Skip to content

P
PaddleDetection

PaddlePaddle / PaddleDetection
1 年多 前同步成功

通知 696
Star 11112
Fork 2696
P
PaddleDetection
提交 ec3111ab 编写于 作者: Y Yu Yang 提交者： GitHub
浏览文件
操作

Merge pull request #1938 from reyoung/feature/try_to_remove_parallel_param 

Try to remove ParallelParameter.
上级 5d7fabd3 4725b1c0
隐藏空白更改
内联 并排
Showing with 5 addition and 531 deletion
paddle/gserver/layers/Layer.h
浏览文件 @ ec3111ab
......@@ -14,20 +14,18 @@ limitations under the License. */
#pragma once
#include <paddle/parameter/Argument.h>
#include <functional>
#include <memory>
#include "ModelConfig.pb.h"
#include "paddle/function/Function.h"
#include "paddle/gserver/activations/ActivationFunction.h"
#include "paddle/math/CpuSparseMatrix.h"
#include "paddle/parameter/Argument.h"
#include "paddle/parameter/Parameter.h"
#include "paddle/parameter/Weight.h"
#include "paddle/utils/ClassRegistrar.h"
#include "paddle/utils/Util.h"
#include <paddle/parameter/ParallelParameter.h>
#include <paddle/parameter/Weight.h>
#include "paddle/gserver/activations/ActivationFunction.h"
/// Macro for registering a layer type.
/// Example: REGISTER_LAYER(crf_error, CRFDecodingErrorLayer);
#define REGISTER_LAYER(__type_name, __class_name) \
......
paddle/gserver/tests/test_RecurrentGradientMachine.cpp
浏览文件 @ ec3111ab
......@@ -14,6 +14,7 @@ limitations under the License. */
#include <gtest/gtest.h>
#include <paddle/gserver/gradientmachines/GradientMachine.h>
#include <paddle/parameter/ParameterUpdateFunctions.h>
#include <paddle/trainer/Trainer.h>
#include <paddle/trainer/TrainerInternal.h>
#include <paddle/utils/PythonUtil.h>
......
paddle/parameter/Argument.cpp
浏览文件 @ ec3111ab
......@@ -570,7 +570,7 @@ void Argument::poolSequenceWithStride(const Argument& input,
CHECK(input.sequenceStartPositions);
CHECK_EQ(input.hasSubseq(), 0UL);
CHECK_GT(stride, 0) << "stride must larger than 0";
CHECK_GT(stride, 0UL) << "stride must larger than 0";
size_t numSequences = input.getNumSequences();
ICpuGpuVector::resizeOrCreate(
sequenceStartPositions, numSequences + 1, false);
......
paddle/parameter/ParallelParameter.cpp 已删除 100644 → 0
浏览文件 @ 5d7fabd3
/* 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 <fstream>
#include "paddle/utils/Logging.h"
#include "ParallelParameter.h"
namespace paddle {
UpdateFunction paramUpdateFunctions[UPDATE_TYPE_NUM] = {
nullptr, // &ParallelParameter::singleUpdate, /* single thread */
nullptr, // &ParallelParameter::controlUpdate, /* controller thread */
&ParallelParameter::majorUpdate, /* major thread */
&ParallelParameter::minorUpdate, /* minor thread */
nullptr, /* master */
&ParallelParameter::slaveUpdate, /* slave */
};
ParallelParameterPtr ParallelParameter::create(TrainerRole role,
ParameterPtr localParam,
int asyncCount) {
ParallelParameterPtr ptr = nullptr;
switch (role) {
case TRAINER_ROLE_CONTROL:
case TRAINER_ROLE_MAJOR:
case TRAINER_ROLE_MINOR:
ptr = std::make_shared<SyncParameter>(role, localParam);
break;
case TRAINER_ROLE_MASTER:
case TRAINER_ROLE_SLAVE:
ptr = std::make_shared<AsyncParameter>(role, asyncCount, localParam);
break;
default:
LOG(FATAL) << "unknown role " << role << "\n";
}
