingH,
int paddingW,
real* dataDst) {}
inline void hl_matrix_col2Vol(real* dataDst,
int channels,
int depth,
int height,
int width,
int filterD,
int filterH,
int filterW,
int strideD,
int strideH,
int strideW,
int paddingD,
int paddingH,
int paddingW,
const real* dataSrc,
real alpha,
real beta) {}
#endif // HL_MATRIX_STUB_H_

427
paddle/cuda/src/hl_cuda_cnn.cu
View File

File diff suppressed because it is too large Load Diff

201
paddle/cuda/src/hl_cuda_matrix.cu
Unescape View File

@ -592,3 +592,204 @@ void hl_matrix_rotate(
mat, matRot, dimM, dimN, clockWise);
CHECK_SYNC("hl_matrix_rotate failed");
}
__global__ void keMatrixVol2Col(int num_kernels,
const real* dataSrc,
real* dataDst,
int depth,
int height,
int width,
int filterD,
int filterH,
int filterW,
int strideD,
int strideH,
int strideW,
int paddingD,
int paddingH,
int paddingW,
int depth_col,
int height_col,
int width_col) {
for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < num_kernels;
index += blockDim.x * gridDim.x) {
int w_out = index % width_col;
int h_out = (index / width_col) % height_col;
int d_out = (index / width_col / height_col) % depth_col;
int channel_in = index / width_col / height_col / depth_col;
int channel_out = channel_in * filterD * filterH * filterW;
int w_in = w_out * strideW - paddingW;
int h_in = h_out * strideH - paddingH;
int d_in = d_out * strideD - paddingD;
dataDst +=
((channel_out * depth_col + d_out) * height_col + h_out) * width_col +
w_out;
dataSrc += ((channel_in * depth + d_in) * height + h_in) * width + w_in;
for (int k = 0; k < filterD; ++k) {
for (int i = 0; i < filterH; ++i) {
for (int j = 0; j < filterW; ++j) {
int d = d_in + k;
int h = h_in + i;
int w = w_in + j;
*dataDst = (d >= 0 && d < depth && h >= 0 && h < height && w >= 0 &&
w < width)
? dataSrc[(k * height + i) * width + j]
: 0;
dataDst += depth_col * height_col * width_col;
}
}
}
}
}
void hl_matrix_vol2Col(const real* dataSrc,
int channels,
int depth,
int height,
int width,
int filterD,
int filterH,
int filterW,
int strideD,
int strideH,
int strideW,
int paddingD,
int paddingH,
int paddingW,
real* dataDst) {
int depth_col = (depth + 2 * paddingD - filterD) / strideD + 1;
int height_col = (height + 2 * paddingH - filterH) / strideH + 1;
int width_col = (width + 2 * paddingW - filterW) / strideW + 1;
int num_kernels = channels * depth_col * height_col * width_col;
const int threads = 512;
const int blocks = DIVUP(num_kernels, threads);
keMatrixVol2Col<<<blocks, threads, 0, STREAM_DEFAULT>>>(num_kernels,
dataSrc,
dataDst,
depth,
height,
width,
filterD,
filterH,
filterW,
strideD,
strideH,
strideW,
paddingD,
paddingH,
paddingW,
depth_col,
height_col,
width_col);
CHECK_SYNC("hl_matrix_vol2Col failed");
}
__global__ void keMatrixCol2Vol(int num_kernels,
real* dataDst,
const real* dataSrc,
int depth,
int height,
int width,
int filterD,
int filterH,
int filterW,
int strideD,
int strideH,
int strideW,
int paddingD,
int paddingH,
int paddingW,
int depth_col,
int height_col,
int width_col,
real alpha,
real beta) {
for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < num_kernels;
index += blockDim.x * gridDim.x) {
real srcVal = 0;
real dstVal = dataDst[index];
int w = index % width + paddingW;
int h = (index / width) % height + paddingH;
int d = (index / width / height) % depth + paddingD;
int c = index / width / height / depth;
// compute the start and end of the output
int w_col_start = (w < filterW) ? 0 : (w - filterW) / strideW + 1;
int w_col_end = min(w / strideW + 1, width_col);
int h_col_start = (h < filterH) ? 0 : (h - filterH) / strideH + 1;
int h_col_end = min(h / strideH + 1, height_col);
int d_col_start = (d < filterD) ? 0 : (d - filterD) / strideD + 1;
int d_col_end = min(d / strideD + 1, depth_col);
int offset = (c * filterD * filterW * filterH + d * filterW * filterH +
h * filterW + w) *
depth_col * height_col * width_col;
int coeff_d_col =
(1 - strideD * filterW * filterH * depth_col) * height_col * width_col;
int coeff_h_col =
(1 - strideH * filterW * depth_col * height_col) * width_col;
int coeff_w_col = (1 - strideW * depth_col * height_col * width_col);
for (int d_col = d_col_start; d_col < d_col_end; ++d_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) {
srcVal += dataSrc[offset + d_col * coeff_d_col + h_col * coeff_h_col +
w_col * coeff_w_col];
}
}
}
dataDst[index] = alpha * srcVal + beta * dstVal;
}
}
void hl_matrix_col2Vol(real* dataDst,
int channels,
int depth,
int height,
int width,
int filterD,
int filterH,
int filterW,
int strideD,
int strideH,
int strideW,
int paddingD,
int paddingH,
int paddingW,
const real* dataSrc,
real alpha,
real beta) {
int depth_col = (depth + 2 * paddingD - filterD) / strideD + 1;
int height_col = (height + 2 * paddingH - filterH) / strideH + 1;
int width_col = (width + 2 * paddingW - filterW) / strideW + 1;
int num_kernels = channels * depth * height * width;
const int threads = 512;
const int blocks = DIVUP(num_kernels, threads);
keMatrixCol2Vol<<<blocks, threads, 0, STREAM_DEFAULT>>>(num_kernels,
dataDst,
dataSrc,
depth,
height,
width,
filterD,
filterH,
filterW,
strideD,
strideH,
strideW,
paddingD,
paddingH,
paddingW,
depth_col,
height_col,
width_col,
alpha,
beta);
CHECK_SYNC("hl_matrix_col2Vol failed");
}

