def nc05(img,
n_clusters=5,
n_neighbors=5,
gamma=1,
affinity='nearest_neighbors',
visualize=True,
include_spatial=False):
"""
Normalized cut algorithm for image segmentation
include_spatial : (Bonus)
"""
img = resize(img, (int(img.shape[0] * 0.3), int(img.shape[1] * 0.3)),
anti_aliasing=True)
#img = imresize(img, 0.3) / 255
n = img.shape[0]
m = img.shape[1]
if include_spatial:
xx = np.arange(n)
yy = np.arange(m)
X, Y = np.meshgrid(yy, xx)
img = np.concatenate((Y.reshape(n, m, 1), X.reshape(n, m, 1), img),
axis=2)
print("spectral_segment(:include_spatial) img.shape = {}".format(
img.shape))
img = img.reshape(-1, img.shape[-1])
# Notes:
# gamma is ignored for affinity='nearest_neighbors'
# n_neighbors is ignore for affinity='rbf'
# n_jobs = -1 means using all processors :D
t0 = time.time()
labels = SpectralClustering(n_clusters=n_clusters,
affinity=affinity,
gamma=gamma,
n_neighbors=n_neighbors,
n_jobs=-1,
eigen_solver='arpack').fit_predict(img)
t1 = time.time()
labels = labels.reshape(n, m)
if visualize == True:
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1)
ax.imshow(labels)
plt.axis("off")
plt.show()
plt.imsave("ns05.png", labels)
print('Time taken to nc segment {} size image is: {}'.format(
img.shape, t1 - t0))
return (labels)
def spectral_segment(img,
n_clusters=5,
n_neighbors=5,
gamma=1,
affinity='nearest_neighbors',
visualize=False,
include_spatial=False):
"""
Normalized cut algorithm for image segmentation
include_spatial : (Bonus)
"""
# img = resize(img, (int(img.shape[0] * 0.3), int(img.shape[1] * 0.3)), anti_aliasing=True)
img = imresize(img, 0.3) / 255
n = img.shape[0]
m = img.shape[1]
if include_spatial:
xx = np.arange(n)
yy = np.arange(m)
X, Y = np.meshgrid(yy, xx)
img = np.concatenate((Y.reshape(n, m, 1), X.reshape(n, m, 1), img),
axis=2)
print("spectral_segment(:include_spatial) img.shape = {}".format(
img.shape))
img = img.reshape(-1, img.shape[-1])
# Notes:
# gamma is ignored for affinity='nearest_neighbors'
# n_neighbors is ignore for affinity='rbf'
# n_jobs = -1 means using all processors :D
labels = SpectralClustering(n_clusters=n_clusters,
affinity=affinity,
gamma=gamma,
n_neighbors=n_neighbors,
n_jobs=-1,
eigen_solver='arpack').fit_predict(img)
labels = labels.reshape(n, m)
if visualize:
plt.figure(figsize=(12, 12))
plt.axis('off')
plt.imshow(labels)
return labels
def ncut_clustering(images, k):
result = []
i = 0
for image in images:
print('clustering image ', i)
image2 = image.reshape(
-1,
3) / 255 #divide by 255 to get values between 0-1->normalization
#removes distortions
s_clustering = SpectralClustering(
n_clusters=5,
affinity='nearest_neighbors',
n_neighbors=5,
n_jobs=-1,
eigen_solver='arpack').fit_predict(image2)
result.append(s_clustering.reshape(image.shape[:-1]))
#result.append(affinity_mat.reshape(image.shape[:-1]))
i += 1
return result
def main():
with fileinput() as fs:
data = list(
set(r[4].replace('AND ', '').replace('OF ', '')
for r in reader(fs)))
vec = CountVectorizer(tokenizer=word_tokenize).fit_transform(data)
svd = TruncatedSVD().fit_transform(vec)
fig, axes = plt.subplots(2, 3)
fig.subplots_adjust(wspace=1)
for i, ax in enumerate(np.array(axes).flatten()):
clustering = SpectralClustering(n_clusters=N_CLUSTERS -
i).fit_predict(svd)
labeled = np.append(svd, clustering.reshape(len(data), 1), axis=1)
clusters = defaultdict(set)
for c, r in zip(clustering, data):
clusters[c].add(r)
pprint(clusters)
for c, titles in clusters.items():
# others = set(data) - titles
wc = wordcount(titles)
t = labeled[labeled[:, 2] == c]
ax.scatter(t[:, 0],
t[:, 1],
c=COLORS[c],
label=', '.join(w for w, _ in wc.most_common(3)))
ax.set_title(f'{N_CLUSTERS - i} Clusters')
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
# ax.legend(loc='lower center', bbox_to_anchor=(0.0, -0.3))
# plt.title('Clustering of %d Job Titles' % len(data))
# plt.show()
plt.savefig('clustering.png')
