internal async Task Train(int numLayers, CancellationToken cancellation, IProgress <Tuple <string, int> > taskAndPercentComplete)
{
for (var i = numLayers; i > 0; i--)
{
await Task.Run(() => _trainer.AddLayer(cancellation, taskAndPercentComplete));
}
}
static void Main(string[] args)
{
try
{
for (var noise = 0; noise < 2; noise++)
{
var error = 0.05 * noise;
List<Sample> pointsPos, pointsNeg;
var flippedLabels = GenerateSamples(error, out pointsPos, out pointsNeg);
var t = new Trainer<Sample>(new ILearner<Sample>[] { new Sample.Learner(21) }, pointsPos, pointsNeg);
var convergence = 0;
var prev = 0;
Console.WriteLine("Noise: " + error + " (" + flippedLabels + " label errors)");
for (var i = 0; i < 1000; i++)
{
var cost = t.AddLayer();
if (float.IsNaN(cost))
{
Console.WriteLine("Terminated after " + i + " layers.");
break;
}
//Console.WriteLine("Loss = " + (float)cost);
var testErrors = TestTrainer(t.Classifier, pointsPos, pointsNeg);
var veriErrors = TestTrainer(t.Classifier);
if (testErrors == prev) convergence++;
else convergence = 0;
if (convergence > 19)
{
Console.WriteLine("Fully trained after " + (i - 9) + " layers.");
break;
}
else if ((i + 1) % 10 == 0)
{
Console.WriteLine("[" + testErrors + " / " + veriErrors + " ] errors after " + (i + 1) + " layers.");
}
prev = testErrors;
}
}
}
catch (Exception e)
{
Console.WriteLine(e.ToString());
}
Console.ReadKey();
}
