About diff

[MTandHJ]

Hi, I found the implementation for KL loss

        loss_diff = self.diff(scoresI, scoresS, score_adv2, score_adv1, diff_mask)

is different the Equation (9), where the code shall be

        loss_diff = self.diff(scoresI, scoresS, score_adv1, score_adv2, diff_mask)

Moreover, I am confused about the role of diff_mask. It will be helpful if some references could be given.

Thanks !

[xiaxin1998]

Just read the paper carefully and you will know why.