public UsersController(MausrDb db)
{
this.db = db;
}
public SymbolDrawingsController(MausrDb db, AppSettingsProvider appSettingsProvider)
{
this.db = db;
this.appSettingsProvider = appSettingsProvider;
}
public SymbolsController(MausrDb db)
{
this.db = db;
}
public TrainController(MausrDb db, TrainStorageManager trainStorageManager, CurrentEvaluator evaluator)
{
this.db = db;
this.trainStorageManager = trainStorageManager;
this.evaluator = evaluator;
}
public TeachController(MausrDb db, CurrentEvaluator evaluator)
{
this.db = db;
this.evaluator = evaluator;
}
public HomeController(MausrDb db, CurrentEvaluator evaluator, ICaptcha captcha)
{
this.db = db;
this.evaluator = evaluator;
this.captcha = captcha;
}
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.");
}
private void prepareInOut(MausrDb db)
{
var ic = new NetInputConvertor();
var dbInputs = db.SymbolDrawings.Where(x => x.Approved == true).ToList();
var baseInputsAndDrawings = dbInputs
.Select(sd => new { SymbolId = sd.Symbol.SymbolId, RawDrawing = sd.GetRawDrawing() })
.ToList();
var inputsAndDrawings = baseInputsAndDrawings.ToList();
var dataProc = new RawDataProcessor();
if (trainSettings.GenerateExtraInputsByRotation != 0)
{
inputsAndDrawings.AddRange(baseInputsAndDrawings
.Select(x => new {
SymbolId = x.SymbolId,
RawDrawing = dataProc.Rotate(x.RawDrawing, trainSettings.GenerateExtraInputsByRotation)
}));
inputsAndDrawings.AddRange(baseInputsAndDrawings
.Select(x => new {
SymbolId = x.SymbolId,
RawDrawing = dataProc.Rotate(x.RawDrawing, -trainSettings.GenerateExtraInputsByRotation)
}));
}
if (trainSettings.NormalizeInput)
{
foreach (var iad in inputsAndDrawings)
{
dataProc.NormalizeInPlace(iad.RawDrawing);
}
}
ic.Shuffle(inputsAndDrawings);
var rasterizer = new Rasterizer()
{
ImageSize = trainSettings.InputImgSizePx,
DrawPoints = false,
ExtraMargin = true,
PenSizePerc = trainSettings.PenThicknessPerc / 100f
};
int testSamplesCount = 1 + (trainSettings.TestDataSetSizePerc * inputsAndDrawings.Count / 100);
{
var testSet = inputsAndDrawings.Take(testSamplesCount).ToList();
var testDrawings = testSet.Select(x => x.RawDrawing).ToList();
testInputs = ic.CreateInputsMatrix(testDrawings, rasterizer, false);
testOutIds = testSet.Select(x => x.SymbolId).ToArray();
testOutNeuronIndices = ic.CreateOutIndicesFromIds(testOutIds, network);
}
{
var trainSet = inputsAndDrawings.Skip(testSamplesCount).ToList();
var trainDrawings = trainSet.Select(x => x.RawDrawing).ToList();
trainInputs = ic.CreateInputsMatrix(trainDrawings, rasterizer, false);
trainOutIds = trainSet.Select(x => x.SymbolId).ToArray();
trainOutNeuronIndices = ic.CreateOutIndicesFromIds(trainOutIds, network);
}
}
