/**
* Performs sanity checks on the input data and reshapes internal matrices. By reshaping
* a matrix it will only declare new memory when needed.
*/
public void configure(DMatrixRMaj initParam, DMatrixRMaj X, DMatrixRMaj Y)
{
if (Y.getNumRows() != X.getNumRows())
{
throw new ArgumentException("Different vector lengths");
}
else if (Y.getNumCols() != 1 || X.getNumCols() != 1)
{
throw new ArgumentException("Inputs must be a column vector");
}
int numParam = initParam.getNumElements();
int numPoints = Y.getNumRows();
if (param.getNumElements() != initParam.getNumElements())
{
// reshaping a matrix means that new memory is only declared when needed
this.param.reshape(numParam, 1, false);
this.d.reshape(numParam, 1, false);
this.H.reshape(numParam, numParam, false);
this.negDelta.reshape(numParam, 1, false);
this.tempParam.reshape(numParam, 1, false);
this.A.reshape(numParam, numParam, false);
}
param.set(initParam);
// reshaping a matrix means that new memory is only declared when needed
temp0.reshape(numPoints, 1, false);
temp1.reshape(numPoints, 1, false);
tempDH.reshape(numPoints, 1, false);
jacobian.reshape(numParam, numPoints, false);
}
/**
* Computes the d and H parameters. Where d is the average error gradient and
* H is an approximation of the hessian.
*/
private void computeDandH(DMatrixRMaj param, DMatrixRMaj x, DMatrixRMaj y)
{
func.compute(param, x, tempDH);
CommonOps_DDRM.subtractEquals(tempDH, y);
computeNumericalJacobian(param, x, jacobian);
int numParam = param.getNumElements();
int length = x.getNumElements();
// d = average{ (f(x_i;p) - y_i) * jacobian(:,i) }
for (int i = 0; i < numParam; i++)
{
double total = 0;
for (int j = 0; j < length; j++)
{
total += tempDH.get(j, 0) * jacobian.get(i, j);
}
d.set(i, 0, total / length);
}
// compute the approximation of the hessian
CommonOps_DDRM.multTransB(jacobian, jacobian, H);
CommonOps_DDRM.scale(1.0 / length, H);
}
public static void inv(DMatrixRMaj mat, DMatrixRMaj inv)
{
double max = Math.Abs(mat.data[0]);
int N = mat.getNumElements();
for (int i = 1; i < N; i++)
{
double a = Math.Abs(mat.data[i]);
if (a > max)
{
max = a;
}
}
switch (mat.numRows)
{
case 2:
inv2(mat, inv, 1.0 / max);
break;
case 3:
inv3(mat, inv, 1.0 / max);
break;
case 4:
inv4(mat, inv, 1.0 / max);
break;
case 5:
inv5(mat, inv, 1.0 / max);
break;
default: throw new ArgumentException("Not supported");
}
}
public static void convert(DMatrixRMaj src, FMatrixRMaj dst)
{
int N = src.getNumElements();
for (int i = 0; i < N; i++)
{
dst.data[i] = (float)src.data[i];
}
}
/**
* <p>
* Adds random values to each element in the matrix from an uniform distribution.<br>
* <br>
* a<sub>ij</sub> = a<sub>ij</sub> + U(min,max)<br>
* </p>
*
* @param A The matrix who is to be randomized. Modified
* @param min The minimum value each element can be.
* @param max The maximum value each element can be..
* @param rand Random number generator used to fill the matrix.
*/
public static void addUniform(DMatrixRMaj A, double min, double max, IMersenneTwister rand)
{
double[] d = A.getData();
int size = A.getNumElements();
double r = max - min;
for (int i = 0; i < size; i++)
{
d[i] += r * rand.NextDouble() + min;
}
}
/**
* <p>
* Creates a reflector from the provided vector and gamma.<br>
* <br>
* Q = I - γ u u<sup>T</sup><br>
* </p>
*
* <p>
* In practice {@link VectorVectorMult_DDRM#householder(double, DMatrixD1, DMatrixD1, DMatrixD1)} multHouseholder}
* should be used for performance reasons since there is no need to calculate Q explicitly.
* </p>
*
* @param u A vector. Not modified.
* @param gamma To produce a reflector gamma needs to be equal to 2/||u||.
* @return An orthogonal reflector.
*/
public static DMatrixRMaj createReflector(DMatrixRMaj u, double gamma)
{
if (!MatrixFeatures_DDRM.isVector(u))
{
throw new ArgumentException("u must be a vector");
}
DMatrixRMaj Q = CommonOps_DDRM.identity(u.getNumElements());
CommonOps_DDRM.multAddTransB(-gamma, u, u, Q);
return(Q);
}
/**
* Computes the QR decomposition of the provided matrix.
*
* @param A Matrix which is to be decomposed. Not modified.
