Closed RylieWeaver closed 2 months ago
Note: I changed the MAE threshold for MFC because it was failing with 0.2571 > 0.2 ,and assumed this was cause by just higher losses rather than a structural error.
It looks good to me. @allaffa and @pzhanggit how do you think?
@pzhanggit Is this good to merge?
This PR adds the PNAPlus stack I've recently created and experimented with. It replicates DimeNet's use of Bessel Functions and Element-Wise product of "message_ij" in its message propagation. In my intuition, and according to some small experiments, this is more effective than raw concatenation as can be done with edge features. Otherwise, this operates the same as regular PNA.
This is an Invariant Stack.