def find_initial_labels(self):
# stage 1 - find initial number of clusters
K = 0
d_copy = self.dataset.view() # create view of dataset
N = self.N
M = self.M
beta = self.beta
centres = np.zeros((self.MAX_CLUSTERS, N - 2),
dtype=np.float32) # temporary variable for centres
counts = np.ones((self.MAX_CLUSTERS),
dtype=np.int32) # temporary variable for counts
res = clustered_object()
# add first point to first cluster
centres[K] += (d_copy[0, 1:N - 1] - centres[K]) / counts[K]
counts[K] += 1
d_copy[0, N - 1] = K
K += 1
for i in range(1, M):
data_tmp = d_copy[i, 1:N - 1]
#print("current point ",data_tmp)
(best_key, max_prob) = find_cluster_s1(
data_tmp,
centres[0:K],
counts[0:K],
) #find max probability and return corresponding index
if max_prob < -beta:
best_key = K
K += 1
centre_tmp = centres[best_key]
new_centre = update_centre(data_tmp,
centre_tmp,
counts[best_key],
sign=1)
centres[best_key] = new_centre
counts[best_key] = counts[best_key] + 1
d_copy[i, N - 1] = best_key
res.clustered_set = self.dataset
res.cluster_labels, res.cluster_counts = np.unique(d_copy[:, N - 1],
return_counts=True)
res.cluster_centres = centres[0:K]
return res
def assign_labels(self,
label_ids=None,
prev_counts=None,
prev_centres=None,
converge=False,
numIter=10):
if label_ids is None:
# no previously assigned labels -> stage 1
return self.find_initial_labels()
else:
if not converge:
# to run just one iteration of stage 2 (Academic purposes only)
K = len(label_ids)
centres = prev_centres
new_counts = np.ones(
(K, ), dtype=np.int32) # temporary variable for counts
d_copy = self.dataset.view() # create view of dataset
N = self.N
M = self.M
likelihood = (np.log(prev_counts) * self.r).astype(np.float32)
res = clustered_object()
for i in range(M):
data_tmp = d_copy[i, 1:N - 1]
(best_key) = find_cluster_s2(
data_tmp, centres, new_counts, likelihood
) #find max probability and return corresponding index
centre_tmp = centres[best_key]
new_centre = update_centre(data_tmp,
centre_tmp,
new_counts[best_key],
sign=1)
centres[best_key] = new_centre
new_counts[best_key] = new_counts[best_key] + 1
d_copy[i, N - 1] = label_ids[best_key]
active_idx = new_counts != 1 # get mask for non zero clusters
# remove counts and centriods of empty or inactive clusters
res.clustered_set = self.dataset
res.cluster_labels, self.indices = np.unique(
d_copy[:, N - 1], return_inverse=True)
res.cluster_counts = new_counts[active_idx]
res.cluster_centres = centres[active_idx]
return res
else:
# stage 2 - iterate till convergence
d_copy = self.dataset.view()
N = self.N
M = self.M
counts = prev_counts
cur_centres = prev_centres.copy()
labels = d_copy[:, N - 1]
res = clustered_object()
for n in range(numIter):
likelihood = (np.log(counts) * self.r).astype(np.float32)
for i in range(M):
data_tmp = d_copy[i, 1:N - 1]
old_idx = self.indices[i] #idx[labels[i]]
# unassign from the previously assigned cluster
if counts[old_idx] >= 1:
centre_tmp = cur_centres[old_idx]
new_centre = update_centre(data_tmp,
centre_tmp,
counts[old_idx],
sign=-1)
cur_centres[old_idx] = new_centre
counts[old_idx] -= 1
# calculate posterior probability
(best_key) = find_cluster_s2(
data_tmp, cur_centres, counts, likelihood
) #find max probability and return corresponding index
# reassign to correct cluster
centre_tmp = cur_centres[best_key]
new_centre = update_centre(data_tmp,
centre_tmp,
counts[best_key],
sign=1)
cur_centres[best_key] = new_centre
counts[best_key] += 1
# d_copy[i, N-1] = label_ids[best_key]
labels[i] = label_ids[best_key]
# remove inactive clusters
active_idx = counts > 1
