pair< SparseMatrix<double>, SparseVector<double> > SVD_Nystrom_columnSpace_sparse(SparseMatrix<double> A, SparseMatrix<double> B){
// right singular vectors of SVD of A' * B, where A \in R^(n * d_A), B \in R^(n * d_B);
double norm_denom = 1.0 / (double)A.rows();
pair< SparseMatrix<double>, SparseVector<double> > USigma;
// Generate random_mat;
int k_prime = 2 * KHID;
SparseMatrix<double> random_mat = (B.cols() > 20 * KHID) ? random_embedding_mat((long)B.cols(), k_prime) : random_embedding_mat_dense((long)B.cols(), k_prime);
SparseMatrix<double> tmp = B * random_mat;
SparseMatrix<double> Q = A.transpose() * tmp; Q.makeCompressed(); Q.prune(TOLERANCE); tmp.resize(0, 0);
Q = orthogonalize_cols(Q);
SparseMatrix<double> tmp2 = B * Q;
SparseMatrix<double> C = A.transpose() * tmp2;
C = norm_denom * C;
MatrixXd Z = (MatrixXd)C.transpose() * C;
pair<pair<Eigen::MatrixXd, Eigen::MatrixXd>, Eigen::VectorXd> svd_Z = latenttree_svd(Z);
MatrixXd V = (svd_Z.first.second).leftCols(KHID);
SparseMatrix<double> V_sparse = V.sparseView();
VectorXd S = svd_Z.second.head(KHID);
USigma.second = S.cwiseSqrt().sparseView(); // S.array().sqrt();
MatrixXd diag_inv_S_sqrt = pinv_vector(S.cwiseSqrt()).asDiagonal();
SparseMatrix<double> diag_inv_S_sqrt_s = diag_inv_S_sqrt.sparseView();
USigma.first = C * (V_sparse)* diag_inv_S_sqrt_s;
return USigma;
}
pair< SparseMatrix<double>, SparseVector<double> > SVD_symNystrom_sparse(SparseMatrix<double> M)
{
pair< SparseMatrix<double>, SparseVector<double> > USigma;
int k_prime = 2 * KHID;
if (M.rows() > 20 * KHID)
{
SparseMatrix<double> Q;
if (M.cols() > k_prime){
SparseMatrix<double> random_mat = random_embedding_mat((long)M.cols(), k_prime);
Q = M * random_mat;
}
else{
Q = M;
}
Q = orthogonalize_cols((MatrixXd) Q);
SparseMatrix<double> C;
if (M.cols() < Q.rows()){
SparseMatrix<double> QQ = Q.block(0, 0, M.cols(), Q.cols());
C = M * QQ;
}
else
{
C = M * Q;
}
MatrixXd Z = (MatrixXd)C.transpose() * C;
// cout << "-----------Z:nonZeros()" << Z.nonZeros() << "-------------" << endl;
pair<pair<Eigen::MatrixXd, Eigen::MatrixXd>, Eigen::VectorXd> svd_Z = latenttree_svd(Z);
// cout << "-----------svd_Z.first.first:nonZeros()" << svd_Z.first.first.nonZeros() << "-------------" << endl;
// cout << "-----------svd_Z.first.second:nonZeros()" << svd_Z.first.second.nonZeros() << "-------------" << endl;
// cout << "-----------svd_Z.second:nonZeros()" << svd_Z.second.nonZeros() << "-------------" << endl;
MatrixXd V = (svd_Z.first.second).leftCols(KHID);
// cout << "-----------V:nonZeros()" << V.nonZeros() << "-------------" << endl;
SparseMatrix<double> V_sparse = V.sparseView();
VectorXd S = svd_Z.second.head(KHID);
// cout << "-----------S:nonZeros()" << S.nonZeros() << "-------------" << endl;
USigma.second = S.cwiseSqrt().sparseView(); // S.array().sqrt();
// cout << "-----------S.cwiseSqrt():nonZeros()" << S.cwiseSqrt().nonZeros() << "-------------" << endl;
// cout << "S" << endl << S << endl;
MatrixXd diag_inv_S_sqrt = pinv_vector(S.cwiseSqrt()).asDiagonal();
// cout << "-----------diag_inv_S_sqrt:nonZeros()" << diag_inv_S_sqrt.nonZeros() << "-------------" << endl;
SparseMatrix<double> diag_inv_S_sqrt_s = diag_inv_S_sqrt.sparseView();
