public new static Kernel <T> FromMatlabStruct(MatlabStruct s)
{
string className = s.GetString("className");
if (className.Equals(KEGaussian <T> .MATLAB_CLASS))
{
return(KEGaussian <T> .FromMatlabStruct(s));
}
else if (className.Equals(KGGaussian <T> .MATLAB_CLASS))
{
return(KGGaussian <T> .FromMatlabStruct(s));
}
else
{
String msg = String.Format("Unknown class: {0}", className);
throw new ArgumentException(msg);
}
}
/** - Generate a cell array of FeatureMap candidates from medf,
* a list of factors to be multiplied with the
* diagonal of the average covariance matrices.
*
* - subsamples can be used to limit the samples used to compute
* median distance.
* See RFGJointKGG in Matlab
*/
public override List <RandomFeatureMap> GenCandidates(
List <IKEPDist[]> msgs, int[] numFeatures, double[] medianFactors, Random rng)
{
if (numFeatures.Length != 2)
{
throw new ArgumentException("numFeatures must have length = 2");
}
List <DVectorNormal> joints = msgs.Select(m => ToJointGaussian(m)).ToList();
int n = joints.Count;
int subsamples = Math.Min(1500, n);
List <IKEPDist> jointsBase = joints.Cast <IKEPDist>().ToList();
Vector avgCov = RFGEProdMap.GetAverageDiagCovariance(jointsBase, subsamples, rng);
double[] embedWidth2 = avgCov.ToArray();
// generate candidates
double meanEmbedMedian2 = KGGaussian <DVectorNormal> .MedianPairwise(
joints, embedWidth2);
var mapCandidates = new List <RandomFeatureMap>(medianFactors.Length);
for (int i = 0; i < medianFactors.Length; i++)
{
double medf = medianFactors[i];
if (medf <= 0)
{
string text = String.Format("medf must be strictly positive. Found i={0}, medf[i]={1}",
i, medf);
throw new ArgumentException(text);
}
double width2 = meanEmbedMedian2 * medf;
mapCandidates.Add(new RFGJointKGG(embedWidth2, width2,
numFeatures[0], numFeatures[1]));
}
return(mapCandidates);
}