return ptr;
}
void ParallelParameter::syncUpdate(TrainerRole role, real learnRate) {
if (paramUpdateFunctions[role]) {
(this->*paramUpdateFunctions[role])(learnRate);
}
}
void SyncParameter::attachControlParam(ParallelParameterPtr controler) {
controlParam_ = controler;
}
void SyncParameter::attachMajorParam(ParallelParameterPtr partner) {
majorPartners_.push_back(partner);
if (role_ == TRAINER_ROLE_CONTROL) {
localParam_->setSharedCount(majorPartners_.size());
}
// partnerParam_ = partner;
}
void SyncParameter::attachMinorParam(ParallelParameterPtr partner,
int deviceId) {
minorPartners_.push_back(partner);
minorDeviceIds_.push_back(deviceId);
// partnerParam_ = partner;
}
void SyncParameter::waitAllMajorGradReady() {
for (size_t i = 0; i < majorPartners_.size(); i++) {
majorPartners_[i]->waitGradReady();
partnerParam_ = majorPartners_[i]->getLocalParameter();
VectorPtr localGrad = localParam_->getBuf(PARAMETER_GRADIENT);
VectorPtr patnrGrad = partnerParam_->getBuf(PARAMETER_GRADIENT);
if (FLAGS_use_gpu) hl_set_device(minorDeviceIds_[i]);
localGrad->add(*patnrGrad);
}
}
void SyncParameter::synchronizeParamter() {
valueSem_->wait();
if (role_ == TRAINER_ROLE_MINOR) {
/* copy the value from controller */
VectorPtr cntrlVec =
(controlParam_->getLocalParameter())->getBuf(PARAMETER_VALUE);
VectorPtr localVec = localParam_->getBuf(PARAMETER_VALUE);
localVec->copyFrom(*cntrlVec);
/* dispatch the value to major */
for (size_t i = 0; i < majorPartners_.size(); i++) {
VectorPtr majorVec =
(majorPartners_[i]->getLocalParameter())->getBuf(PARAMETER_VALUE);
majorVec->copyFrom(*localVec);
majorPartners_[i]->postValueReady();
}
}
}
void SyncParameter::singleUpdate(real learnRate) {
CHECK(role_ == TRAINER_ROLE_SINGLE);
localParam_->updateWithGradient(learnRate);
}
void SyncParameter::controlUpdate(const UpdateCallback &callBack) {
CHECK(role_ == TRAINER_ROLE_CONTROL);
CHECK(gradSem_ != NULL && valueSem_ != NULL);
CHECK(majorPartners_.size());
/* update */
if (callBack) {
callBack(localParam_.get());
localParam_->clearGradient();
}
for (size_t i = 0; i < minorPartners_.size(); i++) {
minorPartners_[i]->postValueReady();
}
}
void SyncParameter::majorUpdate(real learnRate) {
(void)learnRate;
CHECK(role_ == TRAINER_ROLE_MAJOR);
CHECK(gradSem_ != NULL && valueSem_ != NULL);
CHECK(minorPartners_.size() && controlParam_);
/* wait the minor-Gradient is ready */
for (size_t i = 0; i < minorPartners_.size(); i++) {
minorPartners_[i]->waitGradReady();
partnerParam_ = minorPartners_[i]->getLocalParameter();
VectorPtr localGrad = localParam_->getBuf(PARAMETER_GRADIENT);
VectorPtr minorGrad = partnerParam_->getBuf(PARAMETER_GRADIENT);
localGrad->add(*minorGrad);
}
/* notice the controller that the gradient is ready */
gradSem_->post();
}
void SyncParameter::minorUpdate(real learnRate) {
(void)learnRate;
CHECK(role_ == TRAINER_ROLE_MINOR);
CHECK(gradSem_ != NULL && valueSem_ != NULL);
// notice the major that the gradient is ready
gradSem_->post();
}
AsyncParameter::AsyncParameter(TrainerRole role,
int asyncCount,
ParameterPtr localParam)
: ParallelParameter(role, localParam) {
asyncCount_ = asyncCount;