244
paddle/gserver/layers/Conv3DLayer.cpp
Unescape View File

@ -0,0 +1,244 @@
/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "Conv3DLayer.h"
#include "paddle/utils/Logging.h"
#include "paddle/utils/Stat.h"
namespace paddle {
REGISTER_LAYER(conv3d, Conv3DLayer);
bool Conv3DLayer::init(const LayerMap &layerMap,
const ParameterMap &parameterMap) {
if (!ConvBaseLayer::init(layerMap, parameterMap)) return false;
int index = 0;
for (auto &inputConfig : config_.inputs()) {
const ConvConfig &conf = inputConfig.conv_conf();
M_.push_back(numFilters_ / conf.groups());
K_.push_back(filterPixels_[index] * filterChannels_[index]);
// create a new weight
size_t height, width;
width = filterPixels_[index] * filterChannels_[index];
height = numFilters_;
CHECK_EQ(parameters_[index]->getSize(), width * height);
Weight *w = new Weight(height, width, parameters_[index]);
weights_.emplace_back(w);
++index;
}
if (biasParameter_.get()) {
if (sharedBiases_) {
CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
biases_ =
std::unique_ptr<Weight>(new Weight(1, numFilters_, biasParameter_));
} else {
biases_ =
std::unique_ptr<Weight>(new Weight(1, getSize(), biasParameter_));
}
}
return true;
}
size_t Conv3DLayer::getSize() {
CHECK_NE(inputLayers_.size(), 0UL);
outputH_.clear();
outputW_.clear();
outputD_.clear();
N_.clear();
size_t layerSize = 0;
for (size_t i = 0; i < inputLayers_.size(); ++i) {
outputW_.push_back(outputSize(
imgSizeW_[i], filterSize_[i], padding_[i], stride_[i], true));
outputH_.push_back(outputSize(
imgSizeH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true));
outputD_.push_back(outputSize(
imgSizeD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true));
N_.push_back(outputD_[i] * outputH_[i] * outputW_[i]);
CHECK(layerSize == 0 || N_[i] * size_t(numFilters_) == layerSize);
layerSize += N_[i] * numFilters_;
}
getOutput().setFrameHeight(outputH_[0]);
getOutput().setFrameWidth(outputW_[0]);
getOutput().setFrameDepth(outputD_[0]);
return layerSize;
}
void Conv3DLayer::forward(PassType passType) {
Layer::forward(passType);
int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
int outWidth = getSize();
resetOutput(batchSize, outWidth);
for (size_t i = 0; i != inputLayers_.size(); ++i) {
REGISTER_TIMER_INFO("FwdConv3D", getName().c_str());
const MatrixPtr &inMat = getInputValue(i);
const MatrixPtr &outMat = getOutputValue();
int M = M_[i];
int N = N_[i];
int K = K_[i];
Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
MatrixPtr wMat = weights_[i]->getW();
for (int n = 0; n < batchSize; ++n) {
colBuf_->vol2Col(inMat->getData() + n * inMat->getStride(),
channels_[i],
imgSizeD_[i],
imgSizeH_[i],
imgSizeW_[i],
filterSizeZ_[i],
filterSizeY_[i],
filterSize_[i],
strideZ_[i],
strideY_[i],
stride_[i],
paddingZ_[i],
paddingY_[i],
padding_[i]);
real *outData = outMat->getData() + n * outMat->getStride();
MatrixPtr outMatSub =
Matrix::create(outData, groups_[i] * M, N, false, useGpu_);
for (int g = 0; g < groups_[i]; g++) {
MatrixPtr wMatSub = wMat->subMatrix(g * M, M);
MatrixPtr in = colBuf_->subMatrix(g * K, K);
MatrixPtr out = outMatSub->subMatrix(g * M, M);
out->mul(*wMatSub, *in, 1.0, 1.0);
}
}
}
if (nullptr != this->biasParameter_) {
REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str());
this->addBias();
}
forwardActivation();
}
void Conv3DLayer::backward(const UpdateCallback &callback) {
backwardActivation();
if (biases_ && biases_->getWGrad()) {
bpropBiases();
biases_->getParameterPtr()->incUpdate(callback);
}
for (size_t i = 0; i != inputLayers_.size(); ++i) {
REGISTER_TIMER_INFO("BwdConv3D", getName().c_str());
if (weights_[i]->getWGrad()) {
bpropWeights(i);
}
if (getInputGrad(i)) {
bpropData(i);
}
REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
weights_[i]->getParameterPtr()->incUpdate(callback);
}
}
void Conv3DLayer::bpropWeights(int i) {
int M = M_[i];
int N = N_[i];
int K = K_[i];
const MatrixPtr &inMat = getInputValue(i);
Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
MatrixPtr wGradMat = weights_[i]->getWGrad();
int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
for (int n = 0; n < batchSize; ++n) {
colBuf_->vol2Col(inMat->getData() + n * inMat->getStride(),
channels_[i],
imgSizeD_[i],
imgSizeH_[i],
imgSizeW_[i],
filterSizeZ_[i],
filterSizeY_[i],
filterSize_[i],
strideZ_[i],
strideY_[i],
stride_[i],
paddingZ_[i],
paddingY_[i],
padding_[i]);
real *outGradData =
getOutputGrad()->getData() + n * getOutputGrad()->getStride();
MatrixPtr outGradSub =
Matrix::create(outGradData, groups_[i] * M, N, false, useGpu_);
for (int g = 0; g < groups_[i]; ++g) {
MatrixPtr inMatSub = colBuf_->subMatrix(g * K, K);
MatrixPtr outG = outGradSub->subMatrix(g * M, M);
MatrixPtr wGradSub = wGradMat->subMatrix(g * M, M);
wGradSub->mul(*outG, *(inMatSub->getTranspose()), 1.0, 1.0);
}
}
}
void Conv3DLayer::bpropData(int i) {
int M = M_[i];
int N = N_[i];
int K = K_[i];
Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
MatrixPtr wMat = weights_[i]->getW();
int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
for (int n = 0; n < batchSize; ++n) {
real *outGradData =
getOutputGrad()->getData() + n * getOutputGrad()->getStride();
real *preGradData =
getInputGrad(i)->getData() + n * getInputGrad(i)->getStride();
MatrixPtr outGradSub =
Matrix::create(outGradData, M * groups_[i], N, false, useGpu_);
for (int g = 0; g < groups_[i]; ++g) {
MatrixPtr wMatSub = wMat->subMatrix(g * M, M);
MatrixPtr outG = outGradSub->subMatrix(g * M, M);
MatrixPtr inGradMatSub = colBuf_->subMatrix(g * K, K);
inGradMatSub->mul(*(wMatSub->getTranspose()), *outG, 1.0, 0.0);
}
colBuf_->col2Vol(preGradData,
channels_[i],
imgSizeD_[i],
imgSizeH_[i],
imgSizeW_[i],
filterSizeZ_[i],
filterSizeY_[i],
filterSize_[i],
strideZ_[i],
strideY_[i],
stride_[i],
paddingZ_[i],
paddingY_[i],
padding_[i],
1.0,
1.0);
}
}
void Conv3DLayer::bpropBiases() {
MatrixPtr outGradMat = getOutputGrad();
if (this->sharedBiases_) {
biases_->getWGrad()->collectSharedBias(*outGradMat, 1.0f);
} else {
biases_->getWGrad()->collectBias(*outGradMat, 1.0f);
}
}
void Conv3DLayer::addBias() {
MatrixPtr outMat = getOutputValue();
if (this->sharedBiases_) {
outMat->addSharedBias(*(biases_->getW()), 1.0f);
} else {
outMat->addBias(*(biases_->getW()), 1.0f);
}
}
} // namespace paddle

