def cluster(self, data, n_clusters):
n_samples, _ = shape(data)
assert self.n_ants < n_samples, "number of ants must be lower than number of samples"
bestscore = float('inf')
bestcentroids = None
bestweights = None
pheromone = self.t0 * ones((n_samples, n_clusters))
for it in range(self.n_iter):
for _ in range(self.n_ants):
# memory = -1 * ones(n_samples, dtype='int')
weights = zeros((n_samples, n_clusters), dtype='bool')
centroids = array([data[randint(n_samples), :] for _ in range(n_clusters)], copy=True)
for i in permutation(n_samples):
scores = self.centroidscore(data[i, :], centroids, pheromone[i, :], self.beta)
if randfloat() < self.q0:
j = argmax(scores) # exploit
else:
j = choice(n_clusters, p=(scores / sum(scores))) # explore
weights[i, j] = True
centroids[j, :] = average(data, axis=0, weights=weights[:, j])
currentscore = score(data, centroids=centroids, norm=self.beta)
if currentscore < bestscore:
bestscore = currentscore
bestcentroids = copy(centroids)
bestweights = copy(weights)
pheromone = (self.ro * pheromone) + ((1.0 / bestscore) * bestweights)
if it % self.printfreq == 0:
print "Ant Colony iteration", it, "best score:", bestscore
return getlabels(data, centroids=bestcentroids, norm=self.beta)
def getnewcentroids(self, data, locations, k):
a = randint(self.n_bees)
theta = randfloat(-1.0, 1.0)
newcentroids = locations[k, :, :] + (theta * (locations[k, :, :] - locations[a, :, :]))
newscore = score(data, centroids=newcentroids, norm=self.norm)
return newcentroids, newscore
