private void CreateCheckPoint(ParallelCorpus validCorpus, List <IMetric> metrics, IModelMetaData modelMetaData, Func <IComputeGraph, List <List <string> >, List <List <string> >, int, bool, float> ForwardOnSingleDevice, double avgCostPerWordInTotal)
{
if (validCorpus != null)
{
// The valid corpus is provided, so evaluate the model.
if (RunValid(validCorpus, ForwardOnSingleDevice, metrics) == true)
{
SaveModel(modelMetaData);
}
}
else if (m_avgCostPerWordInTotalInLastEpoch > avgCostPerWordInTotal)
{
// We don't have valid corpus, so if we could have lower cost, save the model
SaveModel(modelMetaData);
}
TensorAllocator.FreeMemoryAllDevices();
}
internal void TrainOneEpoch(int ep, ParallelCorpus trainCorpus, ParallelCorpus validCorpus, ILearningRate learningRate, AdamOptimizer solver, List <IMetric> metrics, IModelMetaData modelMetaData,
Func <IComputeGraph, List <List <string> >, List <List <string> >, int, bool, float> ForwardOnSingleDevice)
{
int processedLineInTotal = 0;
DateTime startDateTime = DateTime.Now;
DateTime lastCheckPointDateTime = DateTime.Now;
double costInTotal = 0.0;
long srcWordCntsInTotal = 0;
long tgtWordCntsInTotal = 0;
double avgCostPerWordInTotal = 0.0;
TensorAllocator.FreeMemoryAllDevices();
Logger.WriteLine($"Start to process training corpus.");
List <SntPairBatch> sntPairBatchs = new List <SntPairBatch>();
foreach (SntPairBatch sntPairBatch in trainCorpus)
{
sntPairBatchs.Add(sntPairBatch);
if (sntPairBatchs.Count == m_deviceIds.Length)
{
// Copy weights from weights kept in default device to all other devices
CopyWeightsFromDefaultDeviceToAllOtherDevices();
int batchSplitFactor = 1;
bool runNetwordSuccssed = false;
while (runNetwordSuccssed == false)
{
try
{
(float cost, int sWordCnt, int tWordCnt, int processedLine) = RunNetwork(ForwardOnSingleDevice, sntPairBatchs, batchSplitFactor);
processedLineInTotal += processedLine;
srcWordCntsInTotal += sWordCnt;
tgtWordCntsInTotal += tWordCnt;
//Sum up gradients in all devices, and kept it in default device for parameters optmization
SumGradientsToTensorsInDefaultDevice();
//Optmize parameters
float lr = learningRate.GetCurrentLearningRate();
List <IWeightTensor> models = GetParametersFromDefaultDevice();
solver.UpdateWeights(models, processedLine, lr, m_regc, m_weightsUpdateCount + 1);
costInTotal += cost;
avgCostPerWordInTotal = costInTotal / tgtWordCntsInTotal;
m_weightsUpdateCount++;
if (IterationDone != null && m_weightsUpdateCount % 100 == 0)
{
IterationDone(this, new CostEventArg()
{
LearningRate = lr,
CostPerWord = cost / tWordCnt,
AvgCostInTotal = avgCostPerWordInTotal,
Epoch = ep,
Update = m_weightsUpdateCount,
ProcessedSentencesInTotal = processedLineInTotal,
ProcessedWordsInTotal = srcWordCntsInTotal + tgtWordCntsInTotal,
StartDateTime = startDateTime
});
}
runNetwordSuccssed = true;
}
catch (Exception err)
{
batchSplitFactor *= 2;
Logger.WriteLine($"Increase batch split factor to {batchSplitFactor}, and retry it.");
if (batchSplitFactor >= sntPairBatchs[0].BatchSize)
{
Logger.WriteLine($"Batch split factor is larger than batch size, give it up.");
throw err;
}
}
}
// Evaluate model every hour and save it if we could get a better one.