51
paddle/gserver/layers/Conv3DLayer.h
Unescape View File

@ -0,0 +1,51 @@
/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <vector>
#include "ConvBaseLayer.h"
#include "paddle/math/MathUtils.h"
#include "paddle/math/Matrix.h"
namespace paddle {
/**
* @brief A subclass of convolution layer.
* This layer expands input and use matrix multiplication to
* calculate convolution operation.
*/
class Conv3DLayer : public ConvBaseLayer {
public:
explicit Conv3DLayer(const LayerConfig& config) : ConvBaseLayer(config) {}
~Conv3DLayer() {}
bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
void forward(PassType passType);
void addBias();
void backward(const UpdateCallback& callback);
void bpropBiases();
void bpropData(int i);
void bpropWeights(int i);
size_t getSize();
protected:
// Figure out the dimensions for individual gemms.
IntV M_; /// numFilters_ / filter_group_;
IntV N_; /// channels_ * filterSizeZ_ * filterSize_ * filterSizeY_
IntV K_; /// outputD_ * outputH_ * outputW_
MatrixPtr colBuf_;
};
} // namespace paddle

32
paddle/gserver/layers/ConvBaseLayer.cpp
Unescape View File

@ -38,7 +38,6 @@ bool ConvBaseLayer::init(const LayerMap& layerMap,
strideY_.push_back(conf.stride_y());
dilationY_.push_back(conf.dilation_y());
filterSizeY_.push_back(conf.filter_size_y());
filterPixels_.push_back(filterSize_.back() * filterSizeY_.back());
channels_.push_back(conf.channels());
imgSizeH_.push_back(conf.has_img_size_y() ? conf.img_size_y()
: conf.img_size());
@ -47,31 +46,20 @@ bool ConvBaseLayer::init(const LayerMap& layerMap,
filterChannels_.push_back(conf.filter_channels());
outputH_.push_back(conf.has_output_y() ? conf.output_y() : conf.output_x());
outputW_.push_back(conf.output_x());
paddingZ_.push_back(conf.padding_z());
strideZ_.push_back(conf.stride_z());
filterSizeZ_.push_back(conf.filter_size_z());
imgSizeD_.push_back(conf.img_size_z());
outputD_.push_back(conf.output_z());
filterPixels_.push_back(filterSize_.back() * filterSizeY_.back() *
filterSizeZ_.back());
}
CHECK(inputLayers_.size() == parameters_.size());
for (size_t i = 0; i < inputLayers_.size(); i++) {
size_t height, width;
height = filterPixels_[i] * filterChannels_[i];
width = (!isDeconv_) ? numFilters_ : channels_[i];
// create a new weight
CHECK_EQ(parameters_[i]->getSize(), width * height);
Weight* w = new Weight(height, width, parameters_[i]);
weights_.emplace_back(w);
}
/* initialize the biases_ */
if (biasParameter_.get()) {
if (sharedBiases_) {
CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
biases_ =
std::unique_ptr<Weight>(new Weight(numFilters_, 1, biasParameter_));
} else {
biases_ =
std::unique_ptr<Weight>(new Weight(getSize(), 1, biasParameter_));
}
}
// create new weights_ in derived class
// create new biases_ in derived class
// default caffe model
caffeMode_ = true;

