public void Run(int firstHiddenLayerSize, int secondHiddenLayerSize, int iterations, float learningRate, int batchSize)
{
var digitReader = new HandwrittenDigitReader();
var trainDigits = digitReader.Iterate(
MnistFiles.FindByType(MNISTFileType.TrainImages).FileName,
MnistFiles.FindByType(MNISTFileType.TrainLabels).FileName);
var testDigits = digitReader.Iterate(
MnistFiles.FindByType(MNISTFileType.TestImages).FileName,
MnistFiles.FindByType(MNISTFileType.TestLabels).FileName);
var trainingData = trainDigits.Select(im => im.ToDataPoint());
var testData = testDigits.Select(im => im.ToDataPoint());
var cancellationToken = new CancellationTokenSource();
_model = new PerceptronModel(
HandwrittenDigit.ImageSize * HandwrittenDigit.ImageSize,
firstHiddenLayerSize,
secondHiddenLayerSize,
10);
_classifier = new PerceptronClassifier(
_model,
learningRate,
batchSize);
var watch = Stopwatch.StartNew();
_classifier.IterationComplete += j =>
{
//Console.WriteLine($"Epoch {j}: {MNISTEvaulate(net, testData)} / {testData.Count}");
OnLog(
$"Iteration #{j} has completed: {EvaluateClassifier(_classifier, testData)} / {testData.Count()} [{watch.ElapsedMilliseconds} ms]");
watch = Stopwatch.StartNew();
if (j == iterations)
{
cancellationToken.Cancel();
}
EpochComplete?.Invoke(j);
};
_classifier.Train(trainingData, cancellationToken.Token);
}
public void EndEpoch(double trainingError)
{
_timer.Stop();
var delta = trainingError - LastTrainingError;
LastTrainingError = trainingError;
_trainingErrorDelta.Push(delta);
EpochComplete?.Invoke(this);
float trainingRate;
if (_scheduledTrainingRate.TryGetValue(CurrentEpoch, out trainingRate))
{
TrainingRate = trainingRate;
Console.WriteLine("Training rate changed to " + trainingRate);
}
}
