def euclidean_cluster(points,base_radius=None,n_clusters=None,rgen=default_chooser):
msg = "Must specify either a radius or a number of clusters."
assert not(base_radius is None and n_clusters is None), msg
iters=10
unassigned = range(np.shape(points)[1])
assigned = []
centers= []
while len(unassigned) > 0:
new_center = rgen(unassigned)
centers.append(new_center)
assigned.append(new_center)
unassigned.remove(new_center)
center_2d = np.reshape(points[:,new_center],[1,-1])
ua_dists = spatial.distance.cdist(points[:,unassigned].T,
center_2d)
ball = np.array(unassigned)[ua_dists.flatten() < base_radius]
for pt in ball:
assigned.append(pt)
unassigned.remove(pt)
pts_centers = points[:,centers]
for i in xrange(iters):
idxs = qmain.nn_search(pts_centers, points, 1)[1]
partition = idxs.flatten()
pts_centers = partition_centers(points,partition)
idxs = qmain.nn_search(pts_centers, points, 1)[1]
partition = idxs.flatten()
return partition, pts_centers
def cluster(eigvecs,eigvals,base_radius=None,knn=10):
n_points = np.shape(eigvecs)[0]
orphans = []
for i in xrange(n_points):
orphans.append(ClusterTreeNode([i]))
d = np.diag(eigvals)
points = d.dot(eigvecs.T)
new_parents = []
if base_radius is None:
knn = min(np.shape(points)[1],knn)
dists = qmain.nn_search(points,points,knn)[0]
base_radius = np.median(dists)
print "The estimated radius for building folders is {}.".format(base_radius)
partition_centers = np.array([[1,2],[1,2]]) #run loop at least once
while np.shape(partition_centers)[1] > 1:
partition, partition_centers = euclidean_cluster(points,base_radius)
for center in np.unique(partition):
new_parents.append(ClusterTreeNode([]))
for idx,orphan in enumerate(orphans):
orphan.assign_to_parent(new_parents[partition[idx]])
orphans = new_parents[:]
new_parents = []
d = d.dot(d)
points = d.dot(partition_centers)
assert len(orphans) == 1
orphans[0].make_index()
return orphans[0]
