private TrainingParameters GetParameters()
{
int epochs = Int32.Parse(TrainingEpochs.Text);
int iterations = Int32.Parse(TrainingIterations.Text);
TrainingParameters p = p = new KMeansTrainingParameters()
{
Epochs = epochs,
TrainingData = TrainingData
};
if (Network is SelfOrganizingMap)
{
int kmax = TrainingData.Count * epochs;
double minLearningRate = double.Parse(TrainingLearningRateMin.Text, CultureInfo.InvariantCulture);
double maxLearningRate = double.Parse(TrainingLearningRateMax.Text, CultureInfo.InvariantCulture);
p = new SomTrainingParameters()
{
LearningRate = new DeclineExponentially(kmax, minLearningRate, maxLearningRate),
TirednessMechanism = GetTiredMechanism(),
NeighbourhoodFunction = GetNeighbourhoodFunction(kmax),
NumberOfIterations = iterations,
TrainingData = TrainingData
};
}
return(p);
}
public override void Train(TrainingParameters parameters)
{
SomTrainingParameters p = (SomTrainingParameters)parameters;
try
{
parameters.Validate();
}
catch (TrainingParametersException e)
{
throw new SomLogicalException($"Improper training parameters. {e.Message}");
}
var trainingData = p.TrainingData;
var iterations = p.NumberOfIterations;
var neighborhoodFunction = p.NeighbourhoodFunction;
var learningRate = p.LearningRate;
var tiredness = p.TirednessMechanism;
if (trainingData.First().Count != Dimensions)
{
throw new SomLogicalException("Training data points dimensions doesn't match som dimensions.");
}
tiredness?.Initialize(Neurons);
Observer?.SaveState(Neurons);
int numberOfShuffles = (iterations / trainingData.Count) + 1;
for (int j = 0; j < numberOfShuffles; j++)
{
var shuffled = trainingData.Shuffle();
for (int i = 0; i < shuffled.Count; i++)
{
if (j * shuffled.Count + i == iterations)
{
return;
}
Dictionary <Neuron, double> neuronsDistances = new Dictionary <Neuron, double>();
for (int k = 0; k < Neurons.Count; k++)
{
//store neurons with their distances
neuronsDistances.Add(Neurons[k],
DistanceCalculator.CalculateDistance(shuffled[i], Neurons[k].CurrentWeights));
}
var potentialWinners = tiredness == null ? Neurons : tiredness.SelectPotentialWinners(Neurons);
Neuron winner = neuronsDistances.Where(pair => potentialWinners.Contains(pair.Key))
.OrderBy(pair => pair.Value).First().Key;
tiredness?.Update(winner, Neurons);
var neuronsNeighbourhoodCoefficients =
neighborhoodFunction.CalculateNeighborhoodValues(winner, neuronsDistances, DistanceCalculator,
j * trainingData.Count + i);
foreach (var neuron in Neurons)
{
var diffVec = shuffled[i].Subtract(neuron.CurrentWeights);
var deltaWeight = diffVec.Multiply(learningRate.GetValue(j * trainingData.Count + i) *
neuronsNeighbourhoodCoefficients[neuron]);
neuron.CurrentWeights = neuron.CurrentWeights + deltaWeight;
}
Observer?.SaveState(Neurons);
}
}
}