USigma.first = C * (V_sparse)* diag_inv_S_sqrt_s;
// cout << "-----------USigma.first.:nonZeros()" << USigma.first.nonZeros() << "-------------" << endl;
// cout << "-----------USigma.second:nonZeros()" << USigma.second.nonZeros() << "-------------" << endl;
}
else
{
pair<pair<SparseMatrix<double>, SparseMatrix<double> >, SparseVector<double> > mv = latenttree_svd(M);
USigma.first = mv.first.first.leftCols(KHID);
USigma.second = mv.second.head(KHID);
}
return USigma;
}
MNESourceEstimate MinimumNorm::calculateInverse(const MatrixXd &data, float tmin, float tstep) const
{
if(!inverseSetup)
{
qWarning("Inverse not setup -> call doInverseSetup first!");
return MNESourceEstimate();
}
MatrixXd sol = K * data; //apply imaging kernel
if (inv.source_ori == FIFFV_MNE_FREE_ORI)
{
printf("combining the current components...");
MatrixXd sol1(sol.rows()/3,sol.cols());
for(qint32 i = 0; i < sol.cols(); ++i)
{
VectorXd* tmp = MNEMath::combine_xyz(sol.block(0,i,sol.rows(),1));
sol1.block(0,i,sol.rows()/3,1) = tmp->cwiseSqrt();
delete tmp;
}
sol.resize(sol1.rows(),sol1.cols());
sol = sol1;
}
if (m_bdSPM)
{
printf("(dSPM)...");
sol = inv.noisenorm*sol;
}
else if (m_bsLORETA)
{
printf("(sLORETA)...");
sol = inv.noisenorm*sol;
}
printf("[done]\n");
//Results
VectorXi p_vecVertices(inv.src[0].vertno.size() + inv.src[1].vertno.size());
p_vecVertices << inv.src[0].vertno, inv.src[1].vertno;
// VectorXi p_vecVertices();
// for(qint32 h = 0; h < inv.src.size(); ++h)
// t_qListVertices.push_back(inv.src[h].vertno);
return MNESourceEstimate(sol, p_vecVertices, tmin, tstep);
}
MatrixXd sample_MME_multiple_diagR(
MatrixXd Y,
SpMat W,
SpMat chol_C,
VectorXd pe,
MSpMat chol_K_inv,
VectorXd tot_Eta_prec,
MatrixXd randn_theta,
MatrixXd randn_e
){
MatrixXd theta_star = chol_K_inv.triangularView<Upper>().solve(randn_theta);
MatrixXd e_star = randn_e * pe.cwiseSqrt().cwiseInverse().asDiagonal();
MatrixXd W_theta_star = W * theta_star;
MatrixXd Y_resid = Y - W_theta_star - e_star;
MatrixXd WtRiy = W.transpose() * (Y_resid * pe.asDiagonal());
MatrixXd theta_tilda = chol_C.transpose().triangularView<Upper>().solve(chol_C.triangularView<Lower>().solve(WtRiy));
MatrixXd theta = theta_tilda * tot_Eta_prec.cwiseInverse().asDiagonal() + theta_star;
return theta;
}
/**
* The function main marks the entry point of the program.
* By default, main has the storage class extern.
*
* @param [in] argc (argument count) is an integer that indicates how many arguments were entered on the command line when the program was started.
* @param [in] argv (argument vector) is an array of pointers to arrays of character objects. The array objects are null-terminated strings, representing the arguments that were entered on the command line when the program was started.
* @return the value that was set to exit() (which is 0 if exit() is called via quit()).
*/
int main(int argc, char *argv[])
{
QCoreApplication a(argc, argv);
// fname_data - Name of the data file
// setno - Data set number
// fname_inv - Inverse operator file name
// nave - Number of averages (scales the noise covariance)
// If negative, the number of averages in the data will be
// used
// lambda2 - The regularization factor
// dSPM - do dSPM?
