public override void InitialiseClusters()
{
epochCounter = 1;
iterator = 1;
Random Randy = new Random();
map = new KohonenNeuron[x, y];
for (int i = 0; i < x; i++)
{
for (int j = 0; j < y; j++)
{
//randomly initialise weights (in "Centroid" field) of each neuron
DenseVector InitialCentroid = new DenseVector(data[0].Item2.Count);
Parallel.For(0, InitialCentroid.Count, iter => {
InitialCentroid[iter] = Randy.NextDouble();
});
List <Tuple <string, DenseVector> > EmptyList = new List <Tuple <string, DenseVector> >();
map[i, j] = new KohonenNeuron(EmptyList, $"{i}, {j}", new Tuple <int, int>(i, j));
map[i, j].SetCentroid(InitialCentroid);
}
}
Tuple <string, DenseVector>[] ShuffleList = new Tuple <string, DenseVector> [data.Count];
data.CopyTo(ShuffleList);
shuffleList = ShuffleList.ToList();
}
public override void IterateOnce()
{
if (!stopped)
{
if (Decimal.Divide(Decimal.Divide(iterator, epochCounter), data.Count) == 1)
{
//maintaining separate list allows sampling without replacement at random throughout each epoch
//of training. This needs to be reset to match all data at start of each epoch.
// 1 epoch = 1 pass over all datapoints
epochCounter++;
Tuple <string, DenseVector>[] ShuffleList = new Tuple <string, DenseVector> [data.Count];
data.CopyTo(ShuffleList);
shuffleList = ShuffleList.ToList();
}
//randomly select datapoint with replacement
Random CurrentRand = new Random();
int PickIndex = CurrentRand.Next(shuffleList.Count);
Tuple <string, DenseVector> Picked = shuffleList[PickIndex];
shuffleList.RemoveAt(PickIndex);
//select most appropriate neuron by distance (distance squared to reduce overhead)
double BestDistance = HelperFunctions.GetEuclideanDistanceSquared(Picked.Item2, map[0, 0].GetCentroid());
Tuple <int, int> BestCoordinate = new Tuple <int, int>(0, 0);
KohonenNeuron Best = map[0, 0];
for (int i = 1; i < x; i++)
{
Parallel.For(0, y, j =>
{
double TestDistance = HelperFunctions.GetEuclideanDistanceSquared(Picked.Item2, map[i, j].GetCentroid());
if (BestDistance > TestDistance)
{
BestDistance = TestDistance;
BestCoordinate = new Tuple <int, int>(i, j);
Best = map[i, j];
}
});
}
//update data coordinates list
currentDataCoordinates[data.IndexOf(Picked)] = BestCoordinate;
//select neighbourhood and adjust neighbours' weights accordingly
//exponential decay formula ensures neighbourhood keeps shrinking
double CurrentRadius = radius * Math.Exp(-(double)Decimal.Divide(iterator, (decimal)timeConstant));
double CurrentLearningRate = learningRate * Math.Exp(-(double)Decimal.Divide(iterator, (decimal)timeConstant));
for (int i = 0; i < x; i++)
{
Parallel.For(0, y, j =>
{
double DistanceFromBest = HelperFunctions.GetEuclideanDistanceSquared(Best.Coordinates, map[i, j].Coordinates);
if (DistanceFromBest < Math.Pow(CurrentRadius, 2))
{
double NeighbourhoodWeighting = Math.Exp(-(double)Decimal.Divide((decimal)DistanceFromBest, (decimal)(2 * Math.Pow(CurrentRadius, 2))));
DenseVector CentroidReplacement = map[i, j].GetCentroid() + (NeighbourhoodWeighting * CurrentLearningRate) * (Picked.Item2 - map[i, j].GetCentroid());
map[i, j].SetCentroid(CentroidReplacement);
}
});
}
//housekeeping
iterator++;
stopped = StoppingConditionMet();
if (stopped)
{
//add all datapoints to clusters by coordinates
//add all neurons to cluster list
for (int i = 0; i < x; i++)
{
for (int j = 0; j < y; j++)
{
ConcurrentBag <Tuple <string, DenseVector> > FinalMemberList = new ConcurrentBag <Tuple <string, DenseVector> >();
Parallel.For(0, data.Count, datum =>
{
if (currentDataCoordinates[datum].Item1 == map[i, j].Coordinates[0] && currentDataCoordinates[datum].Item2 == map[i, j].Coordinates[1])
{
FinalMemberList.Add(data[datum]);
}
});
map[i, j].AddMembers(FinalMemberList.OrderBy(o => o.Item1).ToList());
clusters.Add(map[i, j]);
}
}
}
}
}
