DenseSymmetricMatrix compute_distance_matrix(RandomAccessIterator begin, RandomAccessIterator /*end*/,
Landmarks& landmarks, PairwiseCallback callback)
{
timed_context context("Multidimensional scaling distance matrix computation");
const IndexType n_landmarks = landmarks.size();
DenseSymmetricMatrix distance_matrix(n_landmarks,n_landmarks);
#pragma omp parallel shared(begin,landmarks,distance_matrix,callback) default(none)
{
IndexType i_index_iter,j_index_iter;
#pragma omp for nowait
for (i_index_iter=0; i_index_iter<n_landmarks; ++i_index_iter)
{
for (j_index_iter=i_index_iter; j_index_iter<n_landmarks; ++j_index_iter)
{
ScalarType d = callback.distance(begin[landmarks[i_index_iter]],begin[landmarks[j_index_iter]]);
d *= d;
distance_matrix(i_index_iter,j_index_iter) = d;
distance_matrix(j_index_iter,i_index_iter) = d;
}
}
}
return distance_matrix;
}
/**
* Be careful. The number of clusters changes and the number of members in
* each cluster can also change. Make sure memory is being managed well.
* @param n_in Number of input points to cluster
* @param dim_in Dimension of the points to cluster
* @param pts_in The list of n_in points of dimension dim_in
* @param ep_in Epsilon distance to use to search point density
* @param minpts_in Minimum number of points required to form a cluster
* @param cs_out Cluster size vector output
* @param clst_out Cluster list, a pass-by-reference of a int **
* @return Returns the number of clusters found
*/
int cluster_dbscan_density( int n_in, int dim_in, double *pts_in, double (*ep_in)(double*), int minpts_in, int **cs_out, int ***clst_out )
{
/* Variables nc = number of clusters, cs_out = cluster sizes */
int i,j,k,m,nc,nca,nnb,nba,nnbp,nbap,*nb,*nbp,*csa; /* Variable nb is for storing neighbors at each iteration */
int *vtd; /* Variables vtd[i] = -3 if pts_in[i] not visited, vtd[i] = -2 if visited, vtd[i] = -1 if noise, vtd[i] >= 0 if i belongs to cluster vtd[i] */
double tmp,dist,*mat;
/* Allocate space to store if each vertex is visited or not visited */
vtd = (int*) malloc( n_in * sizeof(int) );
for(i=0;i<n_in;i++)
vtd[i] = -3; /* Initialize not visited */
/* Allocate space for temporary neighbor storage */
nba = CLUSTER_MEMBER_INCREMENT;
nb = (int*) malloc( CLUSTER_MEMBER_INCREMENT * sizeof(int) );
nbap = CLUSTER_MEMBER_INCREMENT;
nbp = (int*) malloc( CLUSTER_MEMBER_INCREMENT * sizeof(int) );
/* Allocate base memory for outputting the clusters and cluster sizes */
nca = CLUSTER_INCREMENT; /* Initialize an allocated size for cluster data */
csa = (int*) malloc( CLUSTER_INCREMENT * sizeof(int) ); /* Variable csa stores the amount of space allocated for each cluster individually */
*cs_out = (int*) malloc( CLUSTER_INCREMENT * sizeof(int) );
*clst_out = (int**) malloc( CLUSTER_INCREMENT * sizeof(int*) );
#ifdef CLUSTER_USE_MATRIX
distance_matrix( n_in, dim_in, pts_in, &mat );
#endif
/* Iterate through all unvisited nodes and their neighbors and mark them visited */
nc = 0; /* Start with zero clusters */
for(i=0;i<n_in;i++)
{
/* If point i has been visited then continue */
if( vtd[i] != -3 )
continue;
vtd[i] = -2; /* Mark as visited but not necessarily noise */
/* Get all epsilon-neighbors of pts[i*dim] */
nnb = 0; /* Initialize number of neighbors found to zero */
