HAN的attention里为什么加reduce_sum和reduce_max？

[fengdoudou1895]

在HAN的attention里面看到：
attetion_logits = tf.reduce_sum(hidden_state_context_similarity,axis = 2) attention_logits_max = tf.reduce_max(attention_logits, axis = 1,keep_dims = True) p_attention = tf.nn.softmax(attetion_logits-attention_logits_max)
原论文里没看到这个操作，请问这是为什么呢？

[acadTags]

This is because of the fact that softmax is shift-invariant by a constant offset in the input.

softmax is invariant under translation by the same value in each coordinate
See wikipedia and a StackOverflow answer.

Deducing the maximum value in the attention_logits allows a faster and more stable numerical computation.