// sLORETA - do sLORETA?
QFile t_fileEvoked("./MNE-sample-data/MEG/sample/sample_audvis-ave.fif");
QFile t_fileInv("./MNE-sample-data/MEG/sample/sample_audvis-meg-eeg-oct-6-meg-eeg-inv.fif");
qint32 nave = 40;
float snr = 3.0f;
float lambda2 = pow(1.0f / snr, 2.0f);
bool dSPM = false;
bool sLORETA = true;
//
// Read the data first
//
FiffEvokedSet evokedSet(t_fileEvoked);
//
// Then the inverse operator
//
MNEInverseOperator inv_raw(t_fileInv);
//
// Iterate over found data sets
//
for(qint32 setno = 0; setno < evokedSet.evoked.size(); ++setno)
{
printf(">> Computing inverse for %s data set <<\n", evokedSet.evoked[setno].comment.toLatin1().constData());
//
// Set up the inverse according to the parameters
//
if (nave < 0)
nave = evokedSet.evoked[setno].nave;
MNEInverseOperator inv = inv_raw.prepare_inverse_operator(nave,lambda2,dSPM,sLORETA);
//
// Pick the correct channels from the data
//
FiffEvokedSet newEvokedSet = evokedSet.pick_channels(inv.noise_cov->names);
evokedSet = newEvokedSet;
printf("Picked %d channels from the data\n",evokedSet.info.nchan);
printf("Computing inverse...");
//
// Simple matrix multiplication followed by combination of the
// three current components
//
// This does all the data transformations to compute the weights for the
// eigenleads
//
SparseMatrix<double> reginv(inv.reginv.rows(),inv.reginv.rows());
// put this in the MNE algorithm class derived from inverse algorithm
//ToDo put this into a function of inv data
qint32 i;
for(i = 0; i < inv.reginv.rows(); ++i)
reginv.insert(i,i) = inv.reginv(i,0);
MatrixXd trans = reginv*inv.eigen_fields->data*inv.whitener*inv.proj*evokedSet.evoked[setno].data;
//
// Transformation into current distributions by weighting the eigenleads
// with the weights computed above
//
MatrixXd sol;
if (inv.eigen_leads_weighted)
{
//
// R^0.5 has been already factored in
//
printf("(eigenleads already weighted)...");
sol = inv.eigen_leads->data*trans;
}
else
{
//
// R^0.5 has to factored in
//
printf("(eigenleads need to be weighted)...");
SparseMatrix<double> sourceCov(inv.source_cov->data.rows(),inv.source_cov->data.rows());
for(i = 0; i < inv.source_cov->data.rows(); ++i)
sourceCov.insert(i,i) = sqrt(inv.source_cov->data(i,0));
sol = sourceCov*inv.eigen_leads->data*trans;
}
if (inv.source_ori == FIFFV_MNE_FREE_ORI)
{
printf("combining the current components...");
MatrixXd sol1(sol.rows()/3,sol.cols());
for(i = 0; i < sol.cols(); ++i)
{
VectorXd* tmp = MNE::combine_xyz(sol.block(0,i,sol.rows(),1));
sol1.block(0,i,sol.rows()/3,1) = tmp->cwiseSqrt();
delete tmp;
}
sol.resize(sol1.rows(),sol1.cols());
sol = sol1;
}
if (dSPM)
{
printf("(dSPM)...");
sol = inv.noisenorm*sol;
}
else if (sLORETA)
{
printf("(sLORETA)...");
sol = inv.noisenorm*sol;
}
printf("[done]\n");
//Results
float tmin = ((float)evokedSet.evoked[setno].first) / evokedSet.info.sfreq;
float tstep = 1/evokedSet.info.sfreq;
std::cout << "\npart ( block( 0, 0, 10, 10) ) of the inverse solution:\n" << sol.block(0,0,10,10) << std::endl;
printf("tmin = %f s\n", tmin);
printf("tstep = %f s\n", tstep);
}
return a.exec();
}