for(j=0;j<n_in;j++)
{
if( j == i )
continue;
#ifdef CLUSTER_USE_MATRIX
dist = mat[i*n_in+j];
#else
dist = 0.0;
for(k=0;k<dim_in;k++)
tmp = pts_in[i*dim_in+k] - pts_in[j*dim_in+k], dist += tmp * tmp;
dist = sqrt( dist );
#endif
/* Tack it on to the end of nb for temporary keeping */
if( dist < ep_in( pts_in + i * dim_in ) )
{
/* Allocate more neighbor space if not enough already */
if( nnb + 1 > nba )
{
nba += CLUSTER_MEMBER_INCREMENT;
nb = (int*) realloc( nb, nba * sizeof(int) );
if( nb == NULL )
return -1;
}
nb[nnb++] = j;
}
}
/* Now have all neighbors counted; figure out what to do */
if( nnb < minpts_in )
vtd[i] = -1; /* Mark as noise and move on */
else /* Expand to create a new cluster */
{
/* Create a new cluster */
if( nc + 1 > nca )
{
nca += CLUSTER_INCREMENT; /* Increment nca to show that there is more space allocated */
csa = (int*) realloc( csa, nca * sizeof(int) ); /* Allocate more space for storing individual allocated cluster sizes */
*cs_out = (int*) realloc( *cs_out, nca * sizeof(int) ); /* Allocate more space for actual cluster sizes */
*clst_out = (int**) realloc( *clst_out, nca * sizeof(int*) ); /* Space for the actual cluster members */
}
/* Add onto the end at position nc */
csa[nc] = CLUSTER_MEMBER_INCREMENT;
(*cs_out)[nc] = 1;
(*clst_out)[nc] = (int*) malloc( CLUSTER_MEMBER_INCREMENT * sizeof(int) );
(*clst_out)[nc][0] = i;
vtd[i] = nc; /* Save the index of the cluster to which it belongs */
for(j=0;j<nnb;j++)
{
/* Do epsilon-density for each neighbor nb[j] */
if( vtd[nb[j]] == -3 )
{
vtd[nb[j]] = -2; /* Mark as visited but not necessarily noise */
nnbp = 0;
for(k=0;k<n_in;k++)
{
if( k == nb[j] )
continue;
#ifdef CLUSTER_USE_MATRIX
dist = mat[nb[j]*n_in+k];
#else
dist = 0.0;
for(m=0;m<dim_in;m++)
tmp = pts_in[nb[j]*dim_in+m] - pts_in[k*dim_in+m], dist += tmp * tmp;
dist = sqrt( dist );
#endif
if( dist < ep_in( pts_in + j * dim_in ) )
{
/* Allocate more space if necessary */
if( nnbp + 1 > nbap )
{
nbap += CLUSTER_MEMBER_INCREMENT;
nbp = (int*) realloc( nbp, nbap * sizeof(int) );
if( nbp == NULL )
return -1;
}
/* Add it to the neighbors prime list */
nbp[nnbp++] = k;
}
}
if( nnbp >= minpts_in )
{
/* Combine nb with nbp */
if( nnb + nnbp + 1 > nba )
{
nba += CLUSTER_MEMBER_INCREMENT;
nb = (int*) realloc( nb, nba * sizeof(int) );
if( nb == NULL )
return -1;
}
/* Combine these two neighbor groups */
for(k=0;k<nnbp;k++)
nb[nnb+k] = nbp[k];
nnb += nnbp; /* Important: Note that because nnb can increase, the loop expands! */
}
}
/* If nb[j] is not yet a part of a cluster, then add it to cluster nc */
if( vtd[nb[j]] < 0 ) /* Variable vtd[nb[j]] < 0 means taht nb[j] vertex is not part of a cluster yet */
{
/* Then add it to the new cluster just created */
if( (*cs_out)[nc] + 1 > csa[nc] )
{
csa[nc] += CLUSTER_MEMBER_INCREMENT;
(*clst_out)[nc] = (int*) realloc( (*clst_out)[nc], csa[nc] * sizeof(int) );
if( (*clst_out)[nc] == NULL )
return -1;
}
(*clst_out)[nc][(*cs_out)[nc]++] = nb[j];
vtd[nb[j]] = nc;
}
}
/* Finally increment after all additions made to the output cluster data structures */
++nc;
}
}
/* Clean up here; this will leak memory if you don't */
free( vtd );
free( csa );
free( nb );
free( nbp );
return nc;
}
inline tapkee::ScalarType operator()(int a, int b) const
{
return distance_matrix(a,b);
}