*/
public void decompose(DMatrixRMaj A)
{
Equation.Equation eq = new Equation.Equation();
this.QR = (DMatrixRMaj)A.copy();
int N = Math.Min(A.numCols, A.numRows);
gammas = new double[A.numCols];
for (int i = 0; i < N; i++)
{
// update temporary variables
eq.alias(QR.numRows - i, "Ni", QR, "QR", i, "i");
// Place the column that should be zeroed into v
eq.process("v=QR(i:,i)");
eq.process("maxV=max(abs(v))");
// Note that v is lazily created above. Need direct access to it, which is done below.
DMatrixRMaj v = eq.lookupMatrix("v");
double maxV = eq.lookupDouble("maxV");
if (maxV > 0 && v.getNumElements() > 1)
{
// normalize to reduce overflow issues
eq.process("v=v/maxV");
// compute the magnitude of the vector
double tau = NormOps_DDRM.normF(v);
if (v.get(0) < 0)
{
tau *= -1.0;
}
eq.alias(tau, "tau");
eq.process("u_0 = v(0,0)+tau");
eq.process("gamma = u_0/tau");
eq.process("v=v/u_0");
eq.process("v(0,0)=1");
eq.process("QR(i:,i:) = (eye(Ni) - gamma*v*v')*QR(i:,i:)");
eq.process("QR(i:,i) = v");
eq.process("QR(i,i) = -1*tau*maxV");
// save gamma for recomputing Q later on
gammas[i] = eq.lookupDouble("gamma");
}
}
}
/**
* Normalizes the matrix such that the Frobenius norm is equal to one.
*
* @param A The matrix that is to be normalized.
*/
public static void normalizeF(DMatrixRMaj A)
{
double val = normF(A);
if (val == 0)
{
return;
}
int size = A.getNumElements();
for (int i = 0; i < size; i++)
{
A.div(i, val);
}
}
/**
* <p>
* Performs a rank one update on matrix A using vectors u and w. The results are stored in A.<br>
* <br>
* A = A + γ u w<sup>T</sup><br>
* </p>
* <p>
* This is called a rank1 update because the matrix u w<sup>T</sup> has a rank of 1.
* </p>
*
* @param gamma A scalar.
* @param A A m by m matrix. Modified.
* @param u A vector with m elements. Not modified.
*/
public static void rank1Update(double gamma,
DMatrixRMaj A,
DMatrixRMaj u,
DMatrixRMaj w)
{
int n = u.getNumElements();
int matrixIndex = 0;
for (int i = 0; i < n; i++)
{
double elementU = u.data[i];
for (int j = 0; j < n; j++)
{
A.data[matrixIndex++] += gamma * elementU * w.data[j];
}
}
}
/**
* Computes the QR decomposition of the provided matrix.
*
* @param A Matrix which is to be decomposed. Not modified.
*/
public void decompose(DMatrixRMaj A)
{
this.QR = (DMatrixRMaj)A.copy();
int N = Math.Min(A.numCols, A.numRows);
gammas = new double[A.numCols];
DMatrixRMaj A_small = new DMatrixRMaj(A.numRows, A.numCols);
DMatrixRMaj A_mod = new DMatrixRMaj(A.numRows, A.numCols);
DMatrixRMaj v = new DMatrixRMaj(A.numRows, 1);
DMatrixRMaj Q_k = new DMatrixRMaj(A.numRows, A.numRows);
for (int i = 0; i < N; i++)
{
// reshape temporary variables
A_small.reshape(QR.numRows - i, QR.numCols - i, false);
A_mod.reshape(A_small.numRows, A_small.numCols, false);
v.reshape(A_small.numRows, 1, false);
Q_k.reshape(v.getNumElements(), v.getNumElements(), false);
// use extract matrix to get the column that is to be zeroed
CommonOps_DDRM.extract(QR, i, QR.numRows, i, i + 1, v, 0, 0);
double max = CommonOps_DDRM.elementMaxAbs(v);
if (max > 0 && v.getNumElements() > 1)
{
// normalize to reduce overflow issues
CommonOps_DDRM.divide(v, max);
// compute the magnitude of the vector
double tau = NormOps_DDRM.normF(v);
if (v.get(0) < 0)
{
tau *= -1.0;
}
double u_0 = v.get(0) + tau;
double gamma = u_0 / tau;
CommonOps_DDRM.divide(v, u_0);
v.set(0, 1.0);
// extract the submatrix of A which is being operated on
CommonOps_DDRM.extract(QR, i, QR.numRows, i, QR.numCols, A_small, 0, 0);
// A = (I - γ*u*u<sup>T</sup>)A
CommonOps_DDRM.setIdentity(Q_k);
CommonOps_DDRM.multAddTransB(-gamma, v, v, Q_k);
CommonOps_DDRM.mult(Q_k, A_small, A_mod);
// save the results
CommonOps_DDRM.insert(A_mod, QR, i, i);
CommonOps_DDRM.insert(v, QR, i, i);
QR.unsafe_set(i, i, -tau * max);
// save gamma for recomputing Q later on
gammas[i] = gamma;
}
}
}
