kmeans_analysis_cifar10.py

from fuel.datasets.cifar10 import CIFAR10
from sklearn.cluster import KMeans
from collections import Counter
from scipy.stats.mstats import mode

import numpy as np

def PerformKMeansAnalysis(train_data, train_labels, test_data, test_labels):
    naive_grouping = KMeans(n_clusters=10, n_init=5, n_jobs=8)
    naive_grouping.fit(train_data)
    group_labels = naive_grouping.labels_

    for label in set(group_labels):
        train_mask = group_labels == label
        real_labels = train_labels[train_mask]
        pos_stat = Counter(real_labels).values()
        pos_stat = np.array(pos_stat).astype(np.float)
        pos_stat /= real_labels.shape[0]

        representative = mode(real_labels)[0]
        gini_index = np.dot(pos_stat.T, np.ones(pos_stat.shape[0]) - pos_stat)
        print "For label %d: %s" % (label, gini_index)
        print "Representative: ", representative
        print "======"