public void CanTransferWeightsInsideNetwrok()
{
var o1 = new RMSPropOptimizer <float>(1e-3f);
var n1 = new LayeredNet <float>(1, 1,
new GruLayer <float>(2, 3),
new LinearLayer <float>(3, 2),
new SoftMaxLayer <float>(2))
{
Optimizer = o1
};
var n2 = (LayeredNet <float>)n1.Clone();
var o2 = new RMSPropOptimizer <float>(1e-3f);
n2.Optimizer = o2;
n1.UseGpu();
TrainXor(n1);
n1.TransferStateToHost();
TrainXor(n2);
var w1 = n1.Weights.ToList();
var w2 = n2.Weights.ToList();
for (int i = 0; i < w1.Count; i++)
{
w1[i].Weight.ShouldMatrixEqualWithinError(w2[i].Weight);
w1[i].Gradient.ShouldMatrixEqualWithinError(w2[i].Gradient);
w1[i].Cache1.ShouldMatrixEqualWithinError(w2[i].Cache1);
w1[i].Cache2.ShouldMatrixEqualWithinError(w2[i].Cache2);
w1[i].CacheM.ShouldMatrixEqualWithinError(w2[i].CacheM);
}
}
public void Mnist([DefaultValue(true)] bool gui)
{
const int batchSize = 128;
const int hSize = 20;
MklProvider.TryUseMkl(true, ConsoleProgressWriter.Instance);
string dataDir = Path.Combine(Path.GetTempPath(), "Retia_datasets", "MNIST");
DownloadDataset(dataDir);
Console.WriteLine("Loading training set");
var trainSet = LoadTrainingSet(dataDir);
trainSet.BatchSize = batchSize;
var network = new LayeredNet <float>(batchSize, 1,
new AffineLayer <float>(trainSet.InputSize, hSize, AffineActivation.Sigmoid),
new LinearLayer <float>(hSize, trainSet.TargetSize),
new SoftMaxLayer <float>(trainSet.TargetSize));
var optimizer = new AdamOptimizer <float>();
network.Optimizer = optimizer;
var trainer = new OptimizingTrainer <float>(network, optimizer, trainSet,
new OptimizingTrainerOptions(1)
{
ErrorFilterSize = 100,
MaxEpoch = 1,
ProgressWriter = ConsoleProgressWriter.Instance,
ReportProgress = new EachIteration(100),
ReportMesages = true
}, new OptimizingSession("MNIST"));
RetiaGui retiaGui;
if (gui)
{
retiaGui = new RetiaGui();
retiaGui.RunAsync(() => new TrainingWindow(new TypedTrainingModel <float>(trainer)));
}
var runner = ConsoleRunner.Create(trainer, network);
runner.Run();
}
private void TrainXor(LayeredNet <float> net)
{
var trainer = new OptimizingTrainer <float>(net, net.Optimizer, new XorDataset(true), new OptimizingTrainerOptions(1)
{
ErrorFilterSize = 0,
ReportProgress = ActionSchedule.Disabled,
ReportMesages = false
}, new OptimizingSession(false));
var cts = new CancellationTokenSource();
trainer.SequenceTrained += s =>
{
cts.Cancel();
};
trainer.Train(cts.Token).Wait();
}
public void Xor()
{
MklProvider.TryUseMkl(true, ConsoleProgressWriter.Instance);
var optimizer = new RMSPropOptimizer <float>(1e-3f);
var net = new LayeredNet <float>(1, 1, new AffineLayer <float>(2, 3, AffineActivation.Tanh), new AffineLayer <float>(3, 1, AffineActivation.Tanh)
{
ErrorFunction = new MeanSquareError <float>()
})
{
Optimizer = optimizer
};
var trainer = new OptimizingTrainer <float>(net, optimizer, new XorDataset(true), new OptimizingTrainerOptions(1)
{
ErrorFilterSize = 0,
ReportProgress = new EachIteration(1),
ReportMesages = true,
ProgressWriter = ConsoleProgressWriter.Instance,
LearningRateScaler = new ProportionalLearningRateScaler(new EachIteration(1), 9e-5f)
}, new OptimizingSession("XOR"));
var runner = ConsoleRunner.Create(trainer, net);
trainer.TrainReport += (sender, args) =>
{
if (args.Errors.Last().RawError < 1e-7f)
{
runner.Stop();
Console.WriteLine("Finished training.");
}
};
var gui = new RetiaGui();
gui.RunAsync(() => new TrainingWindow(new TypedTrainingModel <float>(trainer)));
runner.Run();
}
private static LayeredNet <float> CreateNetwork(string fileName)
{
return(LayeredNet <float> .Load(fileName));
}
