kmeansinit.py

import numpy as np
import matplotlib.pyplot as plt
import matplotlib.cm as cm

from sklearn.cluster import KMeans
from sklearn.utils import check_random_state
from sklearn.datasets.samples_generator import make_blobs

RUNS = 1
_CLUSTERS = 4

T_INIT_RANGE = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
                         11, 12, 14, 15, 16, 17, 18, 19, 20])

def kmeansinittest():
    """This script runs the test to determine
    the effect on increasing number of random inits"""

    initfig = plt.figure()
    plots = []
    inertia = np.empty((len(T_INIT_RANGE), RUNS))

    for run in range(RUNS):
        data = make_blobs(n_samples=200, n_features=2, centers=_CLUSTERS,
                          cluster_std=1)
        for i, t_init in enumerate(T_INIT_RANGE):
            km = KMeans(n_clusters=_CLUSTERS, init='random', n_init=t_init,
                        max_iter=300, algorithm='full')
            km.fit(data[0])
            inertia[i, run] = km.inertia_
    errorp = plt.errorbar(T_INIT_RANGE, inertia.mean(axis=1), inertia.std(axis=1))
    print(inertia.mean(axis=1), inertia.std(axis=1))
    plots.append(errorp)

    plt.xlabel('n_init')
    plt.ylabel('inertia')
    plt.title("Mean inertia for various k-means init across %d runs" % RUNS)

    h = .02
    x_min, x_max = data[0][:,0].min()-1, data[0][:,0].max()+1
    y_min, y_max = data[0][:,1].min()-1, data[0][:,1].max()+1
    xx, yy = np.meshgrid(np.arange(x_min, x_max, h), np.arange(y_min, y_max, h))
    Z = km.predict(np.c_[xx.ravel(), yy.ravel()])
    Z = Z.reshape(xx.shape)

    centroids = km.cluster_centers_
    fig2 = plt.figure()
    plt.subplot(111)
    plt.imshow(Z, interpolation='nearest',
           extent=(xx.min(), xx.max(), yy.min(), yy.max()),
           aspect='auto', origin='lower', cmap='Pastel2')
    plt.scatter(data[0][:,0], data[0][:,1], marker='o', c=data[1][:], s=25, cmap='Set1')
    plt.scatter(centroids[:,0], centroids[:,1], marker="x", c="w", s=150, linewidths=3)
    plt.show()

kmeansinittest()

Outputs means and standard deviations in console for inspection, and plt outputs should look something like this (error vs n_init, and final plot/fit)