7
paddle/gserver/layers/ConvBaseLayer.h
Unescape View File

@ -62,6 +62,13 @@ protected:
IntV outputH_;
/// The spatial dimensions of output feature map width.
IntV outputW_;
IntV outputD_;
IntV imgSizeD_;
IntV filterSizeZ_;
IntV strideZ_;
IntV paddingZ_;
/// Group size, refer to grouped convolution in
/// Alex Krizhevsky's paper: when group=2, the first half of the
/// filters are only connected to the first half of the input channels,

18
paddle/gserver/layers/CudnnConvBaseLayer.cpp
Unescape View File

@ -46,8 +46,26 @@ bool CudnnConvBaseLayer::init(const LayerMap &layerMap,
projConf_.emplace_back(conf);
projections_.emplace_back(
Projection::create(*projConf_[i], parameters_[i], useGpu_));
// create a new weight
size_t height, width;
height = filterPixels_[i] * filterChannels_[i];
width = (!isDeconv_) ? numFilters_ : channels_[i];
CHECK_EQ(parameters_[i]->getSize(), width * height);
Weight *w = new Weight(height, width, parameters_[i]);
weights_.emplace_back(w);
}
if (biasParameter_.get()) {
if (sharedBiases_) {
CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
biases_ =
std::unique_ptr<Weight>(new Weight(numFilters_, 1, biasParameter_));
} else {
biases_ =
std::unique_ptr<Weight>(new Weight(getSize(), 1, biasParameter_));
}
}
if (biases_.get() && sharedBiases_) {
hl_create_tensor_descriptor(&biasDesc_);
hl_create_tensor_descriptor(&outputDesc_);

