def train_kmeans(self, prev_score, data):
"""We are using pre-made tensorflow estimators to
train and predict.
Args:
prev_score: a sum of the distance between each sample
to their nearest center
data: list of weights of user model
Return:
updated score
"""
seed = np.random.randint(5667799881, size=1)[0]
self._kmeans = tf.contrib.factorization.KMeansClustering(
random_seed=seed,
num_clusters=self._num_clusters,
use_mini_batch=False)
# composed of weights of every client model
score = self.train_iteation(data)
# evaluate the samples to compute the distance between
# each sample and each center, forming a matrix have each
# row for each sample, and the column is distance to each
# center.
y = self._kmeans.transform(lambda: input_fn(data))
point_distance = list(y)
self._learned = np.argmin(point_distance, 1)
# removing not used files of this model
#shutil.rmtree(temp_modeldir, ignore_errors=True)
return None, score
def train_iteation(self, data):
train_data = lambda: input_fn(data)
self._kmeans.train(train_data)
cluster_centers = self._kmeans.cluster_centers()
# print("cluster centers:", cluster_centers)
score = self._kmeans.score(train_data)
return score