static void Train(
Net net,
Loss lossFn,
Dataset dataset,
int epochs = Config.NUM_EPOCHS,
int logEvery = Config.LOG_EVERY,
float validationRatio = 0.2f
)
{
Console.WriteLine(STARLINE);
Console.WriteLine("TRAINING");
Console.WriteLine(UNDERLINE);
// Reserve some elements for validation
var validationSize = (int)MathF.Round(dataset.Length * validationRatio);
var numberOfExamples = dataset.Length - validationSize;
var validationIndices = Enumerable.Range(0, validationSize).ToArray();
string epochFormat = 'D' + epochs.ToString().Length.ToString();
var learningRate = Config.LEARNING_RATE;
for (int epoch = 0; epoch < epochs; ++epoch)
{
var epochLosses = new float[numberOfExamples];
var epochIndicators = new float[numberOfExamples];
net.SetLearningRate(learningRate);
for (int i = validationSize; i < numberOfExamples; ++i)
{
var dataAndAnswer = dataset.GetItem(i);
var output = net.ForwardPass(dataAndAnswer.Item1);
var outputAndAnswer = new Tuple <float[], float[]>(output, dataAndAnswer.Item2);
var gradient = lossFn.Derivative(outputAndAnswer);
epochLosses[i] = lossFn.Calculate(outputAndAnswer).Average();
epochIndicators[i] = Convert.ToSingle(
(MathF.Round(output[0]) == dataAndAnswer.Item2[0]) &&
(MathF.Round(output[1]) == dataAndAnswer.Item2[1])
);
net.BackwardPass(gradient);
}
Console.WriteLine($"epoch: [{(epoch + 1).ToString(epochFormat)}/{epochs}]\tmean loss: {epochLosses.Average():F5}\taccuracy: {epochIndicators.Average()}");
// Log validation results if needed
if ((epoch + 1) % logEvery == 0)
{
Console.WriteLine(SEPARATOR);
Evaluate(
net: net,
lossFn: lossFn,
dataset: dataset,
indices: validationIndices
);
Console.WriteLine(SEPARATOR);
}
if ((epoch + 1) % (epochs / 3) == 0)
{
var newLearningRate = learningRate * 0.1f;
Console.WriteLine(SEPARATOR);
Console.WriteLine($"Learning rate decayed from: {learningRate} to {newLearningRate}");
learningRate = newLearningRate;
Console.WriteLine(SEPARATOR);
}
}
Console.WriteLine("\nFINISHED");
Console.WriteLine(STARLINE);
}