SourceEstimate MinimumNorm::calculateInverse(const FiffEvoked &p_fiffEvoked, bool pick_normal) const
{
//
// Set up the inverse according to the parameters
//
qint32 nave = p_fiffEvoked.nave;
if(!m_inverseOperator.check_ch_names(p_fiffEvoked.info))
{
qWarning("Channel name check failed.");
return SourceEstimate();
}
MNEInverseOperator inv = m_inverseOperator.prepare_inverse_operator(nave, m_fLambda, m_bdSPM, m_bsLORETA);
//
// Pick the correct channels from the data
//
FiffEvoked t_fiffEvoked = p_fiffEvoked.pick_channels(inv.noise_cov->names);
printf("Picked %d channels from the data\n",t_fiffEvoked.info.nchan);
printf("Computing inverse...");
MatrixXd K;
SparseMatrix<double> noise_norm;
QList<VectorXi> vertno;
Label label;
inv.assemble_kernel(label, m_sMethod, pick_normal, K, noise_norm, vertno);
MatrixXd sol = K * t_fiffEvoked.data; //apply imaging kernel
if (inv.source_ori == FIFFV_MNE_FREE_ORI)
{
printf("combining the current components...");
MatrixXd sol1(sol.rows()/3,sol.cols());
for(qint32 i = 0; i < sol.cols(); ++i)
{
VectorXd* tmp = MNEMath::combine_xyz(sol.block(0,i,sol.rows(),1));
sol1.block(0,i,sol.rows()/3,1) = tmp->cwiseSqrt();
delete tmp;
}
sol.resize(sol1.rows(),sol1.cols());
sol = sol1;
}
if (m_bdSPM)
{
printf("(dSPM)...");
sol = inv.noisenorm*sol;
}
else if (m_bsLORETA)
{
printf("(sLORETA)...");
sol = inv.noisenorm*sol;
}
printf("[done]\n");
//Results
float tmin = ((float)t_fiffEvoked.first) / t_fiffEvoked.info.sfreq;
float tstep = 1/t_fiffEvoked.info.sfreq;
QList<VectorXi> t_qListVertices;
for(qint32 h = 0; h < inv.src.size(); ++h)
t_qListVertices.push_back(inv.src[h].vertno);
return SourceEstimate(sol, t_qListVertices, tmin, tstep);
}
MNEInverseOperator MNEInverseOperator::prepare_inverse_operator(qint32 nave ,float lambda2, bool dSPM, bool sLORETA) const
{
if(nave <= 0)
{
printf("The number of averages should be positive\n");
return MNEInverseOperator();
}
printf("Preparing the inverse operator for use...\n");
MNEInverseOperator inv(*this);
//
// Scale some of the stuff
//
float scale = ((float)inv.nave)/((float)nave);
inv.noise_cov->data *= scale;
inv.noise_cov->eig *= scale;
inv.source_cov->data *= scale;
//
if (inv.eigen_leads_weighted)
inv.eigen_leads->data *= sqrt(scale);
//
printf("\tScaled noise and source covariance from nave = %d to nave = %d\n",inv.nave,nave);
inv.nave = nave;
//
// Create the diagonal matrix for computing the regularized inverse
//
VectorXd tmp = inv.sing.cwiseProduct(inv.sing) + VectorXd::Constant(inv.sing.size(), lambda2);
// if(inv.reginv)
// delete inv.reginv;
inv.reginv = VectorXd(inv.sing.cwiseQuotient(tmp));
printf("\tCreated the regularized inverter\n");
//
// Create the projection operator
//
qint32 ncomp = FiffProj::make_projector(inv.projs, inv.noise_cov->names, inv.proj);
if (ncomp > 0)
printf("\tCreated an SSP operator (subspace dimension = %d)\n",ncomp);
//
// Create the whitener
//
// if(inv.whitener)
// delete inv.whitener;
inv.whitener = MatrixXd::Zero(inv.noise_cov->dim, inv.noise_cov->dim);
qint32 nnzero, k;
if (inv.noise_cov->diag == 0)
{
//
// Omit the zeroes due to projection
//
nnzero = 0;
for (k = ncomp; k < inv.noise_cov->dim; ++k)
{
if (inv.noise_cov->eig[k] > 0)
