public GPOut Compute(Matrix <double> mTr, Matrix <double> mTeTr, Vector <double> vTrY)
{
var Mtr = (mTr).RowCount;
var gpOut = new GPOut();
var Kte = _kernel.Main(DistanceMatrix.Instance.Matrix, _length);
if (Mtr > 0)
{
var Kxx_p_noise = _kernel.Main(mTr, _length);
for (int i = 0; i < Mtr; i++)
{
Kxx_p_noise[i, i] += _noise;
}
var svd1 = Kxx_p_noise.Svd();
for (int i = 0; i < Mtr; i++)
{
svd1.S[i] = svd1.S[i] > Double.Epsilon ? 1.0 / svd1.S[i] : 0;
}
var tmp = _kernel.Main(mTeTr, _length) * svd1.U;
gpOut.mu = tmp * (svd1.S.PointwiseMultiply(svd1.U.Transpose() * (vTrY)));
var cov = tmp * Matrix <double> .Build.DenseOfDiagonalVector(svd1.S.PointwiseSqrt());
cov = cov * cov.Transpose();
cov = Kte /*.SubMatrix(0,100,0,100) */ - cov;
var svd2 = cov.Svd();
for (int i = 0; i < Mte; i++)
{
if (svd2.S[i] < Double.Epsilon)
{
svd2.S[i] = 0.0;
}
}
gpOut.proj = svd2.U * Matrix <double> .Build.DenseOfDiagonalVector(svd2.S.PointwiseSqrt());
gpOut.sd95 = 1.98 * (gpOut.proj * gpOut.proj.Transpose()).Diagonal().PointwiseSqrt();
}
else
{
gpOut.sd95 = 1.98 * (Kte.Diagonal()).PointwiseSqrt();
var svd = Kte.Svd();
gpOut.proj = svd.U * Matrix <double> .Build.DenseOfDiagonalVector(svd.S.PointwiseSqrt());
gpOut.mu = Vector <double> .Build.DenseOfArray(Enumerable.Range(0, Mte).Select(_ => 0d).ToArray());
}
return(gpOut);
}
public Svd <double> GetDefaultSvd()
{
Matrix <double> dmTr = MathHelper.ComputeDistanceMatrix(new[] { 0d }, new[] { 0d });
Vector <double> v = Vector <double> .Build.DenseOfArray(new[] { 0d });
var Kxx_p_noise = kernel.Main(dmTr, length);
for (int i = 0; i < 1; i++)
{
Kxx_p_noise[i, i] += noise;
}
return(Kxx_p_noise.Svd());
}
public GP(IMatrixkernel kernel, double length, double noise)
{
this._kernel = kernel;
this._length = length;
this._noise = noise;
this._Kte = kernel.Main(DistanceMatrix.Instance.Matrix, length);
}
public GPOut Predict(double[] trainingPointsX, double[] trainingPointsY, params double[] TestPointsX)
{
if (TestPointsX.Count() == 1)
{
Kte = kernel.Main(Matrix <double> .Build.Dense(1, 1), length);
}
else
{
Kte = kernel.Main(DistanceMatrix.Instance.Matrix.SubMatrix(0, TestPointsX.Count(), 0, TestPointsX.Count()), length);
}
var svd1 = Task.Run(() =>
trainingPointsX.Length > 0 ? Update(trainingPointsX, trainingPointsY) : GetDefaultSvd());
var gpOut = new GPOut();
Svd <double> svd;
(gpOut.mu, svd) = Task.Run(() =>
{
if (trainingPointsX.Count() > 0)
{
return(Evaluate(TestPointsX, trainingPointsX.ToArray(), trainingPointsY.ToArray(), kernel, length, svd1.Result, Kte));
}
else
{
return(Vector <double> .Build.DenseOfArray(Enumerable.Range(0, TestPointsX.Count()).Select(_ => 0d).ToArray()), Kte.Svd());
//gpOut.sd95 = 1.98 * (Kte.Diagonal()).PointwiseSqrt();
}
}).Result;
gpOut.proj = svd.U * Matrix <double> .Build.DenseOfDiagonalVector(svd.S.PointwiseSqrt());
gpOut.sd95 = 1.98 * (gpOut.proj * gpOut.proj.Transpose()).Diagonal().PointwiseSqrt();
gpOut.X = TestPointsX;
return(gpOut);
}
public static (Vector <double>, Svd <double>) Evaluate(double[] testPointsX, double[] trainingPointsX, double[] trainingPointsY, IMatrixkernel kernel, double length, Svd <double> svd1, Matrix <double> Kte)
{
var gpOut = new GPOut();
Svd <double> svd;
var dmTeTr = MathHelper.ComputeDistanceMatrix(testPointsX, trainingPointsX.ToArray());
var tmp = kernel.Main(dmTeTr, length) * svd1.U;
Vector <double> v = Vector <double> .Build.DenseOfEnumerable(trainingPointsY);
var mu = tmp * (svd1.S.PointwiseMultiply(svd1.U.Transpose() * v));
var cov = tmp * Matrix <double> .Build.DenseOfDiagonalVector(svd1.S.PointwiseSqrt());
cov = cov * cov.Transpose();
cov = Kte - cov;
svd = cov.Svd();
for (int i = 0; i < svd.S.Count; i++)
{
svd.S[i] = svd.S[i] < Double.Epsilon ? 0 : svd.S[i];
}
return(mu, svd);
}
