private Tuple<double[], double[], double[]> TrainModel(ModelSettings settings)
{
var data = GetConvNetSharpData();
_images = data.Item1;
_labels = data.Item2;
var confusionMatrix = new int[settings.NoEpochs, 4, 4];
var validationConfusionMatrix = new int[settings.NoEpochs, 4, 4];
var threshold = (int)Math.Floor(0.9*_images.Count);
for (var k = 0; k < settings.NoEpochs; k++)
{
for (var i = 0; i < threshold; i++)
{
var image = _images[i];
var vol = BuilderInstance.Volume.From(_labels[i], new Shape(1, 1, 4, 1));
try
{
_trainer.Train(BuilderInstance.Volume.From(image, new Shape(80, 60, 1)), vol);
}
catch(ArgumentException)
{
}
var prediction = _trainer.Net.GetPrediction()[0];
confusionMatrix[k, LabelFromOneHot(_labels[i]), prediction]++;
}
for (var i = threshold; i < _images.Count; i++)
{
var image = _images[i];
_trainer.Net.Forward(BuilderInstance.Volume.From(image, new Shape(80, 60, 1)));
var prediction = _trainer.Net.GetPrediction()[0];
validationConfusionMatrix[k, LabelFromOneHot(_labels[i]), prediction]++;
}
}
return GetEpochsAndAccuracies(confusionMatrix, settings.NoEpochs, validationConfusionMatrix, threshold);
}
public Tuple<double[], double[], double[]> Train(ModelSettings settings)
{
_framework = settings.Framework;
if (_framework == "keras")
return TrainKerasModel(settings);
else
return TrainConvNetSharpModel(settings);
}
private Tuple<double[], double[], double[]> TrainKerasModel(ModelSettings settings)
{
var data = GetData();
var xTrain = data.Item1.Take((int)Math.Floor(0.8 * data.Item1.Count));
var yTrain = data.Item2.Take((int)Math.Floor(0.8 * data.Item2.Count));
var xVal = data.Item1.TakeLast(data.Item1.Count - (int)Math.Floor(0.8 * data.Item1.Count));
var yVal = data.Item2.TakeLast(data.Item2.Count - (int)Math.Floor(0.8 * data.Item2.Count));
var xValidation = xVal.Take(xVal.Count() / 2);
var yValidation = yVal.Take(yVal.Count() / 2);
var history = _model.Fit(np.array(xTrain), np.array(yTrain), epochs: settings.NoEpochs, validation_data: new NDarray[]
{ np.array(xValidation), np.array(yValidation)});
var accuracies = history.HistoryLogs["accuracy"];
var valAccuracies = history.HistoryLogs["val_accuracy"];
var epochs = new double[history.Epoch.Length];
for (var i = 0; i < history.Epoch.Length; i++)
epochs[i] = history.Epoch[i] + 1;
return new Tuple<double[], double[], double[]>(epochs, accuracies, valAccuracies);
}
private Tuple<double[], double[], double[]> TrainConvNetSharpModel(ModelSettings settings)
{
EditTrainerSettings(settings);
return TrainModel(settings);
}
private void EditTrainerSettings(ModelSettings settings)
{
_trainer.LearningRate = settings.LearningRate;
_trainer.Momentum = 0.34;
}
