def _build_k_prototypes(self, nprot, data):
# K-Means / K-Medoids algorithm
# should return a N-length array with he indices of the chosen data
grouper = K_Means()
grouper._centroid_mode = "index"
grouper.fit(data,
nprot,
iters=300,
mode="cuda",
cuda_mem='manual',
tol=1e-4,
max_iters=300)
centroids = grouper.centroids
nclusters = centroids.shape[0]
# TODO - very inefficient
k_labels = np.zeros(nclusters, dtype=np.int32)
for k in xrange(nclusters):
dist = data - centroids[k]
dist = dist**2
dist = dist.sum(axis=1)
k_labels[k] = dist.argmin()
return k_labels
