Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into lstm_bp
dangqingqing
8 years ago
commit
2e0298744e
@ -1,39 +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. */
|
||||
|
||||
syntax = "proto2";
|
||||
option optimize_for = LITE_RUNTIME;
|
||||
package paddle.framework;
|
||||
|
||||
import "framework.proto";
|
||||
|
||||
/**
|
||||
* This file contains necessary information for model, checkpoint.
|
||||
* etc.
|
||||
*/
|
||||
|
||||
message LoDInfo { repeated int64 level = 1; }
|
||||
|
||||
/**
|
||||
* Save the LoDTensorDesc information through LoDTensorProto, its data memory
|
||||
* is copyed to c buffer immediately. See model_format.md for details.
|
||||
*/
|
||||
|
||||
message LoDTensorProto {
|
||||
optional DataType data_type = 1;
|
||||
repeated int64 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480]
|
||||
repeated LoDInfo levels = 3;
|
||||
optional int32 lod_level = 4 [ default = 0 ];
|
||||
optional int32 version = 5;
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,138 @@
|
||||
/* Copyright (c) 2017 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 "MKLDNNLayer.h"
|
||||
#include "mkldnn.hpp"
|
||||
|
||||
namespace paddle {
|
||||
typedef mkldnn::batch_normalization_forward bn_fwd;
|
||||
typedef mkldnn::batch_normalization_backward bn_bwd;
|
||||
|
||||
/**
|
||||
* @brief A subclass of MKLDNNLayer BatchNorm layer.
|
||||
*
|
||||
* The config file api is mkldnn_batch_norm
|
||||
*/
|
||||
class MKLDNNBatchNormLayer : public MKLDNNLayer {
|
||||
protected:
|
||||
// save forward primitive_desc, which can be used backward
|
||||
std::shared_ptr<bn_fwd::primitive_desc> fwdPD_;
|
||||
|
||||
// Epsilon value used in the batch normalization formula.
|
||||
static const real EPS;
|
||||
// weight and bias in paddle
|
||||
std::unique_ptr<Weight> weight_;
|
||||
std::unique_ptr<Weight> biases_;
|
||||
// mkldnn use a large buffer store both scale and shift
|
||||
// which are weight and bias in paddle corresponding.
|
||||
MatrixPtr valueScaleShift_;
|
||||
MatrixPtr gradScaleShift_;
|
||||
// Moving average of mean.
|
||||
std::unique_ptr<Weight> movingMean_;
|
||||
// Moving average of variance.
|
||||
std::unique_ptr<Weight> movingVar_;
|
||||
|
||||
// if useGlobalStats_ is true, will use the loaded mean and variance.
|
||||
// otherwise, calculate mean and variance in every mini-batch.
|
||||
bool useGlobalStats_;
|
||||
// used in MKLDNN primitive desc
|
||||
unsigned flags_;
|
||||
// use to compute moving mean and variance.
|
||||
real movingAvgFraction_;
|
||||
// whether the weight has been init
|
||||
bool hasInitedWgt_;
|
||||
|
||||
// local mean and variance
|
||||
// when useGlobalStats_ they are loaded from moving mean and variance
|
||||
// when do not useGlobalStats_ they are calculated from this mini-batch
|
||||
MKLDNNMatrixPtr mean_;
|
||||
MKLDNNMatrixPtr var_;
|
||||
|
||||
public:
|
||||
explicit MKLDNNBatchNormLayer(const LayerConfig& config)
|
||||
: MKLDNNLayer(config), useGlobalStats_(true), hasInitedWgt_(false) {}
|
||||
|
||||
~MKLDNNBatchNormLayer() {}
|
||||
|
||||
bool init(const LayerMap& layerMap,
|
||||
const ParameterMap& parameterMap) override;
|
||||
|
||||
void forward(PassType passType) override;
|
||||
|
||||
void reshape(
|
||||
int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override;
|
||||
|
||||
void resetFwd(std::vector<mkldnn::primitive>& pipeline,
|
||||
MKLDNNMatrixPtr& in,
|
||||
MKLDNNMatrixPtr& wgt,
|
||||
MKLDNNMatrixPtr& bias,
|
||||
MKLDNNMatrixPtr& out) override;
|
||||
|
||||
void resetBwd(std::vector<mkldnn::primitive>& pipeline,
|
||||
MKLDNNMatrixPtr& in,
|
||||
MKLDNNMatrixPtr& wgt,
|
||||
MKLDNNMatrixPtr& bias,
|
||||
MKLDNNMatrixPtr& out) override;
|
||||
|
||||
void updateWeights(const UpdateCallback& callback) override;
|
||||
|
||||
void convertWeightsFromPaddle() override;
|
||||
|
||||
protected:
|
||||
void initWeight();
|
||||
/**
|
||||
* cal moving mean and variance.
|
||||
* moving = moving * AvgFraction + local * (1 - AvgFraction)
|
||||
*/
|
||||
void calMovingMeanAndVar();
|
||||
/**
|
||||
* Forward functions: reset buffers(input, weight, output),
|
||||
* reset primitive descriptor,
|
||||
* reset pipeline.
|
||||
*/
|
||||
void resetFwdBuffers(MKLDNNMatrixPtr& in,
|
||||
MKLDNNMatrixPtr& wgt,
|
||||
MKLDNNMatrixPtr& out);
|
||||
void resetFwdPD(std::shared_ptr<bn_fwd::primitive_desc>& pd,
|
||||
MKLDNNMatrixPtr in,
|
||||
MKLDNNMatrixPtr wgt,
|
||||
MKLDNNMatrixPtr out);
|
||||
void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
|
||||
std::shared_ptr<bn_fwd::primitive_desc>& pd,
|
||||
MKLDNNMatrixPtr& in,
|
||||
MKLDNNMatrixPtr& wgt,
|
||||
MKLDNNMatrixPtr& out);
|
||||
|
||||
/**
|
||||
* Backward functions: reset buffers(input, weight, output),
|
||||
* reset primitive descriptor,
|
||||
* reset pipeline.
|
||||
*/
|
||||
void resetBwdBuffers(MKLDNNMatrixPtr& in,
|
||||
MKLDNNMatrixPtr& wgt,
|
||||
MKLDNNMatrixPtr& out);
|
||||
void resetBwdPD(std::shared_ptr<bn_bwd::primitive_desc>& pd,
|
||||
MKLDNNMatrixPtr& in,
|
||||
MKLDNNMatrixPtr& wgt,
|
||||
MKLDNNMatrixPtr& out);
|
||||
void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
|
||||
std::shared_ptr<bn_bwd::primitive_desc>& pd,
|
||||
MKLDNNMatrixPtr& in,
|
||||
MKLDNNMatrixPtr& wgt,
|
||||
MKLDNNMatrixPtr& out);
|
||||
};
|
||||
|
||||
} // namespace paddle
|
||||
File diff suppressed because it is too large
Load Diff
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue