Why not use FAPE loss? about diffab HOT 3 CLOSED

@luost26 Hi, shitong, I experiment the fape loss and observe that the loss and gradient go nan after one step of gradient descent. You may find out the similar issue and thus give up using fape in your model right?
Would you like to talk about a bit of your experiences about this?
I will keep this issue once I find out the solution or the reason.

from diffab.

We have many internal versions, including FAPE loss-based. All of them are trained successfully and we did not encounter NaN issues. Please double check.

from diffab.

Thanks for the information. I use OpenFold's implementation of FAPE. The problem maybe be it. If fape works, why your paper version does not use it? Because the performance is lower than other versions?

from diffab.