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

Merge pull request #4129 from hedaoyuan/convolution

Browse Source 
Refine ExpandConvLayer
update-doc-pybind
hedaoyuan 8 years ago committed by GitHub
parent 43ec735bed b1bca06667
commit 7bcb1fc3bd
4 changed files with 43 additions and 192 deletions
Show Stats Download Patch File Download Diff File

124
paddle/gserver/layers/ExpandConvBaseLayer.cpp
Unescape View File

@ -1,124 +0,0 @@
/* 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 "ExpandConvBaseLayer.h"
#include "paddle/utils/Logging.h"
namespace paddle {
bool ExpandConvBaseLayer::init(const LayerMap &layerMap,
const ParameterMap &parameterMap) {
/* 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;
}
size_t ExpandConvBaseLayer::getOutputSize() {
CHECK_NE(inputLayers_.size(), 0UL);
size_t layerSize = ConvBaseLayer::calOutputSize();
return layerSize;
}
void ExpandConvBaseLayer::addSharedBias() {
size_t mapW = getOutputSize() / numFilters_;
size_t mapH = getOutputValue()->getElementCnt() / mapW;
MatrixPtr out =
Matrix::create(getOutputValue()->getData(), mapH, mapW, false, useGpu_);
Matrix::resizeOrCreate(transOutValue_, mapW, mapH, false, useGpu_);
out->transpose(transOutValue_, false); // false means no memory allocation
transOutValue_->reshape(transOutValue_->getElementCnt() / numFilters_,
numFilters_);
MatrixPtr bias = Matrix::create(biases_->getW()->getData(),
1,
biases_->getW()->getElementCnt(),
false,
useGpu_);
transOutValue_->addBias(*bias, 1.0f);
transOutValue_->reshape(mapW, mapH);
transOutValue_->transpose(out, false); // false means no memory allocation
out->clear();
bias->clear();
}
void ExpandConvBaseLayer::addUnsharedBias() {
MatrixPtr outValue = getOutputValue();
MatrixPtr bias = Matrix::create(biases_->getW()->getData(),
1,
biases_->getW()->getElementCnt(),
false,
useGpu_);
outValue->addBias(*bias, 1.0f);
}
void ExpandConvBaseLayer::bpropSharedBias(MatrixPtr biases, MatrixPtr v) {
size_t mapW = getOutputSize() / numFilters_;
size_t mapH = v->getElementCnt() / mapW;
MatrixPtr vTmp = Matrix::create(v->getData(), mapH, mapW, false, useGpu_);
Matrix::resizeOrCreate(transOutValue_, mapW, mapH, false, useGpu_);
vTmp->transpose(transOutValue_, false); // false means no memory allocation
transOutValue_->reshape(transOutValue_->getElementCnt() / numFilters_,
numFilters_);
biases->collectBias(*transOutValue_, 1.0f);
}
void ExpandConvBaseLayer::bpropBiases(MatrixPtr v) {
MatrixPtr biases = Matrix::create(biases_->getWGrad()->getData(),
1,
biases_->getWGrad()->getElementCnt(),
false,
useGpu_);
if (sharedBiases_) {
bpropSharedBias(biases, v);
} else {
biases->collectBias(*v, 1.0f);
}
biases->clear();
}
} // namespace paddle

57
paddle/gserver/layers/ExpandConvBaseLayer.h
Unescape View File

@ -1,57 +0,0 @@
/* 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 "ConvBaseLayer.h"
#include "paddle/math/Matrix.h"
namespace paddle {
/**
* @brief A subclass of ConvBaseLayer that is a superclass of both
* ExpandConvLayer and ExpandConvTransLayer
*/
class ExpandConvBaseLayer : public ConvBaseLayer {
protected:
/// The transpose of output, which is an auxiliary matrix.
MatrixPtr transOutValue_;
public:
explicit ExpandConvBaseLayer(const LayerConfig& config)
: ConvBaseLayer(config) {}
~ExpandConvBaseLayer() {}
bool init(const LayerMap& layerMap,
const ParameterMap& parameterMap) override;
size_t getOutputSize();
/**
* Add shared bias.
*/
void addSharedBias();
/**
* Add unshared bias.
*/
void addUnsharedBias();
void bpropSharedBias(MatrixPtr biases, MatrixPtr v);
void bpropBiases(MatrixPtr v);
};
} // namespace paddle