{
inv.whitener(k,k) = 1.0/sqrt(inv.noise_cov->eig[k]);
++nnzero;
}
}
//
// Rows of eigvec are the eigenvectors
//
inv.whitener *= inv.noise_cov->eigvec;
printf("\tCreated the whitener using a full noise covariance matrix (%d small eigenvalues omitted)\n", inv.noise_cov->dim - nnzero);
}
else
{
//
// No need to omit the zeroes due to projection
//
for (k = 0; k < inv.noise_cov->dim; ++k)
inv.whitener(k,k) = 1.0/sqrt(inv.noise_cov->data(k,0));
printf("\tCreated the whitener using a diagonal noise covariance matrix (%d small eigenvalues discarded)\n",ncomp);
}
//
// Finally, compute the noise-normalization factors
//
if (dSPM || sLORETA)
{
VectorXd noise_norm = VectorXd::Zero(inv.eigen_leads->nrow);
VectorXd noise_weight;
if (dSPM)
{
printf("\tComputing noise-normalization factors (dSPM)...");
noise_weight = VectorXd(inv.reginv);
}
else
{
printf("\tComputing noise-normalization factors (sLORETA)...");
VectorXd tmp = (VectorXd::Constant(inv.sing.size(), 1) + inv.sing.cwiseProduct(inv.sing)/lambda2);
noise_weight = inv.reginv.cwiseProduct(tmp.cwiseSqrt());
}
VectorXd one;
if (inv.eigen_leads_weighted)
{
for (k = 0; k < inv.eigen_leads->nrow; ++k)
{
one = inv.eigen_leads->data.block(k,0,1,inv.eigen_leads->data.cols()).cwiseProduct(noise_weight);
noise_norm[k] = sqrt(one.dot(one));
}
}
else
{
// qDebug() << 32;
double c;
for (k = 0; k < inv.eigen_leads->nrow; ++k)
{
// qDebug() << 321;
c = sqrt(inv.source_cov->data(k,0));
// qDebug() << 322;
// qDebug() << "inv.eigen_leads->data" << inv.eigen_leads->data.rows() << "x" << inv.eigen_leads->data.cols();
// qDebug() << "noise_weight" << noise_weight.rows() << "x" << noise_weight.cols();
one = c*(inv.eigen_leads->data.row(k).transpose()).cwiseProduct(noise_weight);//ToDo eigenleads data -> pointer
noise_norm[k] = sqrt(one.dot(one));
// qDebug() << 324;
}
}
// qDebug() << 4;
//
// Compute the final result
//
VectorXd noise_norm_new;
if (inv.source_ori == FIFFV_MNE_FREE_ORI)
{
//
// The three-component case is a little bit more involved
// The variances at three consequtive entries must be squeared and
// added together
//
// Even in this case return only one noise-normalization factor
// per source location
//
VectorXd* t = MNEMath::combine_xyz(noise_norm.transpose());
noise_norm_new = t->cwiseSqrt();//double otherwise values are getting too small
delete t;
//
// This would replicate the same value on three consequtive
// entries
//
// noise_norm = kron(sqrt(mne_combine_xyz(noise_norm)),ones(3,1));
}
VectorXd vOnes = VectorXd::Ones(noise_norm_new.size());
VectorXd tmp = vOnes.cwiseQuotient(noise_norm_new.cwiseAbs());
// if(inv.noisenorm)
// delete inv.noisenorm;
typedef Eigen::Triplet<double> T;
std::vector<T> tripletList;
tripletList.reserve(noise_norm_new.size());
for(qint32 i = 0; i < noise_norm_new.size(); ++i)
tripletList.push_back(T(i, i, tmp[i]));
inv.noisenorm = SparseMatrix<double>(noise_norm_new.size(),noise_norm_new.size());
inv.noisenorm.setFromTriplets(tripletList.begin(), tripletList.end());
printf("[done]\n");
}
else
{
// if(inv.noisenorm)
// delete inv.noisenorm;
inv.noisenorm = SparseMatrix<double>();
}
return inv;
}
