def method_comparison_plot(data, true_labels, k_means_centers, lbg_centers):
k = len(k_means_centers)
l = len(lbg_centers)
plot_data = data + k_means_centers + lbg_centers
D = distances.chordal_distance(plot_data, plot_data)
embed = distances.mds(D)
plt.figure()
for i in np.unique(true_labels):
idx = np.where(true_labels == i)[0]
plt.plot(embed[idx, 0], embed[idx, 1], 'o', label='Class %i' % i)
plt.plot(embed[-l:, 0],
embed[-l:, 1],
'o',
markersize=8,
markeredgecolor='k',
label='LBG Centers')
plt.plot(embed[-(k + l):-l, 0],
embed[-(k + l):-l, 1],
'o',
markersize=8,
markeredgecolor='k',
label='K-means Centers')
plt.legend()
plt.title('Method Comparison Plot')
plt.show()
def embed_plot_results(data, centers, true_labels, eigplot=False):
l = len(data)
plot_data = data + centers
pairwise_dist = distances.chordal_distance(plot_data, plot_data)
embed_coords = distances.mds(pairwise_dist, eigplot=eigplot)
plt.figure()
for i in np.unique(true_labels):
idx = np.where(true_labels == i)[0]
plt.plot(embed_coords[idx, 0],
embed_coords[idx, 1],
'o',
label='Cluster %i' % i)
plt.plot(embed_coords[l + i, 0],
embed_coords[l + i, 1],
'o',
markeredgecolor='k',
markersize=8,
label='Center %i' % i)
plt.legend()
plt.title('Grassmann LBG Results')
plt.show()
def fit(self,
data,
true_labels=None,
supervised=False,
show_cluster_data=True,
center_count=1,
plot_results=False,
eigplot=True,
distortion_plot=True,
numits=10):
if not supervised:
if self.center_select == 'data':
centers = random.sample(data, center_count)
elif self.center_select == 'random':
centers = []
for i in range(center_count):
centers.append(np.linalg.qr(np.random.rand(data.shape))[0])
else:
print(
"Invalid center selection option. Please choose 'data' or 'random'"
)
return
count = 0
self.center_updates.append([centers])
dist = distances.chordal_distance(centers, data)
labels = np.argmin(dist, axis=0) # should be MIN in each column
self.label_change.append([labels])
avg_dist = cluster_distortion(dist, labels, center_count)
self.distortion_change.append(avg_dist)
delta = 1
while count < numits and delta > self.eps:
# Recalculate centers
centers = []
for i in range(center_count):
idx = (labels == i).nonzero()[0]
cluster_subset = []
for q in range(len(idx)):
cluster_subset.append(data[idx[q]])
centers.append(flag_mean.flag_mean(cluster_subset))
count += 1
self.center_updates.append([centers])
dist = distances.chordal_distance(centers, data)
labels = np.argmin(dist,
axis=0) # should be min in each column
self.label_change.append([labels])
avg_dist = cluster_distortion(dist, labels, center_count)
if self.verbosity > 0:
print('Iteration %i cluster distortion: %.8f' %
(count, avg_dist))
delta = np.abs(self.distortion_change[-1] -
avg_dist) / avg_dist
self.distortion_change.append(avg_dist)
# Calculate final post-iteration stuff
print('LBG terminated after %i iterations \n' % count)
else:
pass # gotta add supervised version later
if plot_results:
embed_plot_results(data, centers, labels, eigplot=eigplot)
if show_cluster_data:
print_cluster_data(centers, labels, true_labels)
if distortion_plot:
plt.figure()
plt.plot(self.distortion_change)
plt.xlabel('Iteration')
plt.ylabel('Average distortion')
plt.title('Distortion Change')
return centers, labels
def fit(self,
data,
true_labels=None,
supervised=False,
show_cluster_data=True,
center_count=1,
plot_results=False,
eigplot=True,
distortion_plot=True,
numits=1):
'''
'''
if not supervised:
if self.center_select == 'data':
centers = random.sample(data, center_count)
elif self.center_select == 'random':
centers = []
for i in range(center_count):
centers.append(np.linalg.qr(np.random.rand(data.shape))[0])
else:
print(
"Invalid center selection option. Please choose 'data' or 'random'"
)
return
count = 0
self.center_updates.append([centers])
dist = distances.chordal_distance(centers, data)
labels = np.argmin(dist, axis=0) # should be MIN in each column
self.label_change.append([labels])
avg_dist = cluster_distortion(dist, labels, center_count)
self.distortion_change.append(avg_dist)
delta = 1
n = np.zeros((1, center_count))[0]
while count < numits and delta > self.eps:
if self.verbosity > 1:
print('Begin epoch %i...' % (count + 1))
for i in range(len(data)):
self.center_updates.append([centers])
dist = distances.chordal_distance(centers, [data[i]])
label = np.argmin(
dist, axis=0)[0] # should be min in each column
self.label_change.append([label])
n[label] += 1
centers[label] = distances.geodesic(
centers[label], data[i], 1 / (n[label]))
# Calculate distortion after a single epoch
dist = distances.chordal_distance(centers, data)
labels = np.argmin(dist, axis=0)
avg_dist = cluster_distortion(dist, labels, center_count)
delta = (self.distortion_change[-1] - avg_dist) / avg_dist
self.distortion_change.append(avg_dist)
count += 1
if self.verbosity > 0:
print('Epoch %i cluster distortion: %.8f' %
(count, avg_dist))
# Calculate final post-iteration stuff
print('Kmeans terminated after %i iterations \n' % count)
self.center_updates.append([centers])
dist = distances.chordal_distance(centers, data)
labels = np.argmin(dist, axis=0) # should be min in each column
self.label_change.append([labels])
avg_dist = cluster_distortion(dist, labels, center_count)
self.distortion_change.append(avg_dist)
else:
pass # no supervised version yet
if plot_results:
embed_plot_results(data, centers, labels, eigplot=eigplot)
if show_cluster_data:
print_cluster_data(centers, labels, true_labels)
if distortion_plot:
plt.figure()
plt.plot(self.distortion_change)
plt.xlabel('Iteration')
plt.ylabel('Average distortion')
plt.title('Distortion Change')
return centers, labels