public double GetSDNLL_MDF(ILArray <double> mdfVector)
{
double firstPart = 0;
double secondPart = 0;
double thirdPart = 0;
if (this.items.Count < 2)
{
this.covarianceMatrixMDF = this.GetVarianceMatrix_MDF();
}
int q = this.meanMDF.Length;
// get matrix W
ILArray <double> W = this.GetMatrixW_MDF();
ILArray <double> WInverse = this.GetInverseMatrixOfMatrix(W, q);
/*if (Double.IsNaN(WInverse.ToArray()[0]))
* {
* //throw new InvalidDataException("inverse of covariance matrix MDF is NaN");
* WInverse = ILMath.eye(this.parent.VarianceMDF.Length, this.parent.VarianceMDF.Length);
* Debug.Assert(true, "Problem with covariance gaussian matrix.");
* }*/
ILArray <double> vector1 = mdfVector - this.meanMDF;
ILArray <double> tmpArray = ILMath.multiply(WInverse, vector1);
firstPart = 0.5 * ILMath.multiply(vector1.T, tmpArray).ToArray()[0];
secondPart = q / 2 * ILMath.log(2 * Math.PI).ToArray()[0];
thirdPart = 0.5 * ILMath.log(ILMath.det(W)).ToArray()[0];
return(firstPart + secondPart + thirdPart);
}
private double calcVarianceFunctionIL(double[] xi, double[] xj, ILArray <double> dispersion)
{
ILArray <double> xiIL = xi;
ILArray <double> xjIL = xj;
return((double)ILMath.multiply(ILMath.multiply(xiIL.T, dispersion), xjIL));
}
private double calcVarianceForRun(double[] run, double[,] candidateMatrix)
{
ILArray <double> vector = run;
ILArray <double> matrix = candidateMatrix;
ILArray <double> inverse = ILMath.pinv(ILMath.multiply(matrix, matrix.T));
return((double)ILMath.multiply(ILMath.multiply(vector.T, inverse), vector));
}
/// <summary>
/// Multidimensional scaling/PCoA: transform distances to points in a coordinate system.
/// </summary>
/// <param name="input">A matrix of pairwise distances. Zero indicates identical objects.</param>
/// <returns>A matrix, the columns of which are coordinates in the nth dimension.
/// The rows are in the same order as the input.</returns>
public static ILArray <double> Scale(ILArray <double> input)
{
int n = input.Length;
ILArray <double> p = ILMath.eye <double>(n, n) - ILMath.repmat(1.0 / n, n, n);
ILArray <double> a = -.5 * ILMath.multiplyElem(input, input);
ILArray <double> b = ILMath.multiply(p, a, p);
ILArray <complex> V = ILMath.empty <complex>();
ILArray <complex> E = ILMath.eig((b + b.T) / 2, V);
ILArray <int> i = ILMath.empty <int>();
ILArray <double> e = ILMath.sort(ILMath.diag(ILMath.real(E)), i);
e = ILMath.flipud(e);
i = ILMath.toint32(ILMath.flipud(ILMath.todouble(i)));
ILArray <int> keep = ILMath.empty <int>();
for (int j = 0; j < e.Length; j++)
{
if (e[j] > 0.000000001)
{
keep.SetValue(j, keep.Length);
}
}
ILArray <double> Y;
if (ILMath.isempty(keep))
{
Y = ILMath.zeros(n, 1);
}
else
{
Y = ILMath.zeros <double>(V.S[0], keep.Length);
for (int j = 0; j < keep.Length; j++)
{
Y[ILMath.full, j] = ILMath.todouble(-V[ILMath.full, i[keep[j]]]);
}
Y = ILMath.multiply(Y, ILMath.diag(ILMath.sqrt(e[keep])));
}
ILArray <int> maxind = ILMath.empty <int>();
ILMath.max(ILMath.abs(Y), maxind, 0);