212
paddle/gserver/layers/DeConv3DLayer.cpp
Unescape View File

@ -0,0 +1,212 @@
/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "DeConv3DLayer.h"
#include "paddle/utils/Logging.h"
#include "paddle/utils/Stat.h"
namespace paddle {
REGISTER_LAYER(deconv3d, DeConv3DLayer);
bool DeConv3DLayer::init(const LayerMap &layerMap,
const ParameterMap &parameterMap) {
if (!ConvBaseLayer::init(layerMap, parameterMap)) return false;
// for Deconv, the dimension of Kernel is
// channel * output * depth * height * weigth
// Matrix storage format: (output * depth * height * weigth) x channel
for (int index = 0; index < config_.inputs().size(); ++index) {
M_.push_back(filterChannels_[index]);
K_.push_back(filterPixels_[index] * (numFilters_ / groups_[index]));
// create a new weight
size_t height, width;
height = filterPixels_[index] * numFilters_;
width = filterChannels_[index];
CHECK_EQ(parameters_[index]->getSize(), width * height);
Weight *w = new Weight(height, width, parameters_[index]);
weights_.emplace_back(w);
}
if (biasParameter_.get()) {
if (sharedBiases_) {
CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
biases_ =
std::unique_ptr<Weight>(new Weight(1, numFilters_, biasParameter_));
} else {
biases_ =
std::unique_ptr<Weight>(new Weight(1, getSize(), biasParameter_));
}
}
return true;
}
size_t DeConv3DLayer::getSize() {
CHECK_NE(inputLayers_.size(), 0UL);
outputH_.clear();
outputW_.clear();
outputD_.clear();
N_.clear();
NOut_.clear();
size_t layerSize = 0;
for (size_t i = 0; i < inputLayers_.size(); ++i) {
outputW_.push_back(
imageSize(imgSizeW_[i], filterSize_[i], padding_[i], stride_[i], true));
outputH_.push_back(imageSize(
imgSizeH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true));
outputD_.push_back(imageSize(
imgSizeD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true));
NOut_.push_back(outputD_[i] * outputH_[i] * outputW_[i]);
N_.push_back(imgSizeD_[i] * imgSizeH_[i] * imgSizeW_[i]);
CHECK(layerSize == 0 || N_[i] * size_t(numFilters_) == layerSize);
layerSize += NOut_[i] * numFilters_;
}
getOutput().setFrameHeight(outputH_[0]);
getOutput().setFrameWidth(outputW_[0]);
getOutput().setFrameDepth(outputD_[0]);
return layerSize;
}
void DeConv3DLayer::forward(PassType passType) {
Layer::forward(passType);
int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
int outWidth = getSize();
resetOutput(batchSize, outWidth);
const MatrixPtr outMat = getOutputValue();
for (size_t i = 0; i != inputLayers_.size(); ++i) {
REGISTER_TIMER_INFO("FwdDeConv3D", getName().c_str());
const MatrixPtr &inMat = getInputValue(i);
int M = M_[i];
int N = N_[i];
int K = K_[i];
MatrixPtr wMat = weights_[i]->getW();
Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
for (int n = 0; n < batchSize; ++n) {
real *inData = inMat->getData() + n * inMat->getStride();
for (int g = 0; g < groups_[i]; ++g) {
MatrixPtr inMatSub = Matrix::create(inData, M, N, false, useGpu_);
MatrixPtr wMatSub = wMat->subMatrix(g * K, K);
MatrixPtr colBufDataSub = colBuf_->subMatrix(g * K, K);
colBufDataSub->mul(*wMatSub, *inMatSub, 1.0, 0.0);
inData += M * N;
}
colBuf_->col2Vol(outMat->getData() + n * outMat->getStride(),
numFilters_,
outputD_[i],
outputH_[i],
outputW_[i],
filterSizeZ_[i],
filterSizeY_[i],
filterSize_[i],
strideZ_[i],
strideY_[i],
stride_[i],
paddingZ_[i],
paddingY_[i],
padding_[i],
1.0,
1.0);
}
}
if (nullptr != this->biasParameter_) {
REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str());
this->addBias();
}
forwardActivation();
}
void DeConv3DLayer::backward(const UpdateCallback &callback) {
backwardActivation();
int batchSize = getOutputGrad()->getHeight();
if (biases_ && biases_->getWGrad()) {
bpropBiases();
biases_->getParameterPtr()->incUpdate(callback);
}
for (size_t i = 0; i < inputLayers_.size(); ++i) {
if (weights_[i]->getWGrad() || this->needGradient_) {
int M = M_[i];
int N = N_[i];
int K = K_[i];
REGISTER_TIMER_INFO("BwdDeConv3D", getName().c_str());
Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
const MatrixPtr &inMat = getInputValue(i);
for (int n = 0; n < batchSize; ++n) {
colBuf_->vol2Col(
getOutputGrad()->getData() + n * getOutputGrad()->getStride(),
numFilters_,
outputD_[i],
outputH_[i],
outputW_[i],
filterSizeZ_[i],
filterSizeY_[i],
filterSize_[i],
strideZ_[i],
strideY_[i],
stride_[i],
paddingZ_[i],
paddingY_[i],
padding_[i]);
if (weights_[i]->getWGrad()) {
real *inData = inMat->getData() + n * inMat->getStride();
for (int g = 0; g < groups_[i]; ++g) {
MatrixPtr colBufDataSub = colBuf_->subMatrix(g * K, K);
MatrixPtr wGradMatSub =
weights_[i]->getWGrad()->subMatrix(g * K, K);
MatrixPtr inMatSub = Matrix::create(inData, M, N, false, useGpu_);
wGradMatSub->mul(
*colBufDataSub, *(inMatSub->getTranspose()), 1.0, 1.0);
inData += M * N;
}
}
if (getInputGrad(i)) {
real *preGrad =
getInputGrad(i)->getData() + n * getInputGrad(i)->getStride();
for (int g = 0; g < groups_[i]; ++g) {
MatrixPtr w = weights_[i]->getW()->subMatrix(g * K, K);
MatrixPtr outGradMat = colBuf_->subMatrix(g * K, K);
MatrixPtr inGradMatSub =
Matrix::create(preGrad, M, N, false, useGpu_);
inGradMatSub->mul(*(w->getTranspose()), *outGradMat, 1.0, 1.0);
preGrad += M * N;
}
}
}
REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
weights_[i]->getParameterPtr()->incUpdate(callback);
}
}
}
void DeConv3DLayer::bpropWeights(int i) {}
void DeConv3DLayer::bpropData(int i) {}
void DeConv3DLayer::bpropBiases() {
const MatrixPtr &outGradMat = getOutputGrad();
if (this->sharedBiases_) {
biases_->getWGrad()->collectSharedBias(*outGradMat, 1.0f);
} else {
biases_->getWGrad()->collectBias(*outGradMat, 1.0f);
}
}
void DeConv3DLayer::addBias() {
MatrixPtr outMat = getOutputValue();
if (this->sharedBiases_) {
outMat->addSharedBias(*(biases_->getW()), 1.0f);
} else {
outMat->addBias(*(biases_->getW()), 1.0f);
}
}
} // namespace paddle

52
paddle/gserver/layers/DeConv3DLayer.h
Unescape View File

@ -0,0 +1,52 @@
/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <vector>
#include "ConvBaseLayer.h"
#include "paddle/math/MathUtils.h"
#include "paddle/math/Matrix.h"
namespace paddle {
/**
* @brief A subclass of deconvolution3D layer.
* This layer expands input and use matrix multiplication to
* calculate deconvolution3D operation.
*/
class DeConv3DLayer : public ConvBaseLayer {
public:
explicit DeConv3DLayer(const LayerConfig& config) : ConvBaseLayer(config) {}
~DeConv3DLayer() {}
bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
void forward(PassType passType);
void addBias();
void backward(const UpdateCallback& callback);
void bpropBiases();
void bpropData(int i);
void bpropWeights(int i);
size_t getSize();
protected:
// Figure out the dimensions for individual gemms.
IntV M_; /// numFilters_ / filter_group_;
IntV N_; /// channels_ * filterSizeZ_ * filterSize_ * filterSizeY_
IntV K_; /// outputD_ * outputH_ * outputW_
IntV NOut_;
MatrixPtr colBuf_;
};
} // namespace paddle

