float corr(Eigen::MatrixXf& v1, Eigen::MatrixXf& v2) {
Eigen::MatrixXf v1c = centerMatrix(v1);
Eigen::MatrixXf v2c = centerMatrix(v2);
float v1Norm = v1c.col(0).norm();
float v2Norm = v2c.col(0).norm();
if (v1Norm == 0.0 || v2Norm == 0.0) {
return 0.0;
}
float v1v2 = v1c.col(0).dot(v2c.col(0));
return v1v2 / (v1Norm * v2Norm);
}
DenseSymmetricMatrix compute_centered_kernel_matrix(RandomAccessIterator begin, RandomAccessIterator end,
KernelCallback callback)
{
timed_context context("Constructing kPCA centered kernel matrix");
DenseSymmetricMatrix kernel_matrix(end-begin,end-begin);
for (RandomAccessIterator i_iter=begin; i_iter!=end; ++i_iter)
{
for (RandomAccessIterator j_iter=i_iter; j_iter!=end; ++j_iter)
{
DefaultScalarType k = callback(*i_iter,*j_iter);
kernel_matrix(i_iter-begin,j_iter-begin) = k;
kernel_matrix(j_iter-begin,i_iter-begin) = k;
}
}
centerMatrix(kernel_matrix);
return kernel_matrix;
};
SparseWeightMatrix tangent_weight_matrix(RandomAccessIterator begin, RandomAccessIterator end,
const Neighbors& neighbors, PairwiseCallback callback,
const IndexType target_dimension, const ScalarType shift,
const bool partial_eigendecomposer=false)
{
timed_context context("KLTSA weight matrix computation");
const IndexType k = neighbors[0].size();
SparseTriplets sparse_triplets;
sparse_triplets.reserve((k*k+2*k+1)*(end-begin));
#pragma omp parallel shared(begin,end,neighbors,callback,sparse_triplets) default(none)
{
IndexType index_iter;
DenseMatrix gram_matrix = DenseMatrix::Zero(k,k);
DenseVector rhs = DenseVector::Ones(k);
DenseMatrix G = DenseMatrix::Zero(k,target_dimension+1);
G.col(0).setConstant(1/sqrt(static_cast<ScalarType>(k)));
DenseSelfAdjointEigenSolver solver;
SparseTriplets local_triplets;
local_triplets.reserve(k*k+2*k+1);
#pragma omp for nowait
for (index_iter=0; index_iter<static_cast<IndexType>(end-begin); index_iter++)
{
const LocalNeighbors& current_neighbors = neighbors[index_iter];
for (IndexType i=0; i<k; ++i)
{
for (IndexType j=i; j<k; ++j)
{
ScalarType kij = callback.kernel(begin[current_neighbors[i]],begin[current_neighbors[j]]);
gram_matrix(i,j) = kij;
gram_matrix(j,i) = kij;
}
}
centerMatrix(gram_matrix);
//UNRESTRICT_ALLOC;
#ifdef TAPKEE_WITH_ARPACK
if (partial_eigendecomposer)
{
G.rightCols(target_dimension).noalias() =
eigendecomposition<DenseMatrix,DenseMatrixOperation>(Arpack,gram_matrix,target_dimension,0).first;
}
else
#endif
{
solver.compute(gram_matrix);
G.rightCols(target_dimension).noalias() = solver.eigenvectors().rightCols(target_dimension);
}
//RESTRICT_ALLOC;
gram_matrix.noalias() = G * G.transpose();
SparseTriplet diagonal_triplet(index_iter,index_iter,shift);
local_triplets.push_back(diagonal_triplet);
for (IndexType i=0; i<k; ++i)
{
SparseTriplet neighborhood_diagonal_triplet(current_neighbors[i],current_neighbors[i],1.0);
local_triplets.push_back(neighborhood_diagonal_triplet);
for (IndexType j=0; j<k; ++j)
{
SparseTriplet tangent_triplet(current_neighbors[i],current_neighbors[j],-gram_matrix(i,j));
local_triplets.push_back(tangent_triplet);
}
}
#pragma omp critical
{
copy(local_triplets.begin(),local_triplets.end(),back_inserter(sparse_triplets));
}
local_triplets.clear();
}
}
return sparse_matrix_from_triplets(sparse_triplets, end-begin, end-begin);
}
SparseWeightMatrix hessian_weight_matrix(RandomAccessIterator begin, RandomAccessIterator end,
const Neighbors& neighbors, PairwiseCallback callback,
const IndexType target_dimension)
{
timed_context context("Hessian weight matrix computation");
const IndexType k = neighbors[0].size();
SparseTriplets sparse_triplets;
sparse_triplets.reserve(k*k*(end-begin));
const IndexType dp = target_dimension*(target_dimension+1)/2;
#pragma omp parallel shared(begin,end,neighbors,callback,sparse_triplets) default(none)
{
IndexType index_iter;
DenseMatrix gram_matrix = DenseMatrix::Zero(k,k);
DenseMatrix Yi(k,1+target_dimension+dp);
SparseTriplets local_triplets;
local_triplets.reserve(k*k+2*k+1);
#pragma omp for nowait
for (index_iter=0; index_iter<static_cast<IndexType>(end-begin); index_iter++)
{
const LocalNeighbors& current_neighbors = neighbors[index_iter];
for (IndexType i=0; i<k; ++i)
{
for (IndexType j=i; j<k; ++j)
{
ScalarType kij = callback.kernel(begin[current_neighbors[i]],begin[current_neighbors[j]]);
gram_matrix(i,j) = kij;
gram_matrix(j,i) = kij;
}
}
centerMatrix(gram_matrix);
DenseSelfAdjointEigenSolver sae_solver;
sae_solver.compute(gram_matrix);
Yi.col(0).setConstant(1.0);
Yi.block(0,1,k,target_dimension).noalias() = sae_solver.eigenvectors().rightCols(target_dimension);
IndexType ct = 0;
for (IndexType j=0; j<target_dimension; ++j)
{
for (IndexType p=0; p<target_dimension-j; ++p)
{
Yi.col(ct+p+1+target_dimension).noalias() = Yi.col(j+1).cwiseProduct(Yi.col(j+p+1));
}
ct += ct + target_dimension - j;
}
for (IndexType i=0; i<static_cast<IndexType>(Yi.cols()); i++)
{
for (IndexType j=0; j<i; j++)
{
ScalarType r = Yi.col(i).dot(Yi.col(j));
Yi.col(i) -= r*Yi.col(j);
}
ScalarType norm = Yi.col(i).norm();
Yi.col(i) *= (1.f / norm);
}
for (IndexType i=0; i<dp; i++)
{
ScalarType colsum = Yi.col(1+target_dimension+i).sum();
if (colsum > 1e-4)
Yi.col(1+target_dimension+i).array() /= colsum;
}
// reuse gram matrix storage m'kay?
gram_matrix.noalias() = Yi.rightCols(dp)*(Yi.rightCols(dp).transpose());
for (IndexType i=0; i<k; ++i)
{
for (IndexType j=0; j<k; ++j)
{
SparseTriplet hessian_triplet(current_neighbors[i],current_neighbors[j],gram_matrix(i,j));
local_triplets.push_back(hessian_triplet);
}
}
#pragma omp critical
{
copy(local_triplets.begin(),local_triplets.end(),back_inserter(sparse_triplets));
}
local_triplets.clear();
}
}
return sparse_matrix_from_triplets(sparse_triplets, end-begin, end-begin);
}
