def nn_radius(self, query, radius, **kwargs):
if self.__curindex is None:
raise FLANNException(
'build_index(...) method not called first or current index deleted.'
)
if query.dtype.type not in allowed_types:
raise FLANNException('Cannot handle type: %s' % query.dtype)
if self.__curindex_type != query.dtype.type:
raise FLANNException('Index and query must have the same type')
npts, dim = self.__curindex_data.shape
assert query.shape[0] == dim, 'data and query must have the same dims'
result = np.empty(npts, dtype=index_type)
if self.__curindex_type == np.float64:
dists = np.empty(npts, dtype=np.float64)
else:
dists = np.empty(npts, dtype=np.float32)
self.__flann_parameters.update(kwargs)
nn = flann.radius_search[self.__curindex_type](
self.__curindex, query, result, dists, npts, radius,
pointer(self.__flann_parameters))
return (result[0:nn], dists[0:nn])
def load_index(self, filename, pts):
"""
Loads an index previously saved to disk.
"""
if pts.dtype.type not in allowed_types:
raise FLANNException('Cannot handle type: %s' % pts.dtype)
pts = ensure_2d_array(pts, default_flags)
npts, dim = pts.shape
if self.__curindex is not None:
flann.free_index[self.__curindex_type](
self.__curindex, pointer(self.__flann_parameters))
self.__curindex = None
self.__curindex_data = None
self.__added_data = []
self.__curindex_type = None
self.__curindex = flann.load_index[pts.dtype.type](
c_char_p(to_bytes(filename)), pts, npts, dim)
if self.__curindex is None:
raise FLANNException(
('Error loading the FLANN index with filename=%r.'
' C++ may have thrown more detailed errors') % (filename,))
self.__curindex_data = pts
self.__added_data = []
self.__removed_ids = []
self.__curindex_type = pts.dtype.type
def add_points(self, new_pts, rebuild_threshold=2):
"""
Adds pts to the current index. If the number of added points is more
than a factor of rebuild_threshold larger than the original number of
points, the index is rebuilt.
"""
if new_pts.dtype.type not in allowed_types:
raise FLANNException('Cannot handle type: %s' % new_pts.dtype)
if new_pts.dtype != self.__curindex_type:
raise FLANNException('New points must have the same type')
new_pts = ensure_2d_array(new_pts, default_flags)
rows = new_pts.shape[0]
flann.add_points[self.__curindex_type](self.__curindex, new_pts, rows, rebuild_threshold)
self.__added_data.append(new_pts)
def kmeans(self,
pts,
num_clusters,
max_iterations=None,
dtype=None,
**kwargs):
"""
Runs kmeans on pts with num_clusters centroids. Returns a
numpy array of size num_clusters x dim.
If max_iterations is not None, the algorithm terminates after
the given number of iterations regardless of convergence. The
default is to run until convergence.
If dtype is None (the default), the array returned is the same
type as pts. Otherwise, the returned array is of type dtype.
"""
if int(num_clusters) != num_clusters or num_clusters < 1:
raise FLANNException('num_clusters must be an integer >= 1')
if num_clusters == 1:
if dtype is None or dtype == pts.dtype:
return np.mean(pts, 0).reshape(1, pts.shape[1])
else:
return dtype(np.mean(pts, 0).reshape(1, pts.shape[1]))
return self.hierarchical_kmeans(pts, int(num_clusters), 1,
max_iterations, dtype, **kwargs)
def nn_index(self, qpts, num_neighbors=1, **kwargs):
"""
For each point in querypts, (which may be a single point), it
returns the num_neighbors nearest points in the index built by
calling build_index.
"""
if self.__curindex is None:
raise FLANNException(
'build_index(...) method not called first or current index deleted.'
)
if qpts.dtype.type not in allowed_types:
raise FLANNException('Cannot handle type: %s' % qpts.dtype)
if self.__curindex_type != qpts.dtype.type:
raise FLANNException('Index and query must have the same type')
qpts = ensure_2d_array(qpts, default_flags)
npts, dim = self.__curindex_data.shape
if qpts.size == dim:
qpts.reshape(1, dim)
nqpts = qpts.shape[0]
assert qpts.shape[1] == dim, 'data and query must have the same dims'
assert npts >= num_neighbors, 'more neighbors than there are points'
result = np.empty((nqpts, num_neighbors), dtype=index_type)
if self.__curindex_type == np.float64:
dists = np.empty((nqpts, num_neighbors), dtype=np.float64)
else:
dists = np.empty((nqpts, num_neighbors), dtype=np.float32)
self.__flann_parameters.update(kwargs)
flann.find_nearest_neighbors_index[self.__curindex_type](
self.__curindex, qpts, nqpts, result, dists, num_neighbors,
pointer(self.__flann_parameters))
if num_neighbors == 1:
return (result.reshape(nqpts), dists.reshape(nqpts))
else:
return (result, dists)
def nn(self, pts, qpts, num_neighbors=1, **kwargs):
"""
Returns the num_neighbors nearest points in dataset for each point
in testset.