int d = Y.S[1];
ILArray <int> indices = maxind + ILMath.toint32(ILMath.array <int>(SteppedRange(0, n, (d - 1) * n)));
ILArray <double> colsign = ILMath.sign(Y[indices]);
for (int j = 0; j < Y.S[1]; j++)
{
Y[ILMath.full, j] = Y[ILMath.full, j] * colsign[j];
}
return(Y);
}
private double calcVarianceFunction(double[] xi, double[] xj, double[,] matrix)
{
ILArray <double> xiIL = xi;
ILArray <double> xjIL = xj;
ILArray <double> matrixIL = matrix;
ILArray <double> inverse = ILMath.pinv(ILMath.multiply(matrixIL, matrixIL.T));
return((double)ILMath.multiply(ILMath.multiply(xiIL.T, inverse), xjIL));
}
private double calcVarianceFunction(double[] xi, double[] xj, double[,] matrix)
{
ILArray <double> xiIL = xi;
ILArray <double> xjIL = xj;
ILArray <double> matrixIL = matrix;
ILArray <double> inverse = ((double[, ])MachineLearning.Learning.Regression.FeatureSubsetSelection.toSystemMatrix <double>((ILMath.multiply(matrixIL, matrixIL.T)))).PseudoInverse();
return((double)ILMath.multiply(ILMath.multiply(xiIL.T, inverse), xjIL));
}
public double GetMDFDistance(ILArray <double> vector)
{
ILArray <double> delta = this.valuesMDF - vector;
double result = ILMath.multiply(delta.T, delta).ToArray()[0];
if (result == 0)
{
return(0);
}
return(Math.Sqrt(result));
}
private double calcVarianceForRun(double[] run, double[,] candidateMatrix)
{
ILArray <double> vector = run;
ILArray <double> matrix = candidateMatrix;
ILArray <double> matrixTimes = ILMath.multiply(matrix, matrix.T);
ILArray <double> inverse = ((double[, ])MachineLearning.Learning.Regression.FeatureSubsetSelection.toSystemMatrix <double>(matrixTimes)).PseudoInverse();
return((double)ILMath.multiply(ILMath.multiply(vector.T, inverse), vector));
}
/// <summary>
/// Calculates the eigenvalues and the eigenfaces.
/// </summary>
private void CalculateEigenValues()
{
//25 = magicnumber, just needed to initialize double array
this._eigenValues = new double[25];
_s.Diagonal.ExportValues(ref this._eigenValues);
var eigenValuesPow = new double[_eigenValues.Length];
for (var i = 0; i < eigenValuesPow.Length; i++)
{
eigenValuesPow[i] = (double)Math.Pow((double)_eigenValues[i], -0.5);
}
this._eigenVectors.Clear();
for (var i = 0; i < eigenValuesPow.Length; i++)
{
var oneEigenVector = (ILArray <double>)_v.Subarray(new string[]
{
":",
i.ToString()
});
var oneFace = ILMath.multiply(_a, oneEigenVector);
var oneFace2 = ILMath.multiply(oneFace, eigenValuesPow[i]);
//25 = magicnumber, just needed to initialize double array
var eigenVectorArray = new double[25];
oneFace2.ExportValues(ref eigenVectorArray);
var distance = 0.0;
for (var j = 0; j < eigenVectorArray.Length; j++)
{
distance += Math.Pow((double)eigenVectorArray[j], 2.0);
}
distance = Math.Sqrt(distance);
for (var j = 0; j < eigenVectorArray.Length; j++)
{
eigenVectorArray[j] /= (float)distance;
}
this._eigenVectors.Add(eigenVectorArray);
}
this._eigenWeights.Clear();
for (var i = 0; i < this._vectorSet.Count; i++)
