microsoft
/
CNTK
зеркало из
1
0
Форкнуть 0
Код Задачи Пакеты Проекты Релизы Вики Активность

cleaned up some dead code related to logging

This commit is contained in:
Frank Seide 2016-05-02 22:06:27 -07:00
Родитель 39038a4138
Коммит 16b606c711
4 изменённых файлов: 4 добавлений и 55 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

1
Source/ComputationNetworkLib/MatrixPool.h
Просмотреть файл

@ -36,6 +36,7 @@ public:
if (freeMatrix == nullptr || freeMatrix->GetMatrixType() == SPARSE)
LogicError("MatrixPool::Release: freeMatrix should not be null or sparse.");
//#define SUPRESS_MEMSHARING // #define this to disable memory sharing through this structure
// TODO: Make this a runtime option.
#ifndef SUPRESS_MEMSHARING
vector<shared_ptr<Matrix<ElemType>>>& releasedMatrices = GetReleasedMatrices<ElemType>();
#ifdef _DEBUG

19
Source/SGDLib/Criterion.h
Просмотреть файл

@ -115,7 +115,6 @@ private:
size_t beta = reset ? 0 : 1;
size_t numSamples = GetNumSamples(nodes[i], legacyNumSamples);
#if 1
// For criterion nodes that emit criteria per frame, we will at this point
// do masking and an implicit reduction.
@ -128,8 +127,6 @@ private:
shape.NarrowTo(1, i, i + 1); // narrow to the single element that corresponds to the accumulator value
auto criterionAccumulator = TensorView<ElemType>(m_aggregateCriterionValues, shape);
//fprintf(stderr, "Accumulating %ls, legNumSamp = %d, numSamp = %d\n", nodes[i]->NodeName().c_str(), (int)legacyNumSamples, (int)numSamples);
//node->As<ComputationNode<ElemType>>()->Value().Print(msra::strfun::strprintf("crit[%d] m_value", (int)i).c_str());
if (numSamples > 0) // (if MB is empty, matrix may not have the correct row dmension)
{
auto criterionValue = node->As<ComputationNode<ElemType>>()->ValueTensorFor(SIZE_MAX, fr);
@ -138,23 +135,7 @@ private:
// If count is zero, we lazily consider the numerator as zero as well.
criterionAccumulator.DoCopyOf(m_aggregateSampleCounts[i] ? (float)beta : 0, criterionValue, 1);
}
//m_aggregateCriterionValues->Print(msra::strfun::strprintf("crit[%d] m_aggregateCriterionValues", (int)i).c_str());
m_aggregateSampleCounts[i] = m_aggregateSampleCounts[i] * beta + numSamples;
#else
// temp solution until we add TensorView reduction
if (beta == 0)
{
Matrix<ElemType>::AssignElementToElement(dynamic_pointer_cast<ComputationNode<ElemType>>(node)->Value(),
0, 0, *m_aggregateCriterionValues, 0, i);
m_aggregateSampleCounts[i] = numSamples;
}
else if (numSamples > 0) // avoid unnecessary GPU access
{
Matrix<ElemType>::AddElementToElement(dynamic_pointer_cast<ComputationNode<ElemType>>(node)->Value(),
0, 0, *m_aggregateCriterionValues, 0, i);
m_aggregateSampleCounts[i] += numSamples;
}
#endif
return *this;
}
// get the number of samples

