Is logits_bias trainable?

[audreycs]

Hello, I am confused about the variable logits_bias. It seems that you make logits_bias not trainable in your code. (At line 170 in memn2n_kv.py)

automated-essay-grading/memn2n_kv.py

Line 170 in cda24d8

logits_bias = tf.get_variable('logits_bias', [score_range])

Normally, shouldn't bias be trainable?

Other possible mistakes I found:
1.

automated-essay-grading/cv_train.py

Line 191 in cda24d8

sent_size = sent_size_list.pop(score_idx)

Here sent_size_list is the list of all data(train+test), but score_idx is the index of train data.

automated-essay-grading/memn2n_kv.py

Line 177 in cda24d8

cross_entropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=tf.cast(self._labels, tf.float32), name='cross_entropy')

why here logits are used to calculate cross entropy, but not the softmax result of logits.

Thanks!