calculating the threshold when applying TSA

[shamanez]

In this line you have taken the negative exponential of the supervised cross-entropy loss. Can you please explain why you haven't used a softmax output as the confidence of the prediction?

[SanghunYun]

Hello shamanez!

The TSA formula is as follows

and sup_loss in torch.exp(-sup_loss) is as follows

So, I calculated prob through torch.exp(-sup_loss)
Of course you can use softamx function like line89. But, it is so complex that i used line88