private Function PerformTraining(TrainingConfig config)
{
// Setup the criteria (loss and metric)
var crossEntropy = CNTKLib.CrossEntropyWithSoftmax(model, inputModel.LabelSequence);
var errors = CNTKLib.ClassificationError(model, inputModel.LabelSequence);
// Instantiate the trainer object to drive the model training
var learningRatePerSample = new TrainingParameterScheduleDouble(config.LearningRate, 1);
var momentumTimeConstant = CNTKLib.MomentumAsTimeConstantSchedule(config.MomentumTimeConstant);
var additionalParameters = new AdditionalLearningOptions
{
gradientClippingThresholdPerSample = 5.0,
gradientClippingWithTruncation = true
};
var learner = Learner.MomentumSGDLearner(model.Parameters(), learningRatePerSample, momentumTimeConstant, true, additionalParameters);
trainer = Trainer.CreateTrainer(model, crossEntropy, errors, new List <Learner>()
{
learner
});
for (int i = 0; i < config.Epochs; i++)
{
TrainMinibatch(config);
}
return(model);
}
public Function PerformTraining(TrainingConfig config, TrainingDataSource source)
{
this.source = source;
var modelSequence = CreateModel(source.SymbolsCount, 2, 256);
inputModel = CreateInputs(source.SymbolsCount);
model = modelSequence(inputModel.InputSequence);
return(PerformTraining(config));
}
private void TrainMinibatch(TrainingConfig config)
{
uint minibatchesPerEpoch = (uint)Math.Min(source.Length / config.MinibatchSize, config.MaxNumberOfMinibatches / config.Epochs);
for (int j = 0; j < minibatchesPerEpoch; j++)
{
var trainingData = source.GetData(j);
var features = Value.CreateSequence <float>(inputModel.InputSequence.Shape,
trainingData.InputSequence, device);
var arguments = new Dictionary <Variable, Value>();
arguments.Add(inputModel.InputSequence, features);
var labels = Value.CreateSequence(inputModel.LabelSequence.Shape,
trainingData.OutputSequence, device);
arguments.Add(inputModel.LabelSequence, labels);
trainer.TrainMinibatch(arguments, device);
}
}
public TextDataSource(TrainingConfig config, string filePath)
: base(config)
{
LoadData(filePath);
}
public TrainingDataSource(TrainingConfig config)
{
MinibatchLength = config.MinibatchSize;
}