TimeSpan ts = DateTime.Now - lastCheckPointDateTime;
if (ts.TotalHours > 1.0)
{
CreateCheckPoint(validCorpus, metrics, modelMetaData, ForwardOnSingleDevice, avgCostPerWordInTotal);
lastCheckPointDateTime = DateTime.Now;
}
sntPairBatchs.Clear();
}
}
Logger.WriteLine(Logger.Level.info, ConsoleColor.Green, $"Epoch '{ep}' took '{DateTime.Now - startDateTime}' time to finish. AvgCost = {avgCostPerWordInTotal.ToString("F6")}, AvgCostInLastEpoch = {m_avgCostPerWordInTotalInLastEpoch.ToString("F6")}");
CreateCheckPoint(validCorpus, metrics, modelMetaData, ForwardOnSingleDevice, avgCostPerWordInTotal);
m_avgCostPerWordInTotalInLastEpoch = avgCostPerWordInTotal;
}
internal void TrainOneEpoch(int ep, ParallelCorpus trainCorpus, ParallelCorpus validCorpus, ILearningRate learningRate, AdamOptimizer solver, List <IMetric> metrics, IModelMetaData modelMetaData,
Func <IComputeGraph, List <List <string> >, List <List <string> >, int, bool, float> ForwardOnSingleDevice)
{
int processedLineInTotal = 0;
DateTime startDateTime = DateTime.Now;
DateTime lastCheckPointDateTime = DateTime.Now;
double costInTotal = 0.0;
long srcWordCnts = 0;
long tgtWordCnts = 0;
double avgCostPerWordInTotal = 0.0;
TensorAllocator.FreeMemoryAllDevices();
Logger.WriteLine($"Start to process training corpus.");
List <SntPairBatch> sntPairBatchs = new List <SntPairBatch>();
foreach (SntPairBatch sntPairBatch in trainCorpus)
{
sntPairBatchs.Add(sntPairBatch);
if (sntPairBatchs.Count == m_deviceIds.Length)
{
float cost = 0.0f;
int tlen = 0;
int processedLine = 0;
// Copy weights from weights kept in default device to all other devices
CopyWeightsFromDefaultDeviceToAllOtherDevices();
// Run forward and backward on all available processors
Parallel.For(0, m_deviceIds.Length, i =>
{
SntPairBatch sntPairBatch_i = sntPairBatchs[i];
// Construct sentences for encoding and decoding
List <List <string> > srcTkns = new List <List <string> >();
List <List <string> > tgtTkns = new List <List <string> >();
int sLenInBatch = 0;
int tLenInBatch = 0;
for (int j = 0; j < sntPairBatch_i.BatchSize; j++)
{
srcTkns.Add(sntPairBatch_i.SntPairs[j].SrcSnt.ToList());
sLenInBatch += sntPairBatch_i.SntPairs[j].SrcSnt.Length;
tgtTkns.Add(sntPairBatch_i.SntPairs[j].TgtSnt.ToList());
tLenInBatch += sntPairBatch_i.SntPairs[j].TgtSnt.Length;
}
float lcost = 0.0f;
// Create a new computing graph instance
using (IComputeGraph computeGraph_i = CreateComputGraph(i))
{
// Run forward part
lcost = ForwardOnSingleDevice(computeGraph_i, srcTkns, tgtTkns, i, true);
// Run backward part and compute gradients
computeGraph_i.Backward();
}
lock (locker)
{
cost += lcost;
srcWordCnts += sLenInBatch;
tgtWordCnts += tLenInBatch;
tlen += tLenInBatch;
processedLineInTotal += sntPairBatch_i.BatchSize;
processedLine += sntPairBatch_i.BatchSize;
}
});
//Sum up gradients in all devices, and kept it in default device for parameters optmization
SumGradientsToTensorsInDefaultDevice();
//Optmize parameters
float lr = learningRate.GetCurrentLearningRate();
List <IWeightTensor> models = GetParametersFromDefaultDevice();
solver.UpdateWeights(models, processedLine, lr, m_regc, m_weightsUpdateCount + 1);
//Clear gradient over all devices
ZeroGradientOnAllDevices();
costInTotal += cost;
avgCostPerWordInTotal = costInTotal / tgtWordCnts;
m_weightsUpdateCount++;
if (IterationDone != null && m_weightsUpdateCount % 100 == 0)
{
IterationDone(this, new CostEventArg()
{
LearningRate = lr,
CostPerWord = cost / tlen,
AvgCostInTotal = avgCostPerWordInTotal,
Epoch = ep,
Update = m_weightsUpdateCount,
ProcessedSentencesInTotal = processedLineInTotal,
ProcessedWordsInTotal = srcWordCnts + tgtWordCnts,
StartDateTime = startDateTime
});
}
// Evaluate model every hour and save it if we could get a better one.
TimeSpan ts = DateTime.Now - lastCheckPointDateTime;
if (ts.TotalHours > 1.0)
{
CreateCheckPoint(validCorpus, metrics, modelMetaData, ForwardOnSingleDevice, avgCostPerWordInTotal);
lastCheckPointDateTime = DateTime.Now;
}
sntPairBatchs.Clear();
}
}
Logger.WriteLine(Logger.Level.info, ConsoleColor.Green, $"Epoch '{ep}' took '{DateTime.Now - startDateTime}' time to finish. AvgCost = {avgCostPerWordInTotal.ToString("F6")}, AvgCostInLastEpoch = {m_avgCostPerWordInTotalInLastEpoch.ToString("F6")}");
CreateCheckPoint(validCorpus, metrics, modelMetaData, ForwardOnSingleDevice, avgCostPerWordInTotal);
m_avgCostPerWordInTotalInLastEpoch = avgCostPerWordInTotal;
}