accumCounter_ = 0;
gradientAccum_ = Vector::create(localParam->getSize(), localParam->useGpu());
gradientAccum_->zeroMem();
}
void AsyncParameter::slaveUpdate(real learnRate) {
/* increase the accumCounter_ */
accumCounter_++;
/* accumulate the gradient to the buffer */
VectorPtr grad = localParam_->getBuf(PARAMETER_GRADIENT);
gradientAccum_->add(*grad);
/* if need to be synchronized with the master */
if (accumCounter_ == asyncCount_) {
gradSem_->post();
// accumCounter_ = 0; NOTICE: the upper-function need to reset the counter
} else { // self update
localParam_->updateWithGradient(learnRate);
}
localParam_->clearGradient();
}
bool AsyncParameter::masterUpdate(ParallelParameterPtr slaveParam,
const UpdateCallback &callback) {
CHECK(slaveParam && callback);
/* wait the slave is ready */
if (!slaveParam->timeWaitGradReady(5)) {
return false;
}
AsyncParameter *asyncParam = dynamic_cast<AsyncParameter *>(slaveParam.get());
/* get the accum-gradient to update local parameter */
VectorPtr slaveVec = asyncParam->getAccum();
localParam_->getBuf(PARAMETER_GRADIENT)->copyFrom(*slaveVec);
callback(localParam_.get());
// slaveVec->zeroMem();
/* copy the newest parameter-value to the slave */
slaveVec = (slaveParam->getLocalParameter())->getBuf(PARAMETER_VALUE);
slaveVec->copyFrom(*(localParam_->getBuf(PARAMETER_VALUE)));
/* release the semphore */
slaveParam->postValueReady();
return true;
}
} // namespace paddle
paddle/parameter/ParallelParameter.h 已删除 100644 → 0
浏览文件 @ 5d7fabd3
/* 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 <stdint.h>
#include <sys/time.h>
#include <unistd.h>
#include <iostream>
#include <string>
#include <vector>
#include "hl_gpu.h"
#include "paddle/math/Vector.h"
#include "paddle/parameter/Parameter.h"
#include "paddle/parameter/ParameterUpdateFunctions.h"
#include "paddle/utils/Common.h"
#include "paddle/utils/Flags.h"
#include "paddle/utils/Locks.h"
#include "ParameterConfig.pb.h"
namespace paddle {
class ParallelParameter;
class SyncParameter;
class AsyncParameter;
typedef std::shared_ptr<ParallelParameter> ParallelParameterPtr;
const int UPDATE_TYPE_NUM = 32;
/**
* TrainRole denotes the role of current training, different roles have
* different jobs.
*
* control, major, minor are three kinds of role to support mutiple GPUs
* parallel SGD training. SM on GPU card has two groups, each group
* consist of a major and a minor.
*
* @param single single GPU card single thread training.
*
*
* @param control current parameter updates via control role,
* not participate in real training. control role is
* responsible for merging all major's gradient and
* update parameter value.
*
* @param major major role paticipates in real training, when local
* gradient is ready, merge its corresponding minor's
* gradient and notify controller: this group's gradient
* is already ready.
*
* @param minor minor role participates in real training, when local
* gradient is ready, only notify its corresponding major.
* In order to maximum apportion jobs, after controller
* updates the paramemter value, each group's minior
* reponses to dispatch the latest model into local and
* major.
