sail-sg / volo

VOLO: Vision Outlooker for Visual Recognition

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Finetune with 512 image size

javierrodenas opened this issue · comments

Hello,

I am finetuning a model with an image size of 512 and --token-label-size 24. Is the label size enough for 512 image size? Should I use a higher label size? How do I really know the correct label size?

Thank you in advance!

The label size is related to the number of output tokens. For VOLO, if image size = 224, then the token label size should be 224/16=14. So for your case, when image size = 512, the label size should be 512/16=32.

Thanks, I close the issue.