21
paddle/gserver/layers/ExpandConvBaseLayer.cpp
Unescape View File

@ -22,12 +22,31 @@ bool ExpandConvBaseLayer::init(const LayerMap &layerMap,
/* Initialize the basic convolutional parent class */
ConvBaseLayer::init(layerMap, parameterMap);
int index = 0;
for (auto &inputConfig : config_.inputs()) {
const ConvConfig &conf = inputConfig.conv_conf();
/* Consistent caffe mode for multiple input */
caffeMode_ = conf.caffe_mode();
}
// create a new weight
size_t height, width;
height = filterPixels_[index] * filterChannels_[index];
width = (!isDeconv_) ? numFilters_ : channels_[index];
CHECK_EQ(parameters_[index]->getSize(), width * height);
Weight *w = new Weight(height, width, parameters_[index]);
weights_.emplace_back(w);
index++;
}
if (biasParameter_.get()) {
if (sharedBiases_) {
CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
biases_ =
std::unique_ptr<Weight>(new Weight(numFilters_, 1, biasParameter_));
} else {
biases_ =
std::unique_ptr<Weight>(new Weight(getSize(), 1, biasParameter_));
}
}
getOutputSize();
return true;

178
paddle/gserver/layers/Pool3DLayer.cpp
Unescape View File

@ -0,0 +1,178 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "Pool3DLayer.h"
#include "PoolProjectionLayer.h"
#include "paddle/utils/Logging.h"
namespace paddle {
REGISTER_LAYER(pool3d, Pool3DLayer);
bool Pool3DLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
Layer::init(layerMap, parameterMap);
/* the size of inputs for pool-layer is 1 */
CHECK_EQ(config_.inputs_size(), 1);
const PoolConfig& conf = config_.inputs(0).pool_conf();
poolType_ = conf.pool_type();
channels_ = conf.channels();
sizeX_ = conf.size_x();
sizeY_ = conf.size_y();
sizeZ_ = conf.size_z();
strideW_ = conf.stride();
strideH_ = conf.stride_y();
strideD_ = conf.stride_z();
imgSizeW_ = conf.img_size();
imgSizeH_ = conf.img_size_y();
imgSizeD_ = conf.img_size_z();
paddingW_ = conf.padding();
paddingH_ = conf.padding_y();
paddingD_ = conf.padding_z();
outputW_ = conf.output_x();
outputH_ = conf.output_y();
outputD_ = conf.output_z();
return true;
}
size_t Pool3DLayer::getSize() {
CHECK_EQ(inputLayers_.size(), 1UL);
size_t layerSize = 0;
outputD_ = outputSize(imgSizeD_, sizeZ_, paddingD_, strideD_, false);
outputH_ = outputSize(imgSizeH_, sizeY_, paddingH_, strideH_, false);
outputW_ = outputSize(imgSizeW_, sizeX_, paddingW_, strideW_, false);
layerSize = outputD_ * outputH_ * outputW_ * channels_;
getOutput().setFrameHeight(outputH_);
getOutput().setFrameWidth(outputW_);
getOutput().setFrameDepth(outputD_);
return layerSize;
}
void Pool3DLayer::forward(PassType passType) {
Layer::forward(passType);
const MatrixPtr& inMat = inputLayers_[0]->getOutputValue();
size_t batchSize = inMat->getHeight();
size_t outWidth = getSize();
resetOutput(batchSize, outWidth);
Matrix::resizeOrCreate(maxPoolIdx_, batchSize, outWidth, false, useGpu_);
const MatrixPtr outMat = getOutputValue();
if (poolType_ == "avg") {
outMat->avgPool3DForward(*inMat,
channels_,
imgSizeD_,
imgSizeH_,
imgSizeW_,
outputD_,
outputH_,
outputW_,
sizeZ_,
sizeY_,
sizeX_,
strideD_,
strideH_,
strideW_,
paddingD_,
paddingH_,
paddingW_);
} else if (poolType_ == "max") {
outMat->maxPool3DForward(*inMat,
*maxPoolIdx_,
channels_,
imgSizeD_,
imgSizeH_,
imgSizeW_,
outputD_,
outputH_,
outputW_,
sizeZ_,
sizeY_,
sizeX_,
strideD_,
strideH_,
strideW_,
paddingD_,
paddingH_,
paddingW_);
} else {
LOG(FATAL) << "Unknown pool type: " << poolType_;
}
forwardActivation();
}
void Pool3DLayer::backward(const UpdateCallback& callback) {
backwardActivation();
(void)callback;
if (NULL == getInputGrad(0)) return;
MatrixPtr inMat = inputLayers_[0]->getOutputValue();
MatrixPtr inGradMat = inputLayers_[0]->getOutputGrad();
MatrixPtr outMat = getOutputValue();
MatrixPtr outGradMat = getOutputGrad();
if (poolType_ == "avg") {
inGradMat->avgPool3DBackward(*outGradMat,
imgSizeD_,
imgSizeH_,
imgSizeW_,
outputD_,
outputH_,
outputW_,
sizeZ_,
sizeY_,
sizeZ_,
strideD_,
strideH_,
strideW_,
paddingD_,
paddingH_,
paddingW_,
1.0,
1.0);
} else if (poolType_ == "max") {
inGradMat->maxPool3DBackward(*outGradMat,
*maxPoolIdx_,
imgSizeD_,
imgSizeH_,
imgSizeW_,
outputD_,
outputH_,
outputW_,
sizeZ_,
sizeY_,
sizeZ_,
strideD_,
strideH_,
strideW_,
paddingD_,
paddingH_,
paddingW_,
1.0,
1.0);
} else {
LOG(FATAL) << "Unknown pool type: " << poolType_;
}
}
} // namespace paddle

