minSSE = sys.maxint
minInitialMeans = np.zeros((k, len(points[0])))
minMeans = np.zeros((k, len(points[0])))
minLabels = np.zeros(len(points), dtype=int)
# initialize means
numOfExperiments = 5 #number of random initial means experiments
for count in xrange(numOfExperiments):
if numOfExperiments == 1:
initialMeans[:] = np.loadtxt(os.path.join(os.path.dirname(__file__), fninitcentroids))
else:
initialMeans[:] = Utils.getInitialMeans(points, k)
print "Initial Means: \n%s" % initialMeans
print "Initial SSE: %s" % Utils.calcSSE(points, initialMeans)
means = np.zeros((k, len(points[0])))
means[:] = initialMeans
finalmeans, labels, SSE = km.run(points, means, numberOfClusters=k, threshold=d, maxiterations=i)
print "final means: %s" % finalmeans
if SSE < minSSE:
print '**New min SSE %s' % SSE
minSSE = SSE
minInitialMeans[:] = initialMeans
print 'Min initial means\n%s' % initialMeans
minMeans[:] = finalmeans
import numpy as np import os from sklearn import metrics from sklearn.metrics import pairwise_distances from util.utilities import Utils fn = os.path.join(os.path.dirname(__file__), 'data/zdata.txt') print "opening file %s" % fn points = np.loadtxt(fn) #normalize using z-score or sigmoid points = Utils.zscore(points) #points = Utils.sigmoid(points) #points = whiten(points) minLabels = np.loadtxt(os.path.join(os.path.dirname(__file__), 'data/tmp/minLabels.txt')) #print 'sc kmeans2 %s ' % metrics.silhouette_score(points, idx, metric='euclidean') print 'sc kmeans user %s ' % metrics.silhouette_score(points, minLabels, metric='euclidean')
def run(self, data, means, numberOfClusters, threshold, maxiterations):
# initialize means
# means = np.array(random.sample(data, numberOfClusters))
# print "Initial Means: \n%s" % means
pointsInClusters = np.zeros(numberOfClusters)
SSE = 0
labels = np.zeros(len(data), dtype=int)
iteration = 0
total = 0
#plotting
# self.doPlots(data, labels, means, iteration, title='iteration %i:' % (iteration))
# import pdb;pdb.set_trace()
while iteration < maxiterations:
start = time()
#initialize labels for each iteration
iteration += 1
print "Iteration %d" % iteration
#Distance matrics version
#Using distance matrix calculations
# #Calculating the distance to nearest cluster
meansNew, pointsInClusters = Utils.calcNewMeans(data, means)
#calculate a new mean for each cluster
# meansNew, pointsInClusters = self.calculateMeans(data, labels, numberOfClusters)
#find nearest centroid, where line is a data vector
#feature diff feature version!
# meansNew = np.zeros((numberOfClusters, len(data[0])))
# pointsInClusters = np.zeros(numberOfClusters)
# for i in xrange(len(data)):
# mindist = sys.maxint
# minCentroid = None
# point = data[i]
# for idx in xrange(numberOfClusters):
# d = 0
# for i in xrange(len(means[idx])):
# c = means[idx]
# d += abs(point[i] - c[i])**2
#
# d = math.sqrt(d)
# if(d < mindist or minCentroid == None):
# mindist = d
# minCentroid = idx
# labels[i] = minCentroid
# meansNew[minCentroid] += point
# pointsInClusters[minCentroid] += 1
#
# for i in xrange(len(meansNew)):
# meansNew[i] = meansNew[i] / float(pointsInClusters[i])
#Point diff Point (vectorisation)!
# meansNew = np.zeros((numberOfClusters, len(data[0])))
# pointsInClusters = np.zeros(numberOfClusters)
# for i in xrange(len(data)):
# mindist = sys.maxint
# minCentroid = None
# point = data[i]
# for idx in xrange(numberOfClusters):
# d = np.sqrt(np.sum((point-means[idx])**2))
# if(d < mindist or minCentroid == None):
# mindist = d
# minCentroid = idx
# labels[i] = minCentroid
# meansNew[minCentroid] += point
# pointsInClusters[minCentroid] += 1
#
# for i in xrange(len(meansNew)):
# meansNew[i] = meansNew[i] / float(pointsInClusters[i])
#Point diff Array version (vectorisation)
# Using point and centroids calculations, like used in MR k-means
# Calculating the distance to nearest cluster and new mean
# pointsInClusters = np.zeros(numberOfClusters)
# meansNew = np.zeros((numberOfClusters, len(data[0])))
# for i in xrange(len(data)):
# point = data[i]
# d = np.sqrt(np.sum((point-means)**2,axis=1))
# minCentroidIdx = d.argmin()
# labels[i] = minCentroidIdx
# meansNew[minCentroidIdx] += point
# pointsInClusters[minCentroidIdx] += 1
#
# for i in xrange(len(meansNew)):
# meansNew[i] = meansNew[i] / float(pointsInClusters[i])
#measure calculation time
end = time()
print 'time: %f' % (end-start)
total += (end-start)
#check if the means have changed
meansDiff = 0
for i in xrange(numberOfClusters):
pprint('%s %s' % (i, meansNew[i]))
meansDiff += distance.euclidean(meansNew[i], means[i])
print 'Means difference: %f' % meansDiff
#calculate the within cluster variation, sum of squared distances between all objects in cluster and its centroid
SSE = Utils.calcSSE(data, meansNew)
print "SSE: %0.3f" % SSE
#plotting
# self.doPlots(data, labels, means, iteration, title='iteration %i meansdiff: %f' % (iteration, meansDiff))
means[:] = meansNew
if meansDiff < threshold:
break
#End of While loop
#KMeans iterative process ends here
#If mean difference under threshold or there are max iterations
if iteration == maxiterations:
print "Max iterations reached: %d" % iteration
else:
print "Means difference: %f is under threshold %f" % (meansDiff, threshold)
#print "Clusters:"
for i in xrange(numberOfClusters):
print "Cluster %d, number of points %d" % (i, pointsInClusters[i])
if self.showsubplots:
Plot.subplotClusters(data, labels, means, iteration, title='final means')
pylab.show()
print 'total time: %f' %total
return means, labels, SSE
