[Suggestion] creating a standalone general model only repository or be part of e3nn.

[luwei0917]

Thanks for the great work! This repository is great for those want to reproduce your results. But the equiformer architecture itself is probably more general and versatile that those dataset your tested on. The new "linear, layer norm, DTP and the whole equiformer" modules are probably useful for others wanting to build upon e3nn as well. It will be great if you could separate our the architecture or maybe make it part of the e3nn package.
After a brief check on your nets folder, it seems that the equiformer module is already mostly disentangled from the task specific modification. I guess with slightly more effort to completely separate out the task specific parts and create into a standalone repository or integrate into the e3nn package, this work could have bigger impact.

[mariogeiger]

To answer from the e3nn side, I would be happy to merge such a PR in e3nn

[yilunliao]

Hi @luwei0917 .
Thanks or your suggestion.

Let me discuss this with Mario and Tess.
I am considering including an example of Equiformer in e3nn instead.