How 2 make FLOATER 2D?

[zobeirraisi]

[AncientRemember]

I also want to know

[AncientRemember]

as the emergence of bottleneck-transformer‘s series of paper，More people will try botnet，but the relative position embedding which used by botnet is not Inductive,this limit the generalization

The scenario I am facing is : train with smaller patch of images,but inference with bigger patch of image

[xuanqing94]

I don't know why in the text domain one needs to make FLOATER 2D. But in the image domain, a simple way is to reshape the image (in some way) into a long vector.