TEMPLATE_HEADER
typename LR::Data LR::solveNormalEqn( Coordinate & c, Neighbors & nset ){
if( y.size().rows != Basis::size ){
A.resize( Basis::size, Basis::size );
y.resize( Basis::size, 1);
}
// find bounds for points
Coordinate lb = nset[0].point, rb = nset[0].point, tmp;
for( int j = 0; j < Coordinate::Dims ; j++ ){
for( int i = 1; i < nset.size(); i++ ){
lb[j] = fmin(nset[i].point[j],lb[j]);
rb[j] = fmax(nset[i].point[j],rb[j]);
}
}
// compute weights
std::vector<double> weights(nset.size());
wfunc(weights,nset);
// store values for A
A = 0;
for(int j = 0; j < Basis::size; j++ ){
for( int k = j; k < Basis::size; k++ ){
for( int i = 0; i < nset.size(); i++ ){
scale(lb,rb,nset[i].point,tmp);
A(j,k) += basis(k, tmp)*basis(j, tmp)*weights[i];
}
}
}
// store values for y
y = 0;
for(int j = 0; j < Basis::size; j++ ){
for( int i = 0; i < nset.size(); i++ ){
scale(lb,rb,nset[i].point,tmp);
y[j] += basis(j, tmp)*weights[i]*nset[i].data;
}
}
// solve system of equations
la::solve(A,y,coef);
// compute point
Data output = 0;
for( int i = 0; i < Basis::size; i++ ){
scale(lb,rb,c,tmp);
output += coef[i]*basis(i,tmp);
}
return output;
}
// Zhu et al. "A Rank-Order Distance based Clustering Algorithm for Face Tagging", CVPR 2011
// Ob(x) in eq. 1, modified to consider 0/1 as ground truth imposter/genuine.
static int indexOf(const Neighbors &neighbors, int i)
{
for (int j=0; j<neighbors.size(); j++) {
const Neighbor &neighbor = neighbors[j];
if (neighbor.first == i) {
if (neighbor.second == 0) return neighbors.size()-1;
else if (neighbor.second == 1) return 0;
else return j;
}
}
return -1;
}
SparseMatrix neighbors_distances_matrix(RandomAccessIterator begin, RandomAccessIterator end,
const Neighbors& neighbors, DistanceCallback callback,
ScalarType& average_distance)
{
const IndexType k = neighbors[0].size();
const IndexType n = neighbors.size();
if ((end-begin)!=n)
throw std::runtime_error("Wrong size");
SparseTriplets sparse_triplets;
sparse_triplets.reserve(k*n);
average_distance = 0;
ScalarType current_distance;
for (IndexType i = 0; i < n; ++i)
{
const LocalNeighbors& current_neighbors = neighbors[i];
for (IndexType j = 0; j < k; ++j)
{
current_distance = callback.distance(begin[i], begin[current_neighbors[j]]);
average_distance += current_distance;
SparseTriplet triplet(i, current_neighbors[j], current_distance);
sparse_triplets.push_back(triplet);
}
}
average_distance /= (k*n);
return sparse_matrix_from_triplets(sparse_triplets, n, n);
}
void DBSCAN::dbscan( const DBSCAN::DistanceMatrix& dm )
{
std::vector< uint8_t > visited( dm.size1() );
uint32_t cluster_id = 0;
for ( uint32_t pid = 0; pid < dm.size1(); ++pid ) {
if ( !visited[pid] ) {
visited[pid] = 1;
Neighbors ne = find_neighbors( dm, pid );
if ( ne.size() >= m_min_elems ) {
m_labels[pid] = cluster_id;
for ( uint32_t i = 0; i < ne.size(); ++i ) {
uint32_t nPid = ne[i];
if ( !visited[nPid] ) {
visited[nPid] = 1;
Neighbors ne1 = find_neighbors( dm, nPid );
if ( ne1.size() >= m_min_elems ) {
for ( const auto& n1 : ne1 ) {
ne.push_back( n1 );
}
}
}
if ( m_labels[nPid] == -1 ) {
m_labels[nPid] = cluster_id;
}
}
++cluster_id;
}
}
}
}
void ClustererDBSCAN::run_cluster(Points samples)
{
ClusterId cid = 1;
// foreach pid
for (PointId pid = 0; pid < samples.size(); pid++)
{
// not already visited
if (!_visited[pid]){
_visited[pid] = true;
// get the neighbors
Neighbors ne = findNeighbors(pid, _eps);
// not enough support -> mark as noise
if (ne.size() < _minPts)
{
_noise[pid] = true;
}
else
{
//else it's a core point
_core[pid] = true;
// Add p to current cluster
CCluster c; // a new cluster
c.push_back(pid); // assign pid to cluster
_pointId_to_clusterId[pid]=cid;
// go to neighbors
for (unsigned int i = 0; i < ne.size(); i++)
{
PointId nPid = ne[i];
// not already visited
if (!_visited[nPid])
{
_visited[nPid] = true;
// go to neighbors
Neighbors ne1 = findNeighbors(nPid, _eps);
// enough support
if (ne1.size() >= _minPts)
{
_core[nPid] = true;
// join
BOOST_FOREACH(Neighbors::value_type n1, ne1)
{
// join neighbord
ne.push_back(n1);
}
}
}
// not already assigned to a cluster
if (!_pointId_to_clusterId[nPid])
{
// add it to the current cluster
c.push_back(nPid);
_pointId_to_clusterId[nPid]=cid;
}
}
