private IGradientBasedOptimizer createOptimizer(TrainSettings trainSettings)
{
switch (trainSettings.OptimizationAlgorithm)
{
case OptimizationAlgorithm.BasicGradientDescent:
return(new SteepestDescentBasicOptmizer(trainSettings.LearningRate,
trainSettings.MomentumStartPerc / 100.0, trainSettings.MomentumEndPerc / 100.0,
trainSettings.MaxIteratinosPerBatch));
case OptimizationAlgorithm.NesterovSutskeverGradientDescent:
return(new SteepestDescentAdvancedOptmizer(trainSettings.LearningRate,
trainSettings.MomentumStartPerc / 100.0, trainSettings.MomentumEndPerc / 100.0,
trainSettings.MaxIteratinosPerBatch));
case OptimizationAlgorithm.RpropPlus:
return(new RpropPlusOptmizer(trainSettings.RpropInitStep, trainSettings.RpropMaxStep,
trainSettings.RpropStepUpMult, trainSettings.RpropStepDownMult,
trainSettings.MaxIteratinosPerBatch));
case OptimizationAlgorithm.ImprovedRpropMinus:
default:
return(new ImprovedRpropMinusOptmizer(trainSettings.RpropInitStep, trainSettings.RpropMaxStep,
trainSettings.RpropStepUpMult, trainSettings.RpropStepDownMult,
trainSettings.MaxIteratinosPerBatch));
}
}
public bool SaveTrainSettings(string netId, TrainSettings settings)
{
string baseDir = getSafeBaseDirPath(netId);
if (baseDir == null)
{
return(false);
}
try {
using (var stream = File.Create(Path.Combine(baseDir, TRAIN_SETTINGS_FILE_NAME))) {
trainSettingsSerializer.Serialize(stream, settings);
return(true);
}
}
catch (Exception ex) {
return(false);
}
}
public void TrainNetwork()
{
stopwatch.Reset();
stopwatch.Start();
sendMessage("Loading training settings.");
trainSettings = storageManager.LoadTrainSettings(netId);
if (trainSettings == null)
{
sendMessage("Train settings file was not found.");
return;
}
using (var db = MausrDb.Create()) {
sendMessage("Initializing learning environment.");
trainData = new TrainData();
int inputSize = trainSettings.InputImgSizePx * trainSettings.InputImgSizePx;
int outputSize = db.Symbols.Count();
var layout = new NetLayout(inputSize, trainSettings.HiddenLayersSizes, outputSize);
network = new Net(layout, new SigomidActivationFunc(), new NetCostFunction());
network.SetOutputMap(db.Symbols.Select(s => s.SymbolId).ToArray());
netEvaluator = new NetEvaluator(network);
var optimizer = createOptimizer(trainSettings);
var trainer = new NetTrainer(network, optimizer, trainSettings.RegularizationLambda);
unpackedCoefs = network.Layout.AllocateCoefMatrices();
sendMessage("Preparing inputs and outputs.");
prepareInOut(db);
sendMessage(string.Format("Learning of {0} samples started.", trainInputs.RowCount));
bool converged = trainer.TrainBatch(trainInputs, trainSettings.BatchSize, trainSettings.LearnRounds,
trainOutIds, trainSettings.InitSeed, trainSettings.MinDerivCompMaxMagn, trainIterationCallback, job.CancellationToken);
if (job.CancellationToken.IsCancellationRequested)
{
sendMessage("Training {0}.", job.Canceled ? "canceled" : "stopped");
}
else
{
sendMessage("Training done ({0}converged).", converged ? "" : "not ");
}
if (!job.Canceled)
{
sendMessage("Saving trained data.");
if (!storageManager.SaveNet(netId, network))
{
sendMessage("Failed to save the network.");
}
if (!storageManager.SaveTrainData(netId, trainData))
{
sendMessage("Failed to save training data.");
}
sendMessage("Saving results visualization.");
if (!createAndSaveResultsVis(3, 0.01, 40))
{
sendMessage("Failed visualize results.");
}
}
stopwatch.Stop();
}
sendMessage("All done.");
}
