Update losses.py

bpnet/losses.py

@@ -26,10 +26,11 @@ def multinomial_nll(true_counts, logits):
  dist = tf.contrib.distributions.Multinomial(total_count=counts_per_example,
  logits=logits_perm)
 
- # get the sequence length for normalization
- seqlen = tf.to_float(tf.shape(true_counts)[0])
+ # Normalize by batch size. NOTE: In the paper we say that
+ # the loss is normalized by sequence length.
+ batch_size = tf.to_float(tf.shape(true_counts)[0])
 
- return -tf.reduce_sum(dist.log_prob(true_counts_perm)) / seqlen
+ return -tf.reduce_sum(dist.log_prob(true_counts_perm)) / batch_size
 
 
 @gin.configurable