45
paddle/gserver/layers/ExpandConvLayer.cpp
Unescape View File

@ -36,7 +36,36 @@ inline bool isDepthwiseConv(int channels, int groups) {
bool ExpandConvLayer::init(const LayerMap &layerMap, bool ExpandConvLayer::init(const LayerMap &layerMap,
const ParameterMap &parameterMap) { const ParameterMap &parameterMap) {
/* Initialize the basic convolutional parent class */ /* Initialize the basic convolutional parent class */
ExpandConvBaseLayer::init(layerMap, parameterMap); 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(1, numFilters_, biasParameter_, 0));
} else {
biases_ =
std::unique_ptr<Weight>(new Weight(1, getSize(), biasParameter_, 0));
}
}
getOutputSize();
size_t numInputs = config_.inputs_size(); size_t numInputs = config_.inputs_size();
inputShape_.resize(numInputs); inputShape_.resize(numInputs);
@ -108,6 +137,12 @@ bool ExpandConvLayer::init(const LayerMap &layerMap,
return true; return true;
} }
size_t ExpandConvLayer::getOutputSize() {
CHECK_NE(inputLayers_.size(), 0UL);
size_t layerSize = ConvBaseLayer::calOutputSize();
return layerSize;
}
// i is the index of input layers // i is the index of input layers
#define BACKWARD_INPUT(i, inputs, outputs) \ #define BACKWARD_INPUT(i, inputs, outputs) \
backward_[2 * i]->calc(inputs, outputs) backward_[2 * i]->calc(inputs, outputs)
@ -155,11 +190,7 @@ void ExpandConvLayer::forward(PassType passType) {
/* add the bias-vector */ /* add the bias-vector */
if (biases_.get()) { if (biases_.get()) {
if (sharedBiases_) { output_.value->addBias(*biases_->getW(), 1.0, sharedBiases_);
addSharedBias();
} else {
addUnsharedBias();
}
} }
/* activation */ /* activation */
@ -171,7 +202,7 @@ void ExpandConvLayer::backward(const UpdateCallback &callback) {
MatrixPtr outGrad = getOutputGrad(); MatrixPtr outGrad = getOutputGrad();
if (biases_ && biases_->getWGrad()) { if (biases_ && biases_->getWGrad()) {
bpropBiases(outGrad); biases_->getWGrad()->collectBias(*getOutputGrad(), 1, sharedBiases_);
/* Increasing the number of gradient */ /* Increasing the number of gradient */
biases_->getParameterPtr()->incUpdate(callback); biases_->getParameterPtr()->incUpdate(callback);
} }

9
paddle/gserver/layers/ExpandConvLayer.h
Unescape View File

@ -15,7 +15,7 @@ limitations under the License. */
#pragma once #pragma once
#include <vector> #include <vector>
#include "ExpandConvBaseLayer.h" #include "ConvBaseLayer.h"
#include "paddle/math/Matrix.h" #include "paddle/math/Matrix.h"
namespace paddle { namespace paddle {
@ -28,10 +28,9 @@ namespace paddle {
* The config file api is img_conv_layer. * The config file api is img_conv_layer.
*/ */
class ExpandConvLayer : public ExpandConvBaseLayer { class ExpandConvLayer : public ConvBaseLayer {
public: public:
explicit ExpandConvLayer(const LayerConfig& config) explicit ExpandConvLayer(const LayerConfig& config) : ConvBaseLayer(config) {}
: ExpandConvBaseLayer(config) {}
~ExpandConvLayer() {} ~ExpandConvLayer() {}
@ -41,6 +40,8 @@ public:
void forward(PassType passType) override; void forward(PassType passType) override;
void backward(const UpdateCallback& callback) override; void backward(const UpdateCallback& callback) override;
size_t getOutputSize();
protected: protected:
std::vector<TensorShape> inputShape_; std::vector<TensorShape> inputShape_;
std::vector<TensorShape> filterShape_; std::vector<TensorShape> filterShape_;

Write Preview
Loading…
Cancel
Save