I wonder why do log_softmax twice on the prediction logits?

[su]

As depicted in code bellow .

san_mrc/src/san.py

Line 105 in 1deb1b3

 start_scores = torch.log(start_scores) 

.

san_mrc/src/model.py

Line 92 in 1deb1b3

 loss = F.cross_entropy(start, y[0]) + F.cross_entropy(end, y[1]) 

.

[su]

Is this a special procedure for SAN module?

[Xiaodong]

It is only to make the api consistence, which is output of logits.