*/
enum TrainerRole {
TRAINER_ROLE_SINGLE,
TRAINER_ROLE_CONTROL,
TRAINER_ROLE_MAJOR,
TRAINER_ROLE_MINOR,
TRAINER_ROLE_MASTER,
TRAINER_ROLE_SLAVE
};
typedef void (ParallelParameter::*UpdateFunction)(real learnRate);
class ParallelParameter {
public:
static ParallelParameterPtr create(TrainerRole role,
ParameterPtr localParam,
int asyncCount = 1);
ParallelParameter(TrainerRole role, ParameterPtr localParam) {
role_ = role;
gradSem_.reset(new Semaphore(0));
valueSem_.reset(new Semaphore(0));
localParam_ = localParam;
}
virtual ~ParallelParameter() {}
ParameterPtr getLocalParameter() { return localParam_; }
bool timeWaitGradReady(int sec) {
struct timespec ts;
ts.tv_nsec = 0;
ts.tv_sec = time(NULL) + sec;
return gradSem_->timeWait(&ts);
}
void waitGradReady() { gradSem_->wait(); }
void postValueReady() { valueSem_->post(); }
void syncUpdate(TrainerRole role, real learnRate);
virtual void synchronizeParamter() = 0;
/**
* for synchronous
*/
virtual void singleUpdate(real learnRate) { (void)learnRate; }
virtual void controlUpdate(const UpdateCallback& callback) { (void)callback; }
virtual void majorUpdate(real learnRate) { (void)learnRate; }
virtual void minorUpdate(real learnRate) { (void)learnRate; }
/**
* for asynchronous
*/
virtual void slaveUpdate(real learnRate) { (void)learnRate; }
protected:
TrainerRole role_;
ParameterPtr localParam_;
std::unique_ptr<Semaphore>
gradSem_; /// wether the local parameter-gradient is ready
std::unique_ptr<Semaphore>
valueSem_; /// wether the local parameter-value is updated
};
/**
* this class is designed for multi-threading training.
*
* "Synchronous" means multiple GPUs calculate 1/4 mini-Batch,
* but will get only one gradient
*/
class SyncParameter : public ParallelParameter {
public:
SyncParameter(TrainerRole role, ParameterPtr localParam)
: ParallelParameter(role, localParam) {
controlParam_ = nullptr;
majorPartners_.clear();
minorPartners_.clear();
}
~SyncParameter() {
majorPartners_.clear();
minorPartners_.clear();
}
void attachControlParam(ParallelParameterPtr controler);
void attachMajorParam(ParallelParameterPtr partner);
void attachMinorParam(ParallelParameterPtr partner, int deviceId);
void waitAllMajorGradReady();
void synchronizeParamter();
void singleUpdate(real learnRate);
void controlUpdate(const UpdateCallback& callback);
void majorUpdate(real learnRate);
void minorUpdate(real learnRate);
std::vector<ParallelParameterPtr>& getMajorPartners() {
return majorPartners_;
}
std::vector<ParallelParameterPtr>& getMinorPartners() {
return minorPartners_;
}
private:
// The following variables are used in a multithreaded training situation
// partnerParam_ is local-parameter's partner
// controlParam_ is the controller-thread 's parameter
ParameterPtr partnerParam_;
std::vector<ParallelParameterPtr> majorPartners_;
std::vector<ParallelParameterPtr> minorPartners_;
std::vector<int> minorDeviceIds_;
ParallelParameterPtr controlParam_;
};
class AsyncParameter : public ParallelParameter {
public:
AsyncParameter(TrainerRole role, int asyncCount, ParameterPtr localParam);
void clearCounter() { accumCounter_ = 0; }
VectorPtr getAccum() { return gradientAccum_; }
void synchronizeParamter() {
if (accumCounter_ == asyncCount_) {
valueSem_->wait();
clearCounter();
gradientAccum_->zeroMem();
}
}
/**
* When asynchronous training, update strategy including slave and master.
*
* slave: If in range asyncCount, adopting self-update method.
* If beyond asyncCount, waiting for master to update.
