int MAPGetDir(VecCl &Par,VecCl &MaxStep,MatrCl &DirMat,double ErrorMatr,
double &CurHi)
{
int Dim=Par.Dim(),DimGr=1;
if (MAPVar->ExperPoint.Ptr!=NULL) DimGr=MAPVar->ExperPoint.Dim();
if ((Dim>0) && (DimGr>0))
{
int k;//,k1;
double d;
MatrCl DatM;
VecCl MisV(Dim),Norm(Dim),EigV(Dim);
// SetGrFunc(DatM,MisV,Par,MaxStep);
// SetCorelFunc(DatM,MisV,CurHi); // now DatGr(Dim,Dim) correlations only
// MisV - Derivativ of the function
MAPVar->SetCorrelation(DatM,MisV,CurHi,Par,MaxStep);
for (k=1;k<=Dim;k++)
if ((d=DatM(k,k))>MathZer) Norm[k]=1/sqrt(d);else Norm[k]=0;
DatM=VectorNormalizeMatr(Norm,DatM,Norm);
if (!ReduceLQ(DatM,DatM,EigV.Ptr,ErrorMatr))
{cout<<" Cannot Calculate ReduceLQ. \n";return 0;}
DatM=Transpon(DatM);
for (k=1;k<=Dim;k++)
if ( (fabs(d=EigV[k]))>MathZer) EigV[k]=1/d;else EigV[k]=0;
MaxStep=MaxStep*0-VectorNormalizeMatr(EigV,DatM,Norm)*MisV;
//cout<<" MaxStep "<<MaxStep<<"\n";
DirMat=VectorNormalizeMatr(Norm,Transpon(DatM),Norm*0+1);
}
return 1;
};
void omxCallGREMLFitFunction(omxFitFunction *oo, int want, FitContext *fc){
if (want & (FF_COMPUTE_PREOPTIMIZE)) return;
//Recompute Expectation:
omxExpectation* expectation = oo->expectation;
omxExpectationCompute(expectation, NULL);
omxGREMLFitState *gff = (omxGREMLFitState*)oo->argStruct; //<--Cast generic omxFitFunction to omxGREMLFitState
//Ensure that the pointer in the GREML fitfunction is directed at the right FreeVarGroup
//(not necessary for most compute plans):
if(fc && gff->varGroup != fc->varGroup){
gff->buildParamMap(fc->varGroup);
}
//Declare local variables used in more than one scope in this function:
const double Scale = fabs(Global->llScale); //<--absolute value of loglikelihood scale
const double NATLOG_2PI = 1.837877066409345483560659472811; //<--log(2*pi)
int i;
Eigen::Map< Eigen::MatrixXd > Eigy(omxMatrixDataColumnMajor(gff->y), gff->y->cols, 1);
Eigen::Map< Eigen::MatrixXd > Vinv(omxMatrixDataColumnMajor(gff->invcov), gff->invcov->rows, gff->invcov->cols);
EigenMatrixAdaptor EigX(gff->X);
Eigen::MatrixXd P, Py;
P.setZero(gff->invcov->rows, gff->invcov->cols);
double logdetV=0, logdetquadX=0;
if(want & (FF_COMPUTE_FIT | FF_COMPUTE_GRADIENT | FF_COMPUTE_HESSIAN | FF_COMPUTE_IHESSIAN)){
if(gff->usingGREMLExpectation){
omxGREMLExpectation* oge = (omxGREMLExpectation*)(expectation->argStruct);
//Check that factorizations of V and the quadratic form in X succeeded:
if(oge->cholV_fail_om->data[0]){
oo->matrix->data[0] = NA_REAL;
if (fc) fc->recordIterationError("expected covariance matrix is non-positive-definite");
return;
}
if(oge->cholquadX_fail){
oo->matrix->data[0] = NA_REAL;
if (fc) fc->recordIterationError("Cholesky factorization failed; possibly, the matrix of covariates is rank-deficient");
return;
}
//Log determinant of V:
logdetV = oge->logdetV_om->data[0];
//Log determinant of quadX:
for(i=0; i < gff->X->cols; i++){
logdetquadX += log(oge->cholquadX_vectorD[i]);
}
logdetquadX *= 2;
gff->REMLcorrection = Scale*0.5*logdetquadX;
//Finish computing fit (negative loglikelihood):
P.triangularView<Eigen::Lower>() = Vinv.selfadjointView<Eigen::Lower>() *
(Eigen::MatrixXd::Identity(Vinv.rows(), Vinv.cols()) -
(EigX * oge->quadXinv.selfadjointView<Eigen::Lower>() * oge->XtVinv)); //Vinv * (I-Hatmat)
Py = P.selfadjointView<Eigen::Lower>() * Eigy;
oo->matrix->data[0] = gff->REMLcorrection + Scale*0.5*( (((double)gff->y->cols) * NATLOG_2PI) + logdetV + ( Eigy.transpose() * Py )(0,0));
gff->nll = oo->matrix->data[0];
}
else{ //If not using GREML expectation, deal with means and cov in a general way to compute fit...