49
paddle/gserver/layers/Pool3DLayer.h
Unescape View File

@ -0,0 +1,49 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <vector>
#include "Layer.h"
#include "paddle/math/MathUtils.h"
#include "paddle/math/Matrix.h"
namespace paddle {
/**
* @brief Basic parent layer of pooling
* Pools the input within regions
*/
class Pool3DLayer : public Layer {
public:
explicit Pool3DLayer(const LayerConfig& config) : Layer(config) {}
~Pool3DLayer() {}
bool init(const LayerMap& layerMap,
const ParameterMap& parameterMap) override;
void forward(PassType passType) override;
void backward(const UpdateCallback& callback) override;
size_t getSize();
protected:
int channels_;
int sizeX_, sizeY_, sizeZ_;
int strideW_, strideH_, strideD_;
int paddingW_, paddingH_, paddingD_;
int imgSizeW_, imgSizeH_, imgSizeD_;
int outputW_, outputH_, outputD_;
std::string poolType_;
MatrixPtr maxPoolIdx_;
};
} // namespace paddle

222
paddle/gserver/tests/test_LayerGrad.cpp
Unescape View File

@ -1246,6 +1246,75 @@ TEST(Layer, PoolLayer) {
#endif
}
void setPool3DConfig(TestConfig* config,
PoolConfig* pool,
const string& poolType) {
// filter size
const int NUM_FILTERS = 16;
const int FILTER_SIZE = 3;
const int FILTER_SIZE_Y = 3;
const int FILTER_SIZE_Z = 3;
const int CHANNELS = 16;
(*config).biasSize = 0;
(*config).layerConfig.set_type("pool3d");
(*config).layerConfig.set_num_filters(NUM_FILTERS);
int kw = FILTER_SIZE, kh = FILTER_SIZE_Y, kd = FILTER_SIZE_Z;
int pw = 0, ph = 0, pd = 0;
int sw = 2, sh = 2, sd = 2;
pool->set_pool_type(poolType);
pool->set_pool_type("avg");
pool->set_channels(CHANNELS);
pool->set_size_x(kw);
pool->set_size_y(kh);
pool->set_size_z(kd);
pool->set_padding(0);
pool->set_padding_y(0);
pool->set_padding_z(0);
pool->set_stride(sw);
pool->set_stride_y(sh);
pool->set_stride_z(sd);
pool->set_start(0);
int ow = outputSize(pool->img_size(), kw, pw, sw, /* caffeMode */ false);
int oh = outputSize(pool->img_size_y(), kh, ph, sh, /* caffeMode */ false);
int od = outputSize(pool->img_size_z(), kd, pd, sd, /* caffeMode */ false);
pool->set_output_x(ow);
pool->set_output_y(oh);
pool->set_output_z(od);
}
void testPool3DLayer(const string& poolType, bool trans, bool useGpu) {
TestConfig config;
config.inputDefs.push_back({INPUT_DATA, "layer_0", 11664, 0});
LayerInputConfig* input = config.layerConfig.add_inputs();
PoolConfig* pool = input->mutable_pool_conf();
const int IMAGE_SIZE = 9;
const int IMAGE_SIZE_Y = 9;
const int IMAGE_SIZE_Z = 9;
pool->set_img_size(IMAGE_SIZE);
pool->set_img_size_y(IMAGE_SIZE_Y);
pool->set_img_size_z(IMAGE_SIZE_Z);
setPool3DConfig(&config, pool, poolType);
config.layerConfig.set_size(pool->output_x() * pool->output_y() *
pool->channels());
testLayerGrad(config, "pool3d", 100, trans, useGpu);
}
TEST(Layer, Pool3DLayer) {
testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ false);
testPool3DLayer("max", /* trans= */ false, /* useGpu= */ false);
#ifndef PADDLE_ONLY_CPU
testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ true);
testPool3DLayer("max", /* trans= */ false, /* useGpu= */ true);
#endif
}
void testSppLayer(const string& poolType,
const int pyramidHeight,
bool trans,
@ -2096,6 +2165,159 @@ TEST(Layer, RowL2NormLayer) {
}
}
void test3DConvLayer(const string& type, bool trans, bool useGpu) {
// filter size
const int NUM_FILTERS = 6;
// const int CHANNELS = 3;
const int FILTER_SIZE = 3;
const int FILTER_SIZE_Y = 3;
const int FILTER_SIZE_Z = 3;
// input image
const int CHANNELS = 3;
const int IMAGE_SIZE = 9;
const int IMAGE_SIZE_Y = 9;
const int IMAGE_SIZE_Z = 9;
TestConfig config;
config.biasSize = NUM_FILTERS;
config.layerConfig.set_type(type);
config.layerConfig.set_num_filters(NUM_FILTERS);
config.layerConfig.set_partial_sum(1);
config.layerConfig.set_shared_biases(true);
// Setting up conv3D-trans layer
LayerInputConfig* input = config.layerConfig.add_inputs();
ConvConfig* conv = input->mutable_conv_conf();
conv->set_channels(CHANNELS);
conv->set_filter_size(FILTER_SIZE);
conv->set_filter_size_y(FILTER_SIZE_Y);
conv->set_filter_size_z(FILTER_SIZE_Z);
conv->set_padding(0);
conv->set_padding_y(0);
conv->set_padding_z(0);
conv->set_stride(2);
conv->set_stride_y(2);
conv->set_stride_z(2);
conv->set_img_size(IMAGE_SIZE);
conv->set_img_size_y(IMAGE_SIZE_Y);
conv->set_img_size_z(IMAGE_SIZE_Z);
conv->set_output_x(outputSize(conv->img_size(),
conv->filter_size(),