{
var existingWeight = this.GetEigenWeight((byte[])this._vectorSet[i], this._eigenVectors.Count);
this._eigenWeights.Add(existingWeight);
}
}
private bool fitModel(List <Feature> newModel)
{
ILArray <double> DM = createDataMatrix(newModel);
if (DM.Size.NumberOfElements == 0)
{
return(false);
}
// ILArray<double> DMT = DM.T;
ILArray <double> temparray = null;
double[,] fixSVDwithACCORD;
//var exp = toSystemMatrix<double>(DM.T);
// fixSVDwithACCORD = (double[,])exp;
fixSVDwithACCORD = ((double[, ])toSystemMatrix <double>(DM.T)).PseudoInverse();
temparray = fixSVDwithACCORD;
ILArray <double> constants;
if (temparray.IsEmpty)
{
constants = ILMath.multiply(DM, Y_learning.T);
}
else
{
constants = ILMath.multiply(temparray, Y_learning.T);
}
double[] fittedConstant = constants.ToArray <double>();
for (int i = 0; i < constants.Length; i++)
{
newModel[i].Constant = fittedConstant[i];
//constants.GetValue(i);
}
//the fitting found no further influence
if (newModel[constants.Length - 1].Constant == 0)
{
return(false);
}
return(true);
}
public void UpdateCovarianceMatrixMDF_NonParametric(ILArray <double> vector)
{
#warning this must be remake according to F. Amnesic average with parameters t1, t2
// newCov = t-1/t * cov(t-1) + 1/t * (newVector - mean(t)) * (newVector - mean(t))T
// oldPart = t-1/t * cov(t-1)
// incrementalPart = 1/t * (newVector - mean(t)) * (newVector - mean(t))T
// vector1 = (newVector - mean(t))
// vector2 = (newVector - mean(t))T
// newCovPart = vector1 * vector2
//newVector - mean(t)
//Vector v1 = new Vector(vector.ToArray());
//v1.Subtract(this.meanMD);
ILArray <double> vector1 = (vector - this.meanMDF).ToArray();
ILArray <double> vector2 = vector1.ToArray();
// transpone
vector2 = vector2.T;
#warning need to edit update covariance matrix
double t = (double)this.items.Count;
double fragment1 = (t - 2) / (t - 1);
double fragment2 = t / (Math.Pow((t - 1), 2));
//DenseMatrix oldPart = fragment1 * this.covarianceMatrix;
//DenseMatrix newCovPart = vector1 * vector2;
try
{
//DenseMatrix incrementalPart = newCovPart * fragment2;
this.covarianceMatrixMDF = (this.covarianceMatrixMDF * fragment1) + (ILMath.multiply(vector1, vector2) * fragment2);
}
catch (Exception ee)
{
throw new InvalidCastException();
}
}
private double calcVarianceForRunIL(double[] run, ILArray <double> dispersion)
{
ILArray <double> vector = run;
return((double)ILMath.multiply(ILMath.multiply(vector.T, dispersion), vector));
}
/// <summary>
/// Creates the small Covariancematrix.
/// </summary>
private void CreateCovarianceMatrix()
{
_a = _columnVectorIntMatrix;
_covariance = ILMath.multiply(_a.T, _a);
}
private ILArray <double> calculateDispersion(double[,] matrix)
{
ILArray <double> tmp = matrix;
return(((double[, ])MachineLearning.Learning.Regression.FeatureSubsetSelection.toSystemMatrix <double>((ILMath.multiply(tmp, tmp.T)))).PseudoInverse());
}
public void Test_MatrixMultiply()
{
int errorCode = 0;
// success?
