def find_centroids(centroids, data, labels, pairwise_distances, zero_point, C):
# Get the number of points associated with each centroid
counts = np.bincount(labels, minlength=C)
# more bincounts using the positions as weights produce the unnormalized
# updated centroid locations (have to do each dimension separately since
# a weight cannot be a vector)
for idx in range(data.shape[1]):
centroids[:, idx] = np.bincount(labels,
weights=data[:, idx],
minlength=C)
# would have been nice if numpy offered a combined amin/argmin to avoid
# iterating over pairwise_distances twice
distance_sum = np.sum(np.amin(pairwise_distances, axis=1))
# To avoid introducing divide by zero errors
# If a centroid has no weight, we'll do no normalization
# This will keep its coordinates defined.
counts = np.maximum(counts, np.ones((1, ), dtype=np.uint64))
centroids /= counts[:, np.newaxis]
return distance_sum
def find_centroids(data, labels, C, D):
# Sort the points by their labels
indices = np.argsort(labels)
sorted_points = data[indices]
# Compute counts and indexes for ending of sets of points for each centroid
counts = np.bincount(labels, minlength=C)
indexes = np.cumsum(counts)
# Now we can use the indexes to split the array into sub-arrays and then
# sum across them to create the centroids
centroids = np.empty((C, D), dtype=data.dtype)
ragged_arrays = np.split(sorted_points, indexes)
for idx in xrange(C):
centroids[idx, :] = np.sum(ragged_arrays[idx], axis=0)
# To avoid introducing divide by zero errors
# If a centroid has no weight, we'll do no normalization
# This will keep its coordinates defined.
counts = np.maximum(counts, 1)
return centroids / counts[:, np.newaxis]
def find_centroids(centroids, data, labels, pairwise_distances, zero_point, C,
D):
# Get the number of points associated with each centroid
counts = np.bincount(labels, minlength=C)
# Build label masks for each centroid and sum across all the
# points assocated with each new centroid
distance_sum = 0.0
for idx in range(C):
# Boolean mask indicating where the points are for this center
centroid_mask = labels == idx
centroids[idx, :] = np.sum(np.where(centroid_mask[..., np.newaxis],
data, zero_point),
axis=0)
distance_sum += np.sum(
np.where(centroid_mask, pairwise_distances[:, idx], 0.0))
# To avoid introducing divide by zero errors
# If a centroid has no weight, we'll do no normalization
# This will keep its coordinates defined.
counts = np.maximum(counts, np.ones((1, ), dtype=np.uint64))
centroids /= counts[:, np.newaxis]
return distance_sum