Info Loss on noisy scores

[apurvakokate]

Hi,

Can you shed some light on why you use

loss, loss_dict = self.loss(att, clf_logits, data.y, epoch)

and not loss, loss_dict = self.loss(att_log_logits, clf_logits, data.y, epoch)

Although the noise acts as stochastic attention, the noise may change the scores significantly, making the KLD with the Q distribution meaningless to tune MLP output.

Thanks and Regards