def test_find_neighbor_and_sdist(self, PointSet):
ps = PointSet
fp = FastPair().build(ps)
rando = rand_tuple(len(ps[0]))
neigh = fp._find_neighbor(rando) # Abusing find_neighbor!
dist = fp.dist(rando, neigh["neigh"])
assert abs(dist - neigh["dist"]) < 1e-8
assert len(fp) == len(ps) # Make sure we didn't add a point...
l = [(fp.dist(a, b), b) for a, b in zip(cycle([rando]), ps)]
res = min(l, key=itemgetter(0))
assert abs(res[0] - neigh["dist"]) < 1e-8
assert res[1] == neigh["neigh"]
res = min(fp.sdist(rando), key=itemgetter(0))
assert abs(neigh["dist"] - res[0]) < 1e-8
assert neigh["neigh"] == res[1]
def test_find_neighbor_and_sdist(self):
ps = PointSet()
fp = FastPair().build(ps)
rando = rand_tuple(len(ps[0]))
neigh = fp._find_neighbor(rando) # Abusing find_neighbor!
dist = fp.dist(rando, neigh["neigh"])
assert abs(dist - neigh["dist"]) < 1e-8
assert len(fp) == len(ps) # Make sure we didn't add a point...
l = [(fp.dist(a, b), b) for a, b in zip(cycle([rando]), ps)]
res = min(l, key=itemgetter(0))
assert abs(res[0] - neigh["dist"]) < 1e-8
assert res[1] == neigh["neigh"]
res = min(fp.sdist(rando), key=itemgetter(0))
assert abs(neigh["dist"] - res[0]) < 1e-8
assert neigh["neigh"] == res[1]
class AddC(object):
"""Implements the AddC clustering algorithm.
For each data point arriving, the closest centroid to the incoming point
is moved towards the point. If there are more than `kmax` centroids,
then the two closest centroids are merged. This results in the creation of
a redundant centroid; the redundant centroid is then set equal to the new
data point. At any time, the data-structure can be queried for the current
set of centroids/clusters, or updated with additional data points.
"""
def __init__(self,
kmax=100,
dist=kernel_dist(gaussian),
centroid_factory=KernelCentroid):
"""Initialize an empty FastPair data-structure.
Parameters
----------
kmax : int, default=100
The maximum number of cluster centroids to store (i.e., size of
memory). This parameter controls the 'scale' of the desired
solution, such that larger values of `kmax` will lead to a higher
resolution cluster solution.
"""
self.kmax = kmax
self.npoints = 0
self.centroid_factory = centroid_factory
self.fastpair = FastPair(10, dist=dist)
def __add__(self, p):
"""Add a point to the AddC sketch."""
c = self.centroid_factory(p) # Create an 'empty' centroid at point `p`
self._step_one(c)
self._step_two()
self._step_three(c)
self.npoints += 1 # Update count of points seen so far
return self
def __len__(self):
"""Number of points in the AddC sketch."""
return len(self.fastpair)
def __call__(self):
"""Return the current set of cluster centroids."""
return self.centroids
def __contains__(self, p):
"""Test if a given cluster centroid is in the AddC sketch."""
return p in self.fastpair
def __iter__(self):
return iter(self.fastpair)
def _step_one(self, c):
# Step 1: Move the closest centroid towards the point
if len(self.fastpair) > 0:
# Single pass through list of neighbor points... this could also
# be sped up with a spatial index, though harder to do with
# kernel-induced distances
# Alternatively, if it was possible to insert the new data point
# _before_ querying for the closest pair (`step_two`), then we
# could do it that way...
old = min(self.fastpair.sdist(c), key=itemgetter(0))[1]
self.fastpair._update_point(old, old.add(c))
def _step_two(self):
# Step 2: Merge the two closest centroids
if len(self.fastpair) >= self.kmax and len(self.fastpair) > 1:
dist, (a, b) = self.fastpair.closest_pair()
self.fastpair -= b
self.fastpair._update_point(a, a.merge(b)) # Update point `a`
def _step_three(self, c):
# Step 3: Set redundant centroid equal to new point
self.fastpair += c
def batch(self, points):
# No checks, no nothing... just batch processing, pure and simple
for point in points:
self += point
return self
def trim(self, p=0.01):
"""Return only clusters over threshold."""
sub = [x.size for x in self if x.size > 0]
t = (sum(sub) / len(sub)) * p
return [x for x in self if x.size >= t]
@property
def centroids(self):
"""For plotting."""
return [c.center for c in self.fastpair]
class AddC(object):
"""Implements the AddC clustering algorithm.
For each data point arriving, the closest centroid to the incoming point
is moved towards the point. If there are more than `kmax` centroids,
then the two closest centroids are merged. This results in the creation of
a redundant centroid; the redundant centroid is then set equal to the new
data point. At any time, the data-structure can be queried for the current
set of centroids/clusters, or updated with additional data points.
"""
def __init__(self, kmax=100, dist=kernel_dist(gaussian),
centroid_factory=KernelCentroid):
"""Initialize an empty FastPair data-structure.
Parameters
----------
kmax : int, default=100
The maximum number of cluster centroids to store (i.e., size of
memory). This parameter controls the 'scale' of the desired
solution, such that larger values of `kmax` will lead to a higher
resolution cluster solution.
"""
self.kmax = kmax
self.npoints = 0
self.centroid_factory = centroid_factory
self.fastpair = FastPair(10, dist=dist)
def __add__(self, p):
"""Add a point to the AddC sketch."""
c = self.centroid_factory(p) # Create an 'empty' centroid at point `p`
self._step_one(c)
self._step_two()
self._step_three(c)
self.npoints += 1 # Update count of points seen so far
return self
def __len__(self):
"""Number of points in the AddC sketch."""
return len(self.fastpair)
def __call__(self):
"""Return the current set of cluster centroids."""
return self.centroids
def __contains__(self, p):
"""Test if a given cluster centroid is in the AddC sketch."""
return p in self.fastpair
def __iter__(self):
return iter(self.fastpair)
def _step_one(self, c):
# Step 1: Move the closest centroid towards the point
if len(self.fastpair) > 0:
# Single pass through list of neighbor points... this could also
# be sped up with a spatial index, though harder to do with
# kernel-induced distances
# Alternatively, if it was possible to insert the new data point
# _before_ querying for the closest pair (`step_two`), then we
# could do it that way...
old = min(self.fastpair.sdist(c), key=itemgetter(0))[1]
self.fastpair._update_point(old, old.add(c))
def _step_two(self):
# Step 2: Merge the two closest centroids
if len(self.fastpair) >= self.kmax and len(self.fastpair) > 1:
dist, (a, b) = self.fastpair.closest_pair()
self.fastpair -= b
self.fastpair._update_point(a, a.merge(b)) # Update point `a`
def _step_three(self, c):
# Step 3: Set redundant centroid equal to new point
self.fastpair += c
def batch(self, points):
# No checks, no nothing... just batch processing, pure and simple
for point in points:
self += point
return self
def trim(self, p=0.01):
"""Return only clusters over threshold."""
sub = [x.size for x in self if x.size > 0]
t = (sum(sub)/len(sub)) * p
return [x for x in self if x.size >= t]
@property
def centroids(self):
"""For plotting."""
return [c.center for c in self.fastpair]