public override void Train(string datasetFile)
{
var dataSource = CreateDataSource(datasetFile);
var trainer = CreateTrainer();
const int minibatchSize = 4;
const int maxMinibatches = 100;
var minibatchesSeen = 0;
while (true)
{
var minibatchData = dataSource.MinibatchSource.GetNextMinibatch(minibatchSize, _device);
var arguments = new Dictionary <Variable, MinibatchData>
{
{ _modelWrapper.Input, minibatchData[dataSource.FeatureStreamInfo] },
{ _modelWrapper.TrainingOutput, minibatchData[dataSource.LabelStreamInfo] }
};
trainer.TrainMinibatch(arguments, _device);
var trainingProgressResponse = new TrainingProgressResponse
{
MinibatchesSeen = minibatchesSeen,
Loss = trainer.PreviousMinibatchLossAverage(),
EvaluationCriterion = trainer.PreviousMinibatchEvaluationAverage()
};
OnTrainingIterationPerformed(trainingProgressResponse);
minibatchesSeen++;
if (minibatchesSeen >= maxMinibatches)
{
OnTrainingFinished(new TrainingResultResponse
{
NewModelData = _modelWrapper.Model.Save(),
Progress = trainingProgressResponse
});
break;
}
}
}
protected void OnTrainingIterationPerformed(TrainingProgressResponse trainingProgressResponse) => TrainingIterationPerformed?.Invoke(this, trainingProgressResponse);
private static void TrainingIterationPerformed(
object sender, TrainingProgressResponse trainingProgressResponse) =>
WriteLine(
$"Minibatch: {trainingProgressResponse.MinibatchesSeen} " +
$"CrossEntropyLoss = {trainingProgressResponse.Loss} " +
$"EvaluationCriterion = {trainingProgressResponse.EvaluationCriterion}");