label_ids, counts = label_ids[active_idx], counts[
active_idx]
l_tmp, self.indices = np.unique(d_copy[:, N - 1],
return_inverse=True)
cur_centres = cur_centres[active_idx]
if np.array_equal(cur_centres, prev_centres):
# if centroids are not changing anymore, return after removing inactive clusters
active_idx = counts > 1
print('number of iterations = %d' % n)
res.clustered_set = self.dataset
res.cluster_labels, res.cluster_counts = label_ids[
active_idx], counts[active_idx]
res.cluster_centres = cur_centres[active_idx]
return res
else:
prev_centres = cur_centres.copy()
res.clustered_set = self.dataset
res.cluster_labels = label_ids.astype(int)
res.cluster_counts = counts
res.cluster_centres = cur_centres
return res
def assign_labels(self,
label_ids=None,
prev_counts=None,
prev_centres=None,
converge=False,
numIter=10,
indices=None):
if label_ids is None:
return self.find_initial_labels()
else:
if not converge:
K = len(label_ids)
centres = prev_centres
new_counts = np.ones(
(K, ), dtype=np.int32) # temporary variable for counts
d_copy = self.dataset.view() # create view of dataset
N = self.N
M = self.M
likelihood = (np.log(prev_counts) * self.r).astype(np.float32)
res = clustered_object()
for i in range(M):
data_tmp = d_copy[i, 1:N - 1]
(best_key) = find_cluster_s2(
data_tmp, centres, new_counts, likelihood
) #find max probability and return corresponding index
centre_tmp = centres[best_key]
new_centre = update_centre(data_tmp,
centre_tmp,
new_counts[best_key],
sign=1)
centres[best_key] = new_centre
new_counts[best_key] = new_counts[best_key] + 1
d_copy[i, N - 1] = label_ids[best_key]
active_idx = new_counts != 1
res.clustered_set = self.dataset
res.cluster_labels, self.indices = np.unique(
d_copy[:, N - 1], return_inverse=True)
res.cluster_counts = new_counts[active_idx]
res.cluster_centres = centres[active_idx]
return res
else:
d_copy = self.dataset.view()
N = self.N
M = self.M
counts = prev_counts
cur_centres = prev_centres.copy()
labels = d_copy[:, N - 1]
res = clustered_object()
for n in range(numIter):
likelihood = (np.log(counts) * self.r).astype(np.float32)
#idx = dict(zip(label_ids, range(K))) # dictonary mapping label id to indexes
for i in range(M):
data_tmp = d_copy[i, 1:N - 1]
old_idx = self.indices[i]
(best_key) = find_cluster_s2(
data_tmp, cur_centres, counts, likelihood
) #find max probability and return corresponding index
if best_key != old_idx:
centre_tmp = cur_centres[old_idx]
if counts[old_idx] >= 1:
# not the last point being removed from cluster
new_centre = update_centre(data_tmp,
centre_tmp,
counts[old_idx],
sign=-1)
cur_centres[old_idx] = new_centre
counts[old_idx] -= 1
# reassign to correct cluster
centre_tmp = cur_centres[best_key]
new_centre = update_centre(data_tmp,
centre_tmp,
counts[best_key],
sign=1)
cur_centres[best_key] = new_centre
counts[best_key] += 1
labels[i] = label_ids[best_key]
# remove inactive clusters
active_idx = counts > 1
label_ids, counts = label_ids[active_idx], counts[
active_idx]
l_tmp, self.indices = np.unique(d_copy[:, N - 1],
return_inverse=True)
cur_centres = cur_centres[active_idx]
if np.array_equal(cur_centres, prev_centres):
#print('number of iterations = %d' % n)
res.clustered_set = self.dataset
res.cluster_labels, res.cluster_counts = label_ids[
active_idx], counts[active_idx]
res.cluster_centres = cur_centres[active_idx]
return res
else:
prev_centres = cur_centres.copy()
res.clustered_set = self.dataset
res.cluster_labels = label_ids.astype(int)
res.cluster_counts = counts
res.cluster_centres = cur_centres
return res