maciejskorski

def SparseCategoricalCrossentropy(labels,logits):
  ''' labels: shape [n_batch] contains true classes as numbers from 0 to n_classes-1
      logits: shape [n_batch,n_classes], predicted log probabilities '''
  Z = tf.reduce_logsumexp(logits,axis=-1)
  lookup_labels = tf.stack([tf.range(tf.shape(labels)[0]),tf.cast(labels,tf.int32)],1)
  true_logits = tf.gather_nd(logits,lookup_labels,batch_dims=0)
  return -true_logits + Z

maciejskorski

from itertools import islice,chain
from collections import deque

def gen_skipgrams(itr,window=1,symmetric=False,Q=None):
  itr = iter(itr)
  if not Q:
    Q = deque(islice(itr,window-1),maxlen=window)
  append = Q.append
  for i in itr:
    for j in Q: