public override void SaveModel(string filename)
{
//Save bi-directional model
forwardRNN.Hidden2OutputWeight = Hidden2OutputWeight;
backwardRNN.Hidden2OutputWeight = Hidden2OutputWeight;
forwardRNN.CRFTagTransWeights = CRFTagTransWeights;
backwardRNN.CRFTagTransWeights = CRFTagTransWeights;
forwardRNN.SaveModel(filename + ".forward");
backwardRNN.SaveModel(filename + ".backward");
//Save meta data
using (StreamWriter sw = new StreamWriter(filename))
{
BinaryWriter fo = new BinaryWriter(sw.BaseStream);
fo.Write((int)ModelType);
fo.Write((int)ModelDirection);
// Signiture , 0 is for RNN or 1 is for RNN-CRF
int iflag = 0;
if (IsCRFTraining == true)
{
iflag = 1;
}
fo.Write(iflag);
fo.Write(L0);
fo.Write(L1);
fo.Write(L2);
fo.Write(DenseFeatureSize);
}
}
public void Train()
{
//Create neural net work
Logger.WriteLine("Create a new network according settings in configuration file.");
Logger.WriteLine($"Processor Count = {Environment.ProcessorCount}");
RNN <T> rnn = RNN <T> .CreateRNN(networkType);
if (ModelSettings.IncrementalTrain)
{
Logger.WriteLine($"Loading previous trained model from {modelFilePath}.");
rnn.LoadModel(modelFilePath, true);
}
else
{
Logger.WriteLine("Create a new network.");
rnn.CreateNetwork(hiddenLayersConfig, outputLayerConfig, TrainingSet, featurizer);
//Create tag-bigram transition probability matrix only for sequence RNN mode
if (IsCRFTraining)
{
Logger.WriteLine("Initialize bigram transition for CRF output layer.");
rnn.InitializeCRFWeights(TrainingSet.CRFLabelBigramTransition);
}
}
rnn.MaxSeqLength = maxSequenceLength;
rnn.bVQ = ModelSettings.VQ != 0 ? true : false;
rnn.IsCRFTraining = IsCRFTraining;
int N = Environment.ProcessorCount * 2;
List <RNN <T> > rnns = new List <RNN <T> >();
rnns.Add(rnn);
for (int i = 1; i < N; i++)
{
rnns.Add(rnn.Clone());
}
//Initialize RNNHelper
RNNHelper.LearningRate = ModelSettings.LearningRate;
RNNHelper.vecNormalLearningRate = new Vector <float>(RNNHelper.LearningRate);
RNNHelper.GradientCutoff = ModelSettings.GradientCutoff;
RNNHelper.vecMaxGrad = new Vector <float>(RNNHelper.GradientCutoff);
RNNHelper.vecMinGrad = new Vector <float>(-RNNHelper.GradientCutoff);
RNNHelper.IsConstAlpha = ModelSettings.IsConstAlpha;
RNNHelper.MiniBatchSize = ModelSettings.MiniBatchSize;
Logger.WriteLine("");
Logger.WriteLine("Iterative training begins ...");
var bestTrainTknErrCnt = long.MaxValue;
var bestValidTknErrCnt = long.MaxValue;
var lastAlpha = RNNHelper.LearningRate;
var iter = 0;
ParallelOptions parallelOption = new ParallelOptions();
parallelOption.MaxDegreeOfParallelism = -1;
while (true)
{
if (ModelSettings.MaxIteration > 0 && iter > ModelSettings.MaxIteration)
{
Logger.WriteLine("We have trained this model {0} iteration, exit.");
break;
}
var start = DateTime.Now;
Logger.WriteLine($"Start to training {iter} iteration. learning rate = {RNNHelper.LearningRate}");
//Clean all RNN instances' status for training
foreach (var r in rnns)
{
r.CleanStatusForTraining();
}
Process(rnns, TrainingSet, RunningMode.Training);
var duration = DateTime.Now.Subtract(start);
Logger.WriteLine($"End {iter} iteration. Time duration = {duration}");
Logger.WriteLine("");
if (tknErrCnt >= bestTrainTknErrCnt && lastAlpha != RNNHelper.LearningRate)
{
//Although we reduce alpha value, we still cannot get better result.
