You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi! Thanks for the opensource code. I have the doubts about masking strategy.
In the paper: Uniformly masking stage-1 input tokens from the H/4 × W/4 featuremaps would cause all tokens of stage-3 to have partially visible information and requires keeping all stage-3 tokens. Why the visible information will pass to the stage-3, if the images was masked in the first.
Thanks very much!
The text was updated successfully, but these errors were encountered:
Hi! Thanks for the opensource code. I have the doubts about masking strategy.
In the paper: Uniformly masking stage-1 input tokens from the H/4 × W/4 featuremaps would cause all tokens of stage-3 to have partially visible information and requires keeping all stage-3 tokens. Why the visible information will pass to the stage-3, if the images was masked in the first.
Thanks very much!
The text was updated successfully, but these errors were encountered: