wta.py

import numpy as np
import timeit
import random
import math

def eucdist(a, b):
    d = 0
    if len(a) != len(b):
        return -1
    for i in range(len(a)):
        d += ((a[i] - b[i]) ** 2)
    return math.sqrt(d)

class WTA:
    k = 0
    xdata = 0
    ydata = 0

    def __init__(self, k=6):
        self.k = k
    
    def set_k(self, k):
        self.k = k

    def set_prior(self, data):
        for i in range(0, data.shape[1]):
            x = data[:,i].reshape(self.mean.shape)
            data[:,i] = ((x - self.mean) / self.sigma).reshape(x.shape[0])

    def predict(self, xs, test_labels, e = 0.01, seed = 2, norm=2):
        random.seed(seed)
        xs = xs.T
        clusters = []
        preds = [-1 for x in range(len(xs))]
        for i in range(self.k):
            cluster = []
            for j in range(xs.shape[1]):
                cluster.append(random.random())
            clusters.append([cluster, [], []])

        iters = 0
        old_us = [[0 for x in range(xs.shape[1])] for y in range(self.k)]

        print('STARTING CLUSTERS')
        for cluster in clusters:
            print(cluster[0])
        
        for i in range(len(xs)):
            mindex = -1
            mindist = float('inf')
            for j in range(len(clusters)):
                d = eucdist(xs[i], clusters[j][0])
                if d < mindist:
                    mindex = j
                    mindist = d
            diff = np.subtract(xs[i], clusters[mindex][0])
            clusters[mindex][0] = clusters[mindex][0] + (e * diff)

        for i in range(len(xs)):
            mindex = 0
            mindist = float('inf')
            for j in range(len(clusters)):
                d = np.linalg.norm(xs[i] - clusters[j][0], ord=norm)
                print(d)
                if d < mindist:
                    mindex = j
                    mindist = d
            clusters[mindex][1].append(xs[i])
            clusters[mindex][2].append(test_labels[i])
            preds[i] = mindex

        print('ENDING CLUSTERS')
        for cluster in clusters:
            print(cluster[0])

        num0s = float(len([x for x in test_labels if x == 0]))
        num1s = float(len([x for x in test_labels if x == 1]))
        finals = zip(xs, test_labels)
        y = [x[0] for x in finals if x[1] == 1.0]
        mean1 = np.mean(y, axis=0)
        finals = zip(xs, test_labels)
        y = [x[0] for x in finals if x[1] == 0.0]
        mean0 = np.mean(y, axis=0)
        tpr = 0.0
        tnr = 0.0
        fpr = 0.0
        fnr = 0.0
        for i in range(len(clusters)):
            pred0 = len([x for x in clusters[i][2] if x == 0])
            pred1 = len([x for x in clusters[i][2] if x == 1])
            if eucdist(clusters[i][0], mean0) < eucdist(clusters[i][0], mean1):
                print('Cluster {} ({}) assigned to 0'.format(i, len(clusters[i][2])))
                assigned = 0
                tnr = pred0 / num0s
                fpr = 1 - tnr
            else:
                print('Cluster {} ({}) assigned to 1'.format(i, len(clusters[i][2])))
                assigned = 1
                tpr = pred1 / num1s
                fnr = 1 - tpr
            #print('Cluster {}: 0 = {}/{}'.format(i, pred0, (pred0 + pred1)))
            #print('Cluster {}: 1 = {}/{}'.format(i, pred1, (pred0 + pred1)))

        print('TPR: {}'.format(tpr))
        print('TNR: {}'.format(tnr))
        print('FPR: {}'.format(fpr))
        print('FNR: {}'.format(fnr))

        return preds