Closed yellowcap closed 6 months ago
Could we make this automatic and dependent on inference vs training? Seems like during training it should be always True
, and during inference there is no downside of having it always False
. Or am I not seeing something here?
Shuffling the patches is important during MAE training, but is confusing and requires extra attention when fine tuning or during inference.
We should have an automatic way of handling this, so that the downstream user does not have to worry about the shuffling at all.
Refs
https://github.com/Clay-foundation/model/blob/2a7a6e2b63ba453ab73a192f8560bd724d4f98d5/src/model_clay.py#L598
https://github.com/Clay-foundation/model/pull/118#discussion_r1451790971