public static DVectorNormal ToJointGaussian(params IKEPDist[] msgs)
{
// Stack all messages to form a big Gaussian distribution
Vector[] means = msgs.Select(dist => dist.GetMeanVector()).ToArray();
Matrix[] covs = msgs.Select(dist => dist.GetCovarianceMatrix()).ToArray();
Vector M = MatrixUtils.ConcatAll(means);
Matrix V = MatrixUtils.BlkDiag(covs);
DVectorNormal joint = new DVectorNormal(M, V);
return(joint);
}
public override double[] GenFeatures(params IKEPDist[] msgs)
{
DVectorNormal joint = ToJointGaussian(msgs);
InitMap(joint);
Vector innerFeature = eprodMap.MapToVector(joint);
Vector FWout = innerFeature * Wout;
double[] fea = new double[this.numFeatures];
double scale = Math.Sqrt(2.0 / numFeatures);
for (int i = 0; i < fea.Length; i++)
{
fea[i] = Math.Cos(FWout[i] + Bout[i]) * scale;
}
return(fea);
}
public void InitMap(DVectorNormal joint)
{
// dynamically initialize the random feature map
if (Wout == null)
{
Debug.Assert(Bout == null);
int totalDim = joint.GetDimension();
if (totalDim != flattenEmbedWidth2s.Length)
{
throw new SystemException("Expect total dim. of the joint to be = length of params.");
}
eprodMap = new RFGEProdMap(flattenEmbedWidth2s, innerNumFeatures);
Wout = MatrixUtils.Randn(innerNumFeatures, numFeatures) * (1.0 / Math.Sqrt(outer_width2));
double[] Bvec = MatrixUtils.UniformVector(0, 2.0 * Math.PI, numFeatures);
Bout = Vector.FromArray(Bvec);
// VectorGaussian.SampleFromMeanAndVariance();
}
}
