def __init_kmeansPP(self, X):
distribution = torch.ones(len(X)) / len(X)
frequency = pytorch_categorical.Categorical(distribution)
centroids_index = []
N, D = X.shape
while (len(centroids_index) != self._n_c):
f = frequency.sample(sample_shape=(1, 1)).item()
if (f not in centroids_index):
centroids_index.append(f)
centroids = X[centroids_index]
x = X.unsqueeze(1).expand(N, len(centroids_index), D)
dst = self._distance(centroids, x)
value, indexes = dst.min(-1)
vs = value**2
distribution = vs / (vs.sum())
frequency = pytorch_categorical.Categorical(distribution)
self.centroids_index = torch.tensor(centroids_index,
device=X.device).long()
self.centroids = X[self.centroids_index]
alpha, beta = args.init_alpha, args.init_beta
print("Loading Corpus ")
D, X, Y = dataset_dict[args.dataset]()
print("Creating dataset")
# index of examples dataset
dataset_index = corpora_tools.from_indexable(
torch.arange(0, len(D), 1).unsqueeze(-1))
print("Dataset Size -> ", len(D))
D.set_path(False)
# negative sampling distribution
frequency = D.getFrequency()**(3 / 4)
frequency[:, 1] /= frequency[:, 1].sum()
frequency = pytorch_categorical.Categorical(frequency[:, 1])
# random walk dataset
d_rw = D.light_copy()
rw_log = logger.JSONLogger("ressources/random_walk.conf", mod="continue")
if (args.force_rw):
key = args.dataset + "_" + str(args.context_size) + "_" + str(
args.walk_lenght) + "_" + str(args.seed)
if (key in rw_log):
try:
print('Loading random walks from files')
d_rw = torch.load(rw_log[key]["file"])
print('Loaded')
except:
os.makedirs("/local/gerald/KMEANS_RESULTS/", exist_ok=True)
alpha, beta = args.init_alpha, args.init_beta
print("Loading Corpus ")
D, X, Y = dataset_dict[args.dataset]()
print("Creating dataset")
# index of examples dataset
dataset_index = corpora_tools.from_indexable(
torch.arange(0, len(D), 1).unsqueeze(-1))
print("Dataset Size -> ", len(D))
D.set_path(False)
# negative sampling distribution
frequency = D.getFrequency()**(3 / 4)
frequency[:, 1] /= frequency[:, 1].sum()
frequency = pytorch_categorical.Categorical(
torch.ones(len(frequency)) / len(frequency))
# random walk dataset
d_rw = D.light_copy()
d_rw.set_walk(args.walk_lenght, 1.0)
d_rw.set_path(True)
d_rw = corpora.ContextCorpus(d_rw,
context_size=args.context_size,
precompute=args.precompute_rw)
# neigbhor dataset
d_v = D.light_copy()
d_v.set_walk(1, 1.0)
print(d_rw[1][0].size())
print("Merging dataset")
embedding_dataset = corpora_tools.zip_datasets(