"""
if pts.dtype.type not in allowed_types:
raise FLANNException('Cannot handle type: %s' % pts.dtype)
if qpts.dtype.type not in allowed_types:
raise FLANNException('Cannot handle type: %s' % pts.dtype)
if pts.dtype != qpts.dtype:
raise FLANNException('Data and query must have the same type')
pts = ensure_2d_array(pts, default_flags)
qpts = ensure_2d_array(qpts, default_flags)
npts, dim = pts.shape
nqpts = qpts.shape[0]
assert qpts.shape[1] == dim, 'data and query must have the same dims'
assert npts >= num_neighbors, 'more neighbors than there are points'
result = np.empty((nqpts, num_neighbors), dtype=index_type)
if pts.dtype == np.float64:
dists = np.empty((nqpts, num_neighbors), dtype=np.float64)
else:
dists = np.empty((nqpts, num_neighbors), dtype=np.float32)
self.__flann_parameters.update(kwargs)
flann.find_nearest_neighbors[
pts.dtype.type](
pts, npts, dim, qpts, nqpts, result, dists, num_neighbors,
pointer(self.__flann_parameters))
if num_neighbors == 1:
return (result.reshape(nqpts), dists.reshape(nqpts))
else:
return (result, dists)
def add_points(self, pts, rebuild_threshold=2.0):
"""
Adds points to pre-built index.
Params:
pts: 2D numpy array of points.\n
rebuild_threshold: reallocs index when it grows by factor of \
`rebuild_threshold`. A smaller value results is more space \
efficient but less computationally efficient. Must be greater \
than 1.
"""
if not pts.dtype.type in allowed_types:
raise FLANNException("Cannot handle type: %s" % pts.dtype)
pts = ensure_2d_array(pts, default_flags)
npts, dim = pts.shape
flann.add_points[self.__curindex_type](self.__curindex, pts, npts, dim,
rebuild_threshold)
self.__curindex_data = np.row_stack((self.__curindex_data, pts))
def load_index(self, filename, pts):
"""
Loads an index previously saved to disk.
"""
if pts.dtype.type not in allowed_types:
raise FLANNException('Cannot handle type: %s' % pts.dtype)
pts = ensure_2d_array(pts, default_flags)
npts, dim = pts.shape
if self.__curindex is not None:
flann.free_index[self.__curindex_type](
self.__curindex, pointer(self.__flann_parameters))
self.__curindex = None
self.__curindex_data = None
self.__curindex_type = None
self.__curindex = flann.load_index[pts.dtype.type](c_char_p(
to_bytes(filename)), pts, npts, dim)
self.__curindex_data = pts
self.__curindex_type = pts.dtype.type
def build_index(self, pts, **kwargs):
"""
This builds and internally stores an index to be used for
future nearest neighbor matchings. It erases any previously
stored indexes, so use multiple instances of this class to
work with multiple stored indices. Use nn_index(...) to find
the nearest neighbors in this index.
pts is a 2d numpy array or matrix. All the computation is done
in np.float32 type, but pts may be any type that is convertable
to np.float32.
"""
if pts.dtype.type not in allowed_types:
raise FLANNException('Cannot handle type: %s' % pts.dtype)
pts = ensure_2d_array(pts, default_flags)
npts, dim = pts.shape
self.__ensureRandomSeed(kwargs)
self.__flann_parameters.update(kwargs)
if self.__curindex is not None:
flann.free_index[self.__curindex_type](
self.__curindex, pointer(self.__flann_parameters))
self.__curindex = None
speedup = c_float(0)
self.__curindex = flann.build_index[pts.dtype.type](
pts, npts, dim, byref(speedup), pointer(self.__flann_parameters))
self.__curindex_data = pts
self.__curindex_type = pts.dtype.type
params = dict(self.__flann_parameters)
params['speedup'] = speedup.value
return params
def hierarchical_kmeans(self,
pts,
branch_size,
num_branches,
max_iterations=None,
dtype=None,
**kwargs):
"""
Clusters the data by using multiple runs of kmeans to
recursively partition the dataset. The number of resulting
clusters is given by (branch_size-1)*num_branches+1.
This method can be significantly faster when the number of
desired clusters is quite large (e.g. a hundred or more).
Higher branch sizes are slower but may give better results.
If dtype is None (the default), the array returned is the same
type as pts. Otherwise, the returned array is of type dtype.
"""
# First verify the paremeters are sensible.
if pts.dtype.type not in allowed_types:
raise FLANNException('Cannot handle type: %s' % pts.dtype)
if int(branch_size) != branch_size or branch_size < 2:
raise FLANNException('branch_size must be an integer >= 2.')
branch_size = int(branch_size)
if int(num_branches) != num_branches or num_branches < 1:
raise FLANNException('num_branches must be an integer >= 1.')
num_branches = int(num_branches)
if max_iterations is None:
max_iterations = -1
else:
max_iterations = int(max_iterations)
# init the arrays and starting values
pts = ensure_2d_array(pts, default_flags)
npts, dim = pts.shape
num_clusters = (branch_size - 1) * num_branches + 1
if pts.dtype.type == np.float64:
result = np.empty((num_clusters, dim), dtype=np.float64)
else:
result = np.empty((num_clusters, dim), dtype=np.float32)
# set all the parameters appropriately
self.__ensureRandomSeed(kwargs)
params = {
'iterations': max_iterations,
'algorithm': 'kmeans',
'branching': branch_size,
'random_seed': kwargs['random_seed']
}
self.__flann_parameters.update(params)
numclusters = flann.compute_cluster_centers[pts.dtype.type](
pts, npts, dim, num_clusters, result,
pointer(self.__flann_parameters))
if numclusters <= 0:
raise FLANNException('Error occured during clustering procedure.')
if dtype is None:
return result
else:
return dtype(result)