def neighborhood_graph(D, k=1, eps=0., verbose=True, callback=None):
r'''Neighborhood graph.'''
ds = squareform(D, force='tomatrix')
np.fill_diagonal(ds, np.inf)
min_eps = ds.min(axis=0).max()
if verbose:
print('Minimal epsilon: {0}.'.format(min_eps))
N = len(ds)
G = _csr_graph(N)
UF = union_find(N)
i = 0
for r in range(N):
for c in range(r+1,N):
if D[i]<=eps:
G.reserve_edge(r)
G.reserve_edge(c)
UF.Union(r,c)
i += 1
row1, col1 = np.nonzero(ds<=eps)
if k>1:
d, j = nearest_neighbors_from_dm(D, k, callback=callback)
del d
# Add two edges (both directions) for each pair of kNN
row_col = np.empty((2*N*k,2), dtype=j.dtype)
row_col[:N*k,0] = np.repeat(np.arange(N), k)
row_col[:N*k,1] = j.flat # np.ravel(j) ?
row_col[N*k:,0] = row_col[:N*k,1]
row_col[N*k:,1] = row_col[:N*k,0]
# Remove points as their own nearest neighbors.
row_col = row_col[row_col[:,0]!=row_col[:,1]]
# Keep only edges with distance >eps
s = (ds[row_col[:,0],row_col[:,1]] > eps)
row_col = row_col[s]
# Remove all duplicate edges.
row, col = _unique_rows(row_col).T # TBD timeit
for r in row:
G.reserve_edge(r)
for r,c in zip(row,col):
UF.Union(r,c)
G.finalize()
i = 0
for r in range(N):
for c in range(r+1,N):
if D[i]<=eps:
G.add_edge(r, c, D[i])
G.add_edge(c, r, D[i])
i += 1
if k>1:
for r,c in zip(row,col):
G.add_edge(r,c,ds[r,c])
G.validate()
if UF.ncomp>1:
print('The neighborhood graph has {0} components.'.format(UF.ncomp))
G.ncomp = UF.ncomp
return (G.rowstart, G.targets, G.weights)
def neighborhood_graph(D, k=1, eps=0., verbose=True, callback=None):
r'''Neighborhood graph.'''
ds = squareform(D, force='tomatrix')
np.fill_diagonal(ds, np.inf)
min_eps = ds.min(axis=0).max()
if verbose:
print('Minimal epsilon: {0}.'.format(min_eps))
N = len(ds)
G = _csr_graph(N)
UF = union_find(N)
i = 0
for r in range(N):
for c in range(r + 1, N):
if D[i] <= eps:
G.reserve_edge(r)
G.reserve_edge(c)
UF.Union(r, c)
i += 1
row1, col1 = np.nonzero(ds <= eps)
if k > 1:
d, j = nearest_neighbors_from_dm(D, k, callback=callback)
del d
# Add two edges (both directions) for each pair of kNN
row_col = np.empty((2 * N * k, 2), dtype=j.dtype)
row_col[:N * k, 0] = np.repeat(np.arange(N), k)
row_col[:N * k, 1] = j.flat # np.ravel(j) ?
row_col[N * k:, 0] = row_col[:N * k, 1]
row_col[N * k:, 1] = row_col[:N * k, 0]
# Remove points as their own nearest neighbors.
row_col = row_col[row_col[:, 0] != row_col[:, 1]]
# Keep only edges with distance >eps
s = (ds[row_col[:, 0], row_col[:, 1]] > eps)
row_col = row_col[s]
# Remove all duplicate edges.
row, col = _unique_rows(row_col).T # TBD timeit
for r in row:
G.reserve_edge(r)
for r, c in zip(row, col):
UF.Union(r, c)
G.finalize()
i = 0
for r in range(N):
for c in range(r + 1, N):
if D[i] <= eps:
G.add_edge(r, c, D[i])
G.add_edge(c, r, D[i])
i += 1
if k > 1:
for r, c in zip(row, col):
G.add_edge(r, c, ds[r, c])
G.validate()
if UF.ncomp > 1:
print('The neighborhood graph has {0} components.'.format(UF.ncomp))
G.ncomp = UF.ncomp
return (G.rowstart, G.targets, G.weights)
def nn(k):
return nearest_neighbors_from_dm(D, k, callback)
def nn(k):
return nearest_neighbors_from_dm(D, k, callback)