public Task <TrainingResult> Train(bool liveTracking)
{
return(Task.Run(() =>
{
int inputSize = NetworkSampleInput.InputSize(NumberOfVertices);
Random random = new Random(Seed);
Network network = NetworkBuilder.Build(new NetworkDefinition(FunctionName.ReLU, inputSize, NumberOfVertices, HiddenLayers), new He(Seed));
Optimizer optimizer = new SGDMomentum(network, LearningRate, Momentum);
Trainer trainer = new Trainer(optimizer, random);
GraphPath[] trainingGraphs = GraphGenerator.Generate(random, NumberOfGraphs, NumberOfVertices).Select(Dijkstra.Find).ToArray();
NetworkSample[] trainingData = NetworkSampleGenerator.Generate(trainingGraphs).ToArray();
MeanSquareErrorMonitor monitor = new MeanSquareErrorMonitor();
trainer.Monitors.Add(monitor);
if (liveTracking)
{
trainer.Monitors.Add(new TrainingMonitor(Name));
}
trainer.Train(trainingData, Epoches, BatchSize);
lock (ThreadLock)
{
double error = MeanSquareErrorMonitor.CalculateError(network, trainingData);
Console.WriteLine($"[{Name}] Seed: {Seed,12} Momentum: {Momentum,-6} Learning Rate: {LearningRate,-6} MSE: {error}");
}
return new TrainingResult(Name, monitor.CollectedData as IEnumerable <double>, network, trainingGraphs, trainingData);
}));
}
public void Fit(IEnumerable <MarkovHistoryItem> batch, AgentNetworkTrainingConfiguration configuration, Random random)
{
var optimizer = new SGDMomentum(Network, configuration.LearningRate, configuration.Momentum);
var trainer = new Trainer(optimizer, random);
var nextQ = batch.Select(item => new Projection(item.Input, new double[] { item.Reward + configuration.Gamma * Predict(item.State).Value })).ToArray();
trainer.Train(nextQ, configuration.EpochesPerIteration, configuration.BatchSize);
}
public TicTacToeValueNetwork Train(int epoches, params TrainingMonitor[] monitors)
{
var network = new TicTacToeValueNetwork(HiddenLayerSizes, ActivationFunction, new Random(Seed));
var optimizer = new SGDMomentum(network.Network, LearingRate, Momentum);
var trainer = new Trainer(optimizer, new Random(Seed));
trainer.Monitors.AddRange(monitors);
trainer.Train(TrainingData, epoches, 16);
return(network);
}
private void backgroundWorker1_DoWork(object sender, DoWorkEventArgs e)
{
bitmap = new Bitmap(200, 200);
Random random = new Random(0);
Optimizer optimizer;
Network network;
if (ActivationFunction == Function.ReLU)
{
//var network = NetworkBuilder.Build(Function.ReLU, 2, 3, 64, 64, 32, 32);
//evaluator = new NetworkEvaluator(network);
//optimizer = new SGDMomentum(evaluator, 0.001, 0.04);
network = NetworkBuilder.Build(new NetworkDefinition(FunctionName.ReLU, 2, 3, 72, 72, 72, 36, 36, 36, 18, 18), new He(0));
optimizer = new SGDMomentum(network, 0.001, 0.008);
}
else
{
network = NetworkBuilder.Build(new NetworkDefinition(FunctionName.Sigmoidal, 2, 3, 32, 8), new He(0));
optimizer = new SGDMomentum(network, 0.1, 0.8);
}
var trainer = new Trainer(optimizer, new Random(0));
var mseMonitor = new MeanSquareErrorMonitor();
trainer.Monitors.Add(mseMonitor);
int width;
int height;
lock (bitmap)
{
width = bitmap.Width;
height = bitmap.Height;
}
var imagePointList = new List <double[]>();
for (int y = 0; y < height; y++)
{
for (int x = 0; x < width; x++)
{
imagePointList.Add(new double[] { (double)x / width, (double)y / height });
}
}
trainingData = GenerateTrainingData(random);
for (int epoch = 0; epoch < 15000; epoch++)
{
//if (epoch % 300 == 0)
//{
// GenerateTrainingData(random);
//}
if (epoch % 100 == 0)
{
RefreshBitmap(network, width, height, imagePointList);
}
trainer.Train(trainingData, 1, 1);
Console.WriteLine(mseMonitor.CollectedData.Last());
}
}