Logger.WriteLine(
$"Current token error count({(double)tknErrCnt / (double)processedWordCnt * 100.0}%) is larger than the previous one({(double)bestTrainTknErrCnt / (double)processedWordCnt * 100.0}%) on training set. End training early.");
Logger.WriteLine("Current alpha: {0}, the previous alpha: {1}", RNNHelper.LearningRate, lastAlpha);
break;
}
lastAlpha = RNNHelper.LearningRate;
int trainTknErrCnt = tknErrCnt;
//Validate the model by validated corpus
if (ValidationSet != null)
{
Logger.WriteLine("Verify model on validated corpus.");
Process(rnns, ValidationSet, RunningMode.Validate);
Logger.WriteLine("End model verification.");
Logger.WriteLine("");
if (tknErrCnt < bestValidTknErrCnt)
{
//We got better result on validated corpus, save this model
Logger.WriteLine($"Saving better model into file {modelFilePath}, since we got a better result on validation set.");
Logger.WriteLine($"Error token percent: {(double)tknErrCnt / (double)processedWordCnt * 100.0}%, Error sequence percent: {(double)sentErrCnt / (double)processedSequence * 100.0}%");
rnn.SaveModel(modelFilePath);
bestValidTknErrCnt = tknErrCnt;
}
}
else if (trainTknErrCnt < bestTrainTknErrCnt)
{
//We don't have validate corpus, but we get a better result on training corpus
//We got better result on validated corpus, save this model
Logger.WriteLine($"Saving better model into file {modelFilePath}, although validation set doesn't exist, we have better result on training set.");
Logger.WriteLine($"Error token percent: {(double)trainTknErrCnt / (double)processedWordCnt * 100.0}%, Error sequence percent: {(double)sentErrCnt / (double)processedSequence * 100.0}%");
rnn.SaveModel(modelFilePath);
}
Logger.WriteLine("");
if (trainTknErrCnt >= bestTrainTknErrCnt)
{
//We don't have better result on training set, so reduce learning rate
RNNHelper.LearningRate = RNNHelper.LearningRate / 2.0f;
RNNHelper.vecNormalLearningRate = new Vector <float>(RNNHelper.LearningRate);
}
else
{
bestTrainTknErrCnt = trainTknErrCnt;
}
iter++;
}
}
public void Process(RNN <T> rnn, DataSet <T> trainingSet, RunningMode runningMode, int totalSequenceNum)
{
//Shffle training corpus
trainingSet.Shuffle();
for (var i = 0; i < trainingSet.SequenceList.Count; i++)
{
var pSequence = trainingSet.SequenceList[i];
int wordCnt = 0;
if (pSequence is Sequence)
{
wordCnt = (pSequence as Sequence).States.Length;
}
else
{
SequencePair sp = pSequence as SequencePair;
if (sp.srcSentence.TokensList.Count > rnn.MaxSeqLength)
{
continue;
}
wordCnt = sp.tgtSequence.States.Length;
}
if (wordCnt > rnn.MaxSeqLength)
{
continue;
}
Interlocked.Add(ref processedWordCnt, wordCnt);
int[] predicted;
if (IsCRFTraining)
{
predicted = rnn.ProcessSequenceCRF(pSequence as Sequence, runningMode);
}
else
{
Matrix <float> m;
predicted = rnn.ProcessSequence(pSequence, runningMode, false, out m);
}
int newTknErrCnt;
if (pSequence is Sequence)
{
newTknErrCnt = GetErrorTokenNum(pSequence as Sequence, predicted);
}
else
{
newTknErrCnt = GetErrorTokenNum((pSequence as SequencePair).tgtSequence, predicted);
}
Interlocked.Add(ref tknErrCnt, newTknErrCnt);
if (newTknErrCnt > 0)
{
Interlocked.Increment(ref sentErrCnt);
}
Interlocked.Increment(ref processedSequence);
if (processedSequence % 1000 == 0)
{
Logger.WriteLine("Progress = {0} ", processedSequence / 1000 + "K/" + totalSequenceNum / 1000.0 + "K");
Logger.WriteLine(" Error token ratio = {0}%", (double)tknErrCnt / (double)processedWordCnt * 100.0);
Logger.WriteLine(" Error sentence ratio = {0}%", (double)sentErrCnt / (double)processedSequence * 100.0);
}
if (ModelSettings.SaveStep > 0 && processedSequence % ModelSettings.SaveStep == 0)
{
//After processed every m_SaveStep sentences, save current model into a temporary file
Logger.WriteLine("Saving temporary model into file...");
rnn.SaveModel("model.tmp");
}
}
}