try {
ILPerformer p = new ILPerformer();
double[] data1 = new double[120];
double[] data2 = new double[120];
double[] results = new double[36] {
177840, 180120, 182400, 184680, 186960, 189240, 180120, 182440,
184760, 187080, 189400, 191720, 182400, 184760, 187120, 189480, 191840, 194200, 184680, 187080,
189480, 191880, 194280, 196680, 186960, 189400, 191840, 194280, 196720, 199160, 189240, 191720,
194200, 196680, 199160, 201640
};
for (int i = 0; i < 120; i++)
{
data1[i] = i;
data2[i] = i * 2;
}
ILDA A = new ILDA(data1, 6, 20);
ILDA BR = new ILDA(data2, 6, 20);
ILDA ResultExpect = new ILDA(results, 6, 6);
ILDA Result = null;
ILDA AR = A.T.T;
ILDA B = BR.T;
B.Detach();
BR = B.T.T;
errorCode = 0;
p.Tic();
Result = ILMath.multiply(A, B);
p.Toc();
errorCode = 1;
if (Result.Dimensions.NumberOfDimensions != 2)
{
throw new Exception("Wrong number of results dimensions!");
}
errorCode = 2;
if (Result.Dimensions[0] != 6 || Result.Dimensions[1] != 6)
{
throw new Exception("Wrong result's size!");
}
errorCode = 3;
if (!Result.Equals(ResultExpect))
{
throw new Exception("Wrong results values!");
}
Info("Test_MatrixMultiplyDouble succeeded: phy/phy " + p.ToString() + "ms needed");
errorCode = 4;
p.Tic();
if (!ResultExpect.Equals(ILMath.multiply(AR, BR)))
{
throw new Exception("Wrong data value on ref/ref");
}
p.Toc();
Info("Test_MatrixMultiplyDouble succeeded: ref/ref " + p.ToString() + "ms needed");
// test if wrong dimensions throw exceptions
try {
ILMath.multiply(A, AR);
throw new InvalidOperationException();
} catch (Exception e) {
if (e is InvalidOperationException)
{
throw new Exception("Dimensions should be forced to match!");
}
}
data1 = new double[1000 * 2000];
A = new ILDA(data1, 1000, 2000);
B = new ILDA(data1, 2000, 1000);
p.Tic();
Result = ILMath.multiply(A, B);
p.Toc();
Info("Test_MatrixMultiply 1000x2000: phy/phy " + p.ToString() + "ms needed");
Info("TODO: Implement test MatrixMultiply with reference storages!!");
errorCode = 5;
ILArray <float> fA = new ILArray <float>(new float[12] {
1f, 2f, 3f, 4f, 5f, 6f, 7f, 8f, 9f, 10f, 11f, 12f
}, 4, 3);
ILArray <float> fB = new ILArray <float>(new float[12] {
17f, 18f, 19f, 20f, 21f, 22f, 23f, 24f, 25f, 26f, 27f, 28f
}, 3, 4);
ILArray <float> fRes = new ILArray <float>(new float[16] {
278f, 332f, 386f, 440f, 323f, 386f, 449f, 512f, 368f, 440, 512f, 584f, 413f, 494f, 575f, 656f
}, 4, 4);
if (ILMath.norm(fRes - ILMath.multiply(fA, fB)) > 1.0e-10)
{
throw new Exception("Test_MatrixMultiply: invalid values detected!");
}
errorCode = 6;
ILArray <complex> zA = ILMath.real2complex(ILMath.vector(1, 12), ILMath.vector(-1, -1, -12));
ILArray <complex> zB = ILMath.real2complex(ILMath.vector(17, 28), ILMath.vector(-17, -1, -28));
//{new complex (0 , 0.5560), new complex (0, 0.6640), new complex (0, 0.7720), new complex (0, 0.8800), new complex (0, 0.6460), new complex (0, 0.7720), new complex (0, 0.8980), new complex (0, 1.0240), new complex (0, 0.7360), new complex (0, 0.8800),new complex(0,1.0240),new complex(0,1.1680),new complex(0,0.8260),new complex(0,0.9880),new complex(0,1.1500),new complex(0,1.3120)}
