def knn_search(K, x, y=None, min_dist=None, index=None, algorithm=None,
return_indices=False, dist_double=False, **kwargs):
'''
Calculates Euclidean distances to the first K closest elements of y
for each x, which are row-instance data matrices.
Returns a matrix whose (i, j)th element is the distance from the ith point
in x to the (j+1)th nearest neighbor in y.
If return_indices, also returns a matrix whose (i, j)th element is the
identity of the (j+1)th nearest neighbor in y to the ith point in x.
FLANN returns squared Euclidean distances of the same type as the input
data. If dist_double, if results are float32, cast them to float64 before
square-rooting for more numerical accuracy.
By default, clamps minimum distance to min(1e-2, 1e-100 ** (1/dim));
setting min_dist to a number changes this value. Use 0 for no clamping.
If index is passed, uses a preconstructed FLANN index for the elements of y
(a FLANN() instance where build_index() has been run). Otherwise, constructs
an index here and then deletes it, using the passed algorithm. By default,
uses a single k-d tree for data with dimension 5 or lower, and brute-force
search in higher dimensions (which give exact results). Any other keyword
arguments are also passed to the FLANN() object.
'''
N, dim = x.shape
if y is not None:
M, dim2 = y.shape
if dim != dim2:
raise TypeError("x and y must have same second dimension")
if not is_integer(K) or K < 1:
raise TypeError("K must be a positive integer")
if index is None:
if algorithm is None:
algorithm = pick_flann_algorithm(dim)
index = FLANNIndex(algorithm=algorithm, **kwargs)
index.build_index(x if y is None else x)
idx, dist = index.nn_index(x, K)
if return_indices:
idx = idx.astype(np.uint16)
if dist_double:
dist = np.asarray(dist, dtype=np.float64)
np.sqrt(dist, out=dist)
# protect against identical points
if min_dist is None:
min_dist = default_min_dist(dim)
if min_dist > 0:
np.maximum(min_dist, dist, out=dist)
return (dist, idx) if return_indices else dist
def _build_indices(X, flann_args):
"Builds FLANN indices for each bag."
# TODO: should probably multithread this
logger.info("Building indices...")
indices = [None] * len(X)
for i, bag in enumerate(plog(X, name="index building")):
indices[i] = idx = FLANNIndex(**flann_args)
idx.build_index(bag)
return indices
def pick_kmeans(x, n):
# NOTE: doesn't make sense to do this iteratively
# run k-means clustering
from vlfeat import vl_kmeans
centers = vl_kmeans(x, num_centers=n)
# pick points closest to the cluster centers
from cyflann import FLANNIndex
picked = FLANNIndex().nn(x, centers, num_neighbors=1)[0]
N = x.shape[0]
return picked, N**2 - (N - n)**2
def _make_index(bag):
idx = FLANNIndex(**self.flann_args)
idx.build_index(bag)
return idx
def _make_index(bag):
idx = FLANNIndex(**self.flann_args)
idx.build_index(bag)
return idx