def fit(self, samples):
"""
Apply Spectral Clustering algorithm as described in Ng et al. 2002
Affinity matrix calculated with scaling parameter as described in Zelnik-Manor et al. 2005
Clustering algorithm used: custom simple KMeans
:param samples: data samples to cluster
:return: labels associated with cluster
"""
self.samples = samples
self.samples_len = len(samples)
# Compute affinity matrix (A)
affinity = self._affinity_matrix()
# Square root of diagonal matrix (D) composed of the sum of each of A's rows => D^1/2
d = np.diag(np.power(np.sum(affinity, axis=0), -1 / 2))
# Compute laplacian matrix (L) using formula L = D^1/2 . A . D^1/2
laplacian = d @ affinity @ d
# Eigenvectors of L stacked as a matrix (X)
_, eig_vecs = sp.sparse.linalg.eigs(laplacian, k=self.k)
# Normalize X using formula X / sum(X^2)^1/2 which gives us a data sample representation (Y)
normalized_eig_vecs = eig_vecs / np.linalg.norm(
eig_vecs, axis=1, keepdims=True)
# Fit a KMeans algorithm to Y and receive cluster labels
kmeans = KMeans(k=self.k)
return kmeans.fit(normalized_eig_vecs)
import numpy as np
import matplotlib.pyplot as plt
from unsupervised.kmeans import KMeans
n_samples = 100
n_features = 2
X = np.random.rand(n_samples, n_features)
inertias = []
for k in range(1, 10): # len(X)):
kmeans = KMeans(k=k)
kmeans.fit(X)
print(kmeans.inertia_)
inertias.append(kmeans.inertia_)
plt.figure(figsize=(10, 20))
plt.plot(range(1, len(inertias) + 1), inertias)
plt.show()
parser.add_argument('--center', type=int, help='Number of data centers.', default=3)
parser.add_argument('--random_state', type=int, help='Random state for data generation.', default=42)
parser.add_argument('--n_samples', type=int, help='Number of data points.', default=5000)
args = parser.parse_args()
# Setting parameters
max_iterations = args.max_iter
n_centers = args.center
n_samples = args.n_samples
random_state = args.random_state
# Create the clusters
X, y = make_blobs(n_samples=n_samples, centers=n_centers, n_features=2, random_state=random_state, cluster_std=1.5)
# Clustering
kmeans = KMeans(k=n_centers, iterations=max_iterations, random_state=random_state, track_history=True)
kmeans.fit(X)
# Extract centroids
centroids = kmeans.history_centroids
# Create decision boundary data
h = .1
y_min, y_max = X[:, 1].min() - 1, X[:, 1].max() + 1
x_min, x_max = X[:, 0].min() - 1, X[:, 0].max() + 1
xx, yy = np.meshgrid(np.arange(x_min, x_max, h), np.arange(y_min, y_max, h))
area_data = np.c_[xx.ravel(), yy.ravel()]
# Prepare predictions
predicted_labels = []
predicted_area = []