*/
void slaveUpdate(real learnRate);
/**
* When asynchronous training, update strategy including slave and master.
*
* master: it only polls slaves, do not training data.
* If slave's gradient is ready, fetch it.
* Update master's parameter, then copy it into
* corresponding slave.
*/
bool masterUpdate(ParallelParameterPtr slaveParam,
const UpdateCallback& callback);
private:
/**
* When asynchronous training, every aysnc trainer needs to
* accumulate a number of batch gradient.
*
* gradientAccum_ is used to save the sum of gradients.
*/
VectorPtr gradientAccum_;
/// Asynchronous count.
int asyncCount_;
/// Accumulate counter of current gradients.
int accumCounter_;
};
typedef std::map<std::string, ParallelParameterPtr> ParallelParameterMap;
} // namespace paddle
paddle/parameter/Parameter.cpp
浏览文件 @ ec3111ab
......@@ -271,55 +271,6 @@ SparsePrefetchRowCpuMatrix* Parameter::getPrefetchMatrix() {
return nullptr;
}
void Parameter::updateWithGradient(real learningRate) {
sgdUpdate(learningRate * config_.learning_rate(),
config_.momentum(),
config_.decay_rate(),
bufs_[PARAMETER_VALUE].get(),
bufs_[PARAMETER_GRADIENT].get(),
bufs_[PARAMETER_MOMENTUM].get());
}
void Parameter::updateWithGradient(real learningRate,
MatrixPtr gradMat,
IVectorPtr t0,
int currentTime,
bool fini) {
SparseRowCpuMatrix* sparseMat =
dynamic_cast<SparseRowCpuMatrix*>(gradMat.get());
CHECK(sparseMat);
CHECK_EQ(config_.momentum(), 0.0f)
<< "not support momentum in sparse input sgd";
bool useL1 = (config_.decay_rate_l1() != 0.0f);
sparseMat->sgdUpdate(*bufs_[PARAMETER_VALUE],
*t0,
learningRate * config_.learning_rate(),
currentTime,
useL1 ? config_.decay_rate_l1() : config_.decay_rate(),
useL1,
fini);
}
void Parameter::updateWithGradient(real learningRate,
VectorPtr gradVec,
bool normalUpdate) {
if (normalUpdate) {
sgdUpdate(learningRate * config_.learning_rate(),
config_.momentum(),
config_.decay_rate(),
bufs_[PARAMETER_VALUE].get(),
gradVec.get(),
bufs_[PARAMETER_MOMENTUM].get());
} else {
size_t size = gradVec->getSize();
real* mom = bufs_[PARAMETER_MOMENTUM]->getData();
real* grad = gradVec->getData();
real* value = bufs_[PARAMETER_VALUE]->getData();
hl_matrix_add(mom, grad, mom, 1, size, 1.0f, learningRate);
hl_matrix_add(value, grad, value, 1, size, 1.0f, learningRate);
}
}
void Parameter::incUpdate(const UpdateCallback& callback) {
// Static parameter is fixed, and does not need to be updated
if (isStatic()) {
......
paddle/parameter/Parameter.h
浏览文件 @ ec3111ab
......@@ -223,29 +223,6 @@ public:
bool isValueUpdated() const { return updated_; }
/**
* Update bufs_[PARAMETER_VALUE] using bufs_[PARAMETER_GRADIENT]
*/
void updateWithGradient(real learningRate);
/**
* Update bufs_[PARAMETER_VALUE] using sparse row grad matrix.
*
* @see SparseRowCpuMatrix::sgdUpdate for more information.
*/
void updateWithGradient(real learningRate,
MatrixPtr gradMat,
IVectorPtr t0,
int currentTime,
bool fini = false);
/**
* This function is used to calculate multiple gpus, but only as a candidate
*/
void updateWithGradient(real learningRate,
VectorPtr grad,
bool normalUpdate = true);
/**
* Save parameter value to a file
*/
......
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