fengxin619

Results 10 issues of fengxin619

seq2seq_model.py 108行 需要构建特殊的输出mask,屏蔽掉句子a的影响 预测的值不用取最后sep符号的结果 因此是到-1 predictions = predictions[:, :-1].contiguous() target_mask = token_type_id[:, 1:].contiguous() --------------------- 为什么target_mask是丢掉了[CLS]位,而predictions是丢掉[SEP]位,这在计算Loss的时候不是错位了么?

in keras, i dont find the part ? tf.train.ExponentialMovingAverage .....dose this affect the mode's effect?

![image](https://user-images.githubusercontent.com/30429455/188409935-12712074-8419-4719-b3be-c5acc1667f1e.png)

Could you tell me where does the formula come from?