//Declare locals:
EigenMatrixAdaptor yhat(gff->means);
EigenMatrixAdaptor EigV(gff->cov);
double logdetV=0, logdetquadX=0;
Eigen::MatrixXd Vinv, quadX;
Eigen::LLT< Eigen::MatrixXd > cholV(gff->cov->rows);
Eigen::LLT< Eigen::MatrixXd > cholquadX(gff->X->cols);
Eigen::VectorXd cholV_vectorD, cholquadX_vectorD;
//Cholesky factorization of V:
cholV.compute(EigV);
if(cholV.info() != Eigen::Success){
omxRaiseErrorf("expected covariance matrix is non-positive-definite");
oo->matrix->data[0] = NA_REAL;
return;
}
//Log determinant of V:
cholV_vectorD = (( Eigen::MatrixXd )(cholV.matrixL())).diagonal();
for(i=0; i < gff->X->rows; i++){
logdetV += log(cholV_vectorD[i]);
}
logdetV *= 2;
Vinv = cholV.solve(Eigen::MatrixXd::Identity( EigV.rows(), EigV.cols() )); //<-- V inverse
quadX = EigX.transpose() * Vinv * EigX; //<--Quadratic form in X
cholquadX.compute(quadX); //<--Cholesky factorization of quadX
if(cholquadX.info() != Eigen::Success){
omxRaiseErrorf("Cholesky factorization failed; possibly, the matrix of covariates is rank-deficient");
oo->matrix->data[0] = NA_REAL;
return;
}
cholquadX_vectorD = (( Eigen::MatrixXd )(cholquadX.matrixL())).diagonal();
for(i=0; i < gff->X->cols; i++){
logdetquadX += log(cholquadX_vectorD[i]);
}
logdetquadX *= 2;
gff->REMLcorrection = Scale*0.5*logdetquadX;
//Finish computing fit:
oo->matrix->data[0] = gff->REMLcorrection + Scale*0.5*( ((double)gff->y->rows * NATLOG_2PI) + logdetV +
( Eigy.transpose() * Vinv * (Eigy - yhat) )(0,0));
gff->nll = oo->matrix->data[0];
return;
}
}
if(want & (FF_COMPUTE_GRADIENT | FF_COMPUTE_HESSIAN | FF_COMPUTE_IHESSIAN)){
//This part requires GREML expectation:
omxGREMLExpectation* oge = (omxGREMLExpectation*)(expectation->argStruct);
//Declare local variables for this scope:
int numChildren = fc->childList.size();
int __attribute__((unused)) parallelism = (numChildren == 0) ? 1 : numChildren;
fc->grad.resize(gff->dVlength); //<--Resize gradient in FitContext
//Set up new HessianBlock:
HessianBlock *hb = new HessianBlock;
if(want & (FF_COMPUTE_HESSIAN | FF_COMPUTE_IHESSIAN)){
hb->vars.resize(gff->dVlength);
hb->mat.resize(gff->dVlength, gff->dVlength);
}
//Begin looping thru free parameters:
#pragma omp parallel for num_threads(parallelism)
for(i=0; i < gff->dVlength; i++){
//Declare locals within parallelized region:
int j=0, t1=0, t2=0;
Eigen::MatrixXd PdV_dtheta1;
Eigen::MatrixXd dV_dtheta1(Eigy.rows(), Eigy.rows()); //<--Derivative of V w/r/t parameter i.
Eigen::MatrixXd dV_dtheta2(Eigy.rows(), Eigy.rows()); //<--Derivative of V w/r/t parameter j.
t1 = gff->gradMap[i]; //<--Parameter number for parameter i.
if(t1 < 0){continue;}
if(want & (FF_COMPUTE_HESSIAN | FF_COMPUTE_IHESSIAN)){hb->vars[i] = t1;}
if( oge->numcases2drop ){
dropCasesAndEigenize(gff->dV[i], dV_dtheta1, oge->numcases2drop, oge->dropcase, 1);
}
else{dV_dtheta1 = Eigen::Map< Eigen::MatrixXd >(omxMatrixDataColumnMajor(gff->dV[i]), gff->dV[i]->rows, gff->dV[i]->cols);}
//PdV_dtheta1 = P.selfadjointView<Eigen::Lower>() * dV_dtheta1.selfadjointView<Eigen::Lower>();
PdV_dtheta1 = P.selfadjointView<Eigen::Lower>();
PdV_dtheta1 = PdV_dtheta1 * dV_dtheta1.selfadjointView<Eigen::Lower>();
for(j=i; j < gff->dVlength; j++){
if(j==i){
gff->gradient(t1) = Scale*0.5*(PdV_dtheta1.trace() - (Eigy.transpose() * PdV_dtheta1 * Py)(0,0));
fc->grad(t1) += gff->gradient(t1);
if(want & (FF_COMPUTE_HESSIAN | FF_COMPUTE_IHESSIAN)){
gff->avgInfo(t1,t1) = Scale*0.5*(Eigy.transpose() * PdV_dtheta1 * PdV_dtheta1 * Py)(0,0);
}
}
else{if(want & (FF_COMPUTE_HESSIAN | FF_COMPUTE_IHESSIAN)){
t2 = gff->gradMap[j]; //<--Parameter number for parameter j.
if(t2 < 0){continue;}
if( oge->numcases2drop ){
dropCasesAndEigenize(gff->dV[j], dV_dtheta2, oge->numcases2drop, oge->dropcase, 1);
}
else{dV_dtheta2 = Eigen::Map< Eigen::MatrixXd >(omxMatrixDataColumnMajor(gff->dV[j]), gff->dV[j]->rows, gff->dV[j]->cols);}
gff->avgInfo(t1,t2) = Scale*0.5*(Eigy.transpose() * PdV_dtheta1 * P.selfadjointView<Eigen::Lower>() * dV_dtheta2.selfadjointView<Eigen::Lower>() * Py)(0,0);
gff->avgInfo(t2,t1) = gff->avgInfo(t1,t2);
}}}}
//Assign upper triangle elements of avgInfo to the HessianBlock:
if(want & (FF_COMPUTE_HESSIAN | FF_COMPUTE_IHESSIAN)){
for (size_t d1=0, h1=0; h1 < gff->dV.size(); ++h1) {
for (size_t d2=0, h2=0; h2 <= h1; ++h2) {
hb->mat(d2,d1) = gff->avgInfo(h2,h1);
++d2;
}
++d1;
}
fc->queue(hb);
}}
