def get_centroids(data_pnts,k): centroids = (utils.init_centroids(data_pnts,1).tolist()) # rest of the centroids are determined by how close they are to existing centroids for i in range(1,k): centroids.append(new_center(centroids,data_pnts)) # Finished initializing, continue with regular k means # print centroids centroids = np.array(centroids) return centroids
def main():
datafile = "toydata.txt"
k = 3
# Read in file
data_pnts = utils.read(datafile)
centroids = utils.init_centroids(data_pnts,k)
rv = run(centroids,data_pnts,k,"cost.png","k_means.png")
# run through k_means ~ 19 times to account for random init
for i in range(19):
centroids = utils.init_centroids(data_pnts,k)
temp = run(centroids,data_pnts,k,"cost.png")
rv = np.concatenate((rv,temp),axis = 0)
plt.figure()
for j in rv:
plot_cost(j[0],j[1])
plt.savefig("cost.png")
def initialize(data_pnts,k):
'''Randomly select points to serve as means
Covariance matrix equals covariance of full training set
Each cluster has equal prior probability.
Dataset is equally divided among clusters
Each cluster has a mean, a probability of coming from it,
and a covariance matrix of the zth gaussian
'''
# randomly find k gaussians from n datapoints, with equal prob
means = utils.init_centroids(data_pnts,k)
# probability is uniform, fill pi
pi = [1/float(k) for i in range(k)]
# covar matrix initialized as identity matrix
# initialize it as a 2x2 matrix
covar = [np.identity(2) for j in range(k)]
# covar = [np.dot(means[j] for j in range(k)]
return (means,pi,covar)
import scipy.io as sio
import numpy as np
import matplotlib.pyplot as plt
from utils import init_centroids, get_point_centroid_indices, compute_centroids, calculate_cost
if __name__ == "__main__":
# Load data from the file
data = sio.loadmat('k_means_clustering/data/dataset.mat')
X = np.matrix(data['X'])
# Initial setup
K = 3
max_iterations = 20
max_clusters = 10
centroids = init_centroids(X, K)
# Elbow method (calculate optimal number of clusters)
costs = np.zeros((max_clusters, 1))
for c in range(1, max_clusters + 1):
centroids = init_centroids(X, c)
# Iterate through the centroids
for i in range(0, max_iterations):
indices = get_point_centroid_indices(X, centroids)
centroids = compute_centroids(X, indices, c)
costs[c - 1, 0] = calculate_cost(X, indices, centroids)
# Plot the cost graph
plt.plot(np.arange(1, max_clusters + 1), costs, c='r', marker='o')
plt.grid()
plt.xlabel('Number of clusters')