ILArray <complex> zRes = ILMath.real2complex(ILMath.zeros(1, 16), new ILArray <double>(new double[] { 556, 664, 772, 880, 646, 772, 898, 1024, 736, 880, 1024, 1168, 826, 988, 1150, 1312 }) * (-1));
zA = ILMath.reshape(zA, 4, 3);
zB = ILMath.reshape(zB, 3, 4);
zRes = ILMath.reshape(zRes, 4, 4);
if (ILMath.norm(zRes - ILMath.multiply(zA, zB)) > 1.0e-10)
{
throw new Exception("Test_MatrixMultiply: invalid values detected!");
}
errorCode = 7;
ILArray <fcomplex> cA = ILMath.real2fcomplex(ILMath.vector(1, 12), ILMath.vector(-1, -1, -12));
ILArray <fcomplex> cB = ILMath.real2fcomplex(ILMath.vector(17, 28), ILMath.vector(-17, -1, -28));
ILArray <fcomplex> cRes = ILMath.real2fcomplex(ILMath.zeros(1, 16), new ILArray <double>(new double[] { 556, 664, 772, 880, 646, 772, 898, 1024, 736, 880, 1024, 1168, 826, 988, 1150, 1312 }) * (-1));
cA = ILMath.reshape(cA, 4, 3);
cB = ILMath.reshape(cB, 3, 4);
cRes = ILMath.reshape(cRes, 4, 4);
if (ILMath.norm(cRes - ILMath.multiply(cA, cB)) > 1.0e-10)
{
throw new Exception("Test_MatrixMultiply: invalid values detected!");
}
Success("Test_MatrixMultiply successful.");
} catch (Exception e) {
Error("Test_MatrixMultiply failed at step: " + errorCode + " Msg: " + e.Message);
}
}
/// <summary>
/// Matrix multiplication
/// </summary>
/// <param name="star">'*'</param>
protected static ILArray <double> _m(ILArray <double> ilArray1, char star, ILArray <double> ilArray2)
{
return(ILMath.multiply(ilArray1, ilArray2));
}
public void CountMDF(ILArray <double> GSOmanifold, ILArray <double> C)
{
ILArray <double> scaterPart = this.values - C;
this.valuesMDF = ILMath.multiply(GSOmanifold.T, scaterPart);
}
public static double GetNormalisationNum(ILArray <double> vector)
{
double sum = ILMath.multiply(vector.T, vector).ToArray()[0];
return(Math.Sqrt(sum));
}
private ILArray <double> calculateDispersion(double[,] matrix)
{
ILArray <double> tmp = matrix;
return(ILMath.pinv(ILMath.multiply(tmp, tmp.T)));
}
public void UpdateCovarianceMatrixMDF(ILArray <double> vector)
{
// newCov = t-1/t * cov(t-1) + 1/t * (newVector - mean(t)) * (newVector - mean(t))T
// oldPart = t-1/t * cov(t-1)
// incrementalPart = 1/t * (newVector - mean(t)) * (newVector - mean(t))T
// vector1 = (newVector - mean(t))
// vector2 = (newVector - mean(t))T
// newCovPart = vector1 * vector2
//newVector - mean(t)
//Vector v1 = new Vector(vector.ToArray());
//v1.Subtract(this.meanMD);
ILArray <double> vector1 = (vector - this.meanMDF).ToArray();
ILArray <double> vector2 = vector1.ToArray();
// transpone
vector2 = vector2.T;
double t = (double)this.items.Count;
//double fragment1 = (t - 2) / (t - 1);
//double fragment2 = t / (Math.Pow((t - 1), 2));
double fragment1 = (t - 1 - this.GetAmnesicParameter(t)) / t;
double fragment2 = (1 + this.GetAmnesicParameter(t)) / t;
try
{
this.covarianceMatrixMDF = (this.covarianceMatrixMDF * fragment1) + (fragment2 * ILMath.multiply(vector1, vector2));
}
catch (Exception ee)
{
throw new InvalidCastException();
}
}