Design philosophy behind removal of EMA from Corrdiff

[joshdorrington]

Hi,
I see that in commit #604 EMA was removed from the training loop for the Corrdiff model. As the paper still mentions the use of EMA, which is the same training approach used in the ESD paper, would anyone be able to explain the reasoning behind dropping EMA? I looked to see if there was some discussion about it here in the github history but didn't find anything.

Thanks for any thoughts! I am applying Corrdiff to different problems than in the original paper, and so I'm trying to get the best sense of the sensitivities and rationale behind parameter choices as I can :)

[ram-cherukuri]

@pzharrington Can you share any insights here?

[pzharrington]

Yes, during refactoring it was discovered that EMA did not have a notable impact on the final results so it was removed in hopes of simplifying the recipe a bit.