} //buildModel
protected void InitializeSOMWeights(ModelSelfOrganizingMap model,
long noOfWeights)
{
Parallel.For(0, _xdim,
new ParallelOptions {
MaxDegreeOfParallelism = _maxParallelThreads
},
xdim =>
{
Random rnd = new Random();
model.SomMap[xdim] = new SingleSOMNode[_ydim];
//computedDistValues[xdim] = new double[_ydim];
for (int ydim = 0; ydim < _ydim; ydim++)
{
model.SomMap[xdim][ydim] = new SingleSOMNode(noOfWeights, xdim, ydim);
model.SomMap[xdim][ydim].InitWeights(rnd, _weightBaseValue);
}
});
}
BuildModel(double[][] trainingData,
string[] attributeHeaders,
int indexTargetAttribute = -1)
{
if (indexTargetAttribute == -1)
{
indexTargetAttribute = attributeHeaders.Length - 1;
}
//Verify data and set variables
VerifyData(trainingData, attributeHeaders, indexTargetAttribute);
ModelSelfOrganizingMap model = new ModelSelfOrganizingMap(_missingValue,
indexTargetAttribute,
trainingData.Length);
model.SomMap = new SingleSOMNode[_xdim][];
long noOfWeights = trainingData.Length; //Equal to number of features
//Choose as lower of dimension as distance diamter
long distanceDiameter = (_xdim < _ydim) ? _xdim : _ydim;
double origRadius;
if (_origRadius == double.MaxValue) //Not set by user
{
origRadius = distanceDiameter / 2.0; //Radius of influence
//Make it circluar so that Maps in immediate grid can be accomodated
//origRadius = sqrt(x*x+y*y), x=y=oriRadiud
origRadius = origRadius * Math.Sqrt(2.0);
}
else
{
origRadius = _origRadius;
}
double[][] computedDistValues = new double[_xdim][];
//Initialize SOM Weights
InitializeSOMWeights(model, noOfWeights);
double learningRate = _learningRate;
double radius = origRadius;
double timeConstant = (double)_maxEpoch / Math.Log(origRadius);
//Start Convergence Loops
for (int epoch = 0; epoch < _maxEpoch; epoch++) //Do not parallelize
{
for (int dataRow = 0; dataRow < trainingData[0].Length; dataRow++)
{
double[] lowestDistance = new double[_xdim];
long[] lowestDistanceY = new long[_xdim];
double[] inputDataRow = GetLinearArray(_trainingData, dataRow, _trainingData.Length - 1);
//Find Best Matching Unit (BMU) for training row
//Find lowest distance in xdim
Parallel.For(0, _xdim,
new ParallelOptions {
MaxDegreeOfParallelism = _maxParallelThreads
}, xdim =>
//for (long xdim = 0; xdim < _xdim; xdim++) //Parallelize this loop
{
double computedDistValue = double.MaxValue;
lowestDistance[xdim] = double.MaxValue;
for (long ydim = 0; ydim < _ydim; ydim++)
{
//computedDistValues[xdim][ydim] =
computedDistValue =
model.SomMap[xdim][ydim].GetComputedDistance(inputDataRow);
if (computedDistValue < lowestDistance[xdim])
{
lowestDistance[xdim] = computedDistValue;
lowestDistanceY[xdim] = ydim;
}
}//ydim
});//xdim
long bmuX = 0, bmuY = 0;
double finalDistanceValue = double.MaxValue;
for (int xdim = 0; xdim < _xdim; xdim++) //Parallelize this loop
{
if (lowestDistance[xdim] < finalDistanceValue)
{
bmuX = xdim;
bmuY = lowestDistanceY[xdim];
finalDistanceValue = lowestDistance[xdim];
}
}
//Update Weights in neighborhood accounting for distance from BMU
for (long xdim = 0; xdim < _xdim; xdim++) //Parallelize this loop
{
for (long ydim = 0; ydim < _ydim; ydim++)
{
model.SomMap[xdim][ydim].UpdateWeights(model.SomMap[bmuX][bmuY],
inputDataRow, radius, learningRate, _useDistanceInfluence == 1?true:false);
}
}
} //row
//Update learning rate and distance
//Formula from: http://www.ai-junkie.com/ann/som/som3.html
radius = origRadius * Math.Exp(-(double)epoch / timeConstant); // Math.Pow(origRadius,((double)epoch/(double)_maxEpoch)); //;
learningRate = _learningRate * Math.Exp(-(double)epoch / timeConstant);
}
return(model);
} //buildModel