conv->padding(),
conv->stride(),
/* caffeMode */ true));
conv->set_output_y(outputSize(conv->img_size_y(),
conv->filter_size_y(),
conv->padding_y(),
conv->stride_y(),
/* caffeMode */ true));
conv->set_output_z(outputSize(conv->img_size_z(),
conv->filter_size_z(),
conv->padding_z(),
conv->stride_z(),
/* caffeMode */ true));
config.layerConfig.set_size(conv->output_x() * conv->output_y() *
conv->output_z() * NUM_FILTERS);
conv->set_groups(1);
conv->set_filter_channels(conv->channels() / conv->groups());
config.inputDefs.push_back(
{INPUT_DATA,
"layer_0",
CHANNELS * IMAGE_SIZE * IMAGE_SIZE_Y * IMAGE_SIZE_Z,
conv->filter_channels() * FILTER_SIZE * FILTER_SIZE_Y * FILTER_SIZE_Z *
NUM_FILTERS});
testLayerGrad(config, "conv3D", 10, trans, useGpu);
// Use small batch_size and useWeight=true to test biasGrad
testLayerGrad(config, "conv3D", 2, trans, useGpu, true, 0.02);
}
TEST(Layer, test3DConvLayer) {
test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ false);
#ifndef PADDLE_ONLY_CPU
test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ true);
#endif
}
void test3DDeConvLayer(const string& type, bool trans, bool useGpu) {
// filter size
const int NUM_FILTERS = 6;
// const int CHANNELS = 3;
const int FILTER_SIZE = 3;
const int FILTER_SIZE_Y = 3;
const int FILTER_SIZE_Z = 3;
// input image
const int CHANNELS = 3;
const int IMAGE_SIZE = 4;
const int IMAGE_SIZE_Y = 6;
const int IMAGE_SIZE_Z = 6;
// Setting up conv-trans layer
TestConfig config;
config.biasSize = NUM_FILTERS;
config.layerConfig.set_type("deconv3d");
config.layerConfig.set_num_filters(NUM_FILTERS);
config.layerConfig.set_partial_sum(1);
config.layerConfig.set_shared_biases(true);
LayerInputConfig* input = config.layerConfig.add_inputs();
ConvConfig* conv = input->mutable_conv_conf();
conv->set_channels(CHANNELS);
conv->set_filter_size(FILTER_SIZE);
conv->set_filter_size_y(FILTER_SIZE_Y);
conv->set_filter_size_z(FILTER_SIZE_Z);
conv->set_padding(0);
conv->set_padding_y(0);
conv->set_padding_z(0);
conv->set_stride(2);
conv->set_stride_y(2);
conv->set_stride_z(2);
conv->set_img_size(IMAGE_SIZE);
conv->set_img_size_y(IMAGE_SIZE_Y);
conv->set_img_size_z(IMAGE_SIZE_Z);
conv->set_output_x(imageSize(conv->img_size(),
conv->filter_size(),
conv->padding(),
conv->stride(),
true));
conv->set_output_y(imageSize(conv->img_size_y(),
conv->filter_size_y(),
conv->padding_y(),
conv->stride_y(),
true));
conv->set_output_z(imageSize(conv->img_size_z(),
conv->filter_size_z(),
conv->padding_z(),
conv->stride_z(),
true));
config.layerConfig.set_size(conv->output_x() * conv->output_y() *
conv->output_z() * NUM_FILTERS);
conv->set_groups(1);
conv->set_filter_channels(conv->channels() / conv->groups());
config.inputDefs.push_back(
{INPUT_DATA,
"layer_0",
CHANNELS * IMAGE_SIZE * IMAGE_SIZE_Y * IMAGE_SIZE_Z,
conv->filter_channels() * FILTER_SIZE * FILTER_SIZE_Y * FILTER_SIZE_Z *
NUM_FILTERS});
testLayerGrad(config, "deconv3D", 10, trans, useGpu);
// Use small batch_size and useWeight=true to test biasGrad
testLayerGrad(config, "deconv3D", 2, trans, useGpu, true, 0.02);
}
TEST(Layer, test3DDeConvLayer) {
test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ false);
#ifndef PADDLE_ONLY_CPU
test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ true);
#endif
}
TEST(Layer, ScaleShiftLayer) {
const size_t batchSize = 16;
const size_t size = 32;

640
paddle/math/Matrix.cpp
View File

File diff suppressed because it is too large Load Diff

353
paddle/math/Matrix.h
View File

File diff suppressed because it is too large Load Diff

494
paddle/math/tests/test_matrixCompare.cpp
View File

File diff suppressed because it is too large Load Diff

1
proto/ModelConfig.proto
Unescape View File

@ -515,6 +515,7 @@ message LayerConfig {
// for HuberRegressionLoss
optional double delta = 57 [ default = 1.0 ];
// for 3D data
optional double depth = 58 [ default = 1 ];
}

250
python/paddle/trainer/config_parser.py
View File

File diff suppressed because it is too large Load Diff

0
python/paddle/trainer/recurrent_units.py
Unescape View File

290
python/paddle/trainer_config_helpers/layers.py
View File

File diff suppressed because it is too large Load Diff

0
python/paddle/trainer_config_helpers/networks.py
Unescape View File

Some files were not shown because too many files have changed in this diff Show More

Write Preview
Loading…
Cancel
Save