-
Hi, thanks for your great work. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
In case anyone else reads this, for swin and other transformer arch with fixed or contrained image sizes, it's best to resize/crop/pad the image in preprocessing to fit the network input. That's already handled by default eval transforms for the validation/train scripts here.. but it's not hard to customize for padding or different types of resize (non-aspect preserving, etc). |
Beta Was this translation helpful? Give feedback.
In case anyone else reads this, for swin and other transformer arch with fixed or contrained image sizes, it's best to resize/crop/pad the image in preprocessing to fit the network input. That's already handled by default eval transforms for the validation/train scripts here.. but it's not hard to customize for padding or different types of resize (non-aspect preserving, etc).