37
Source/SGDLib/SGD.cpp
Просмотреть файл

@ -9,7 +9,7 @@
#include "DataReaderHelpers.h"
#include "MatrixQuantizerImpl.h"
#ifdef CNTK_PARALLEL_TRAINING_SUPPORT
static inline bool operator==(const std::pair<double,size_t>& a, double b) { assert(b==0); return a.first == b; }
//static inline bool operator==(const std::pair<double,size_t>& a, double b) { assert(b==0); return a.first == b; }
// ^^ workaround until this line in AggregateGradientsImpl() gets updated: assert(headerCPU->evalErrors[i] == 0);
#include "AllReduceDistGradAggregator.h"
#include "BlockMomentumSGD.h"
@ -483,14 +483,9 @@ void SGD<ElemType>::TrainOrAdaptModel(int startEpoch, ComputationNetworkPtr net,
else
lrControlCriterion = epochCriterion.Average();
#if 1
LOGPRINTF(stderr, "Finished Epoch[%2d of %d]: [Training] ", i + 1, (int)m_maxEpochs);
epochCriterion.LogCriterion(criterionNodes[0]->NodeName());
#else
LOGPRINTF(stderr,
"Finished Epoch[%2d of %d]: [Training Set] TrainLossPerSample = %.8g; TotalSamplesSeen = %d; ",
i + 1, (int)m_maxEpochs, epochCriterion.Average(), (int)totalTrainingSamplesSeen);
#endif
m_lastFinishedEpochTrainLoss = epochCriterion.Average();
for (size_t j = 0; j < epochEvalErrors.size(); j++)
epochEvalErrors[j].LogCriterion(evaluationNodes[j]->NodeName());
@ -1078,8 +1073,6 @@ size_t SGD<ElemType>::TrainOneEpoch(ComputationNetworkPtr net,
}
// epochCriterion aggregates over entire epoch, but we only show difference to last time we logged
// BUGBUG: How does this work if useGradientAggregation? epochCriterion has not been set.
//double trainLossSinceLastLogged = (trainSamplesSinceLastLogged != 0) ? ((epochCriterion - epochCriterionLastLogged) / trainSamplesSinceLastLogged) : 0.0;
EpochCriterion epochCriterionSinceLastLogged = epochCriterion - epochCriterionLastLogged;
let trainLossSinceLastLogged = epochCriterionSinceLastLogged.Average(); // TODO: Check whether old trainSamplesSinceLastLogged matches this ^^ difference
let trainSamplesSinceLastLogged = (int)epochCriterionSinceLastLogged.second;
@ -1113,10 +1106,7 @@ size_t SGD<ElemType>::TrainOneEpoch(ComputationNetworkPtr net,
if (epochNumber > 0 || (int)epochSize > 0) // got anything? --TODO: why cast epochSize to (int) for this comparison?
fprintf(stderr, (", %2." + to_string(mbProgNumPrecision) + "f%%").c_str(), mbProg * 100); // --TODO: use a * format?
fprintf(stderr, "]: ");
//fprintf(stderr, "SamplesSeen = %d; ", (int)trainSamplesSinceLastLogged);
epochCriterionSinceLastLogged.LogCriterion(criterionNodes[0]->NodeName());
//fprintf(stderr, ("%ls = " + GeneratePaddedFloatOrExpFormat(11, 8, trainLossSinceLastLogged) + " * %d").c_str(),
// criterionNodes[0]->NodeName().c_str(), trainLossSinceLastLogged, trainSamplesSinceLastLogged);
for (size_t i = 0; i < epochEvalErrors.size(); i++)
(epochEvalErrors[i] - epochEvalErrorsLastLogged[i]).LogCriterion(evaluationNodes[i]->NodeName());
@ -1280,7 +1270,7 @@ bool SGD<ElemType>::PreCompute(ComputationNetworkPtr net,
// trainSetDataReader->StartMinibatchLoop(m_mbSize[0], 0 , requestDataSize);
// trainSetDataReader->StartMinibatchLoop(m_mbSize[0], 0 , m_epochSize); // only based on one epoch
// [1/12/2015 erw] to support large dataset, we usually partition whole dataset into several epoch's,
// To support large dataset, we usually partition whole dataset into several epoch's,
// so we need to use all the data to do precomputing
if (m_useAllDataForPreComputedNode) // using all the data
trainSetDataReader->StartMinibatchLoop(m_mbSize[0], 0);
@ -1757,27 +1747,6 @@ void SGD<ElemType>::AttemptUtteranceDerivativeFeatures(ComputationNetworkPtr net
}
}
#if 0
template <class ElemType>
int SGD<ElemType>::SGDTrace(FILE* __restrict __stream, bool isPrependTimestamp, const char* __restrict __format, ...)
{
int result = 0;
if (m_traceLevel > 0)
{
va_list args;
va_start(args, __format);
if (isPrependTimestamp)
{
PREPENDTS(__stream);
}
result = vfprintf(__stream, __format, args);
va_end(args);
}
return result;
}
#endif
template <class ElemType>
void SGD<ElemType>::InitDistGradAgg(int numEvalNodes, int traceLevel)
{

2
Source/SGDLib/SGD.h
Просмотреть файл

@ -310,7 +310,6 @@ public:
// TODO: The next few do not belong into SGD any more than the network or reader we operate on. Either move network and reader in here, or move these out.
m_modelPath((const wstring&) configSGD(L"modelPath")),
m_keepCheckPointFiles(configSGD(L"keepCheckPointFiles", false)),
// m_validateAfterModelReloading(configSGD(L"validateAfterModelReloading", true)),
m_trainCriterionNodeName((const wstring&) configSGD(L"trainCriterionNodeName", L"")),
m_evalCriterionNodeName ((const wstring&) configSGD(L"evalCriterionNodeName", L"")),
m_traceNodeNamesReal (configSGD(L"traceNodeNamesReal", ConfigRecordType::Array(stringargvector()))),
@ -546,7 +545,6 @@ public:
protected:
std::wstring m_modelPath;
bool m_keepCheckPointFiles;
// bool m_validateAfterModelReloading; // TODO: remove this. Why would one not validate a model?
std::wstring m_trainCriterionNodeName;
std::wstring m_evalCriterionNodeName;