search

facebookresearch / ov-seg

This is the official PyTorch implementation of the paper Open-Vocabulary Semantic Segmentation with Mask-adapted CLIP.
Other
676 stars 61 forks source link

FREEZE_AT parameter in config #8

Closed goncayilmaz closed 1 year ago

goncayilmaz commented 1 year ago

Hello,

I was looking into the implementation of OVSeg and I cannot see where you used the cfg.BACKBONE.FREEZE_AT. I see that you initialize frozen_stages=-1 in SwinTransformer. Shouldnt it be initialized with this cfg.BACKBONE.FREEZE_AT value?

Jeff-LiangF commented 1 year ago

Hi @goncayilmaz, Thanks for pointing it out. We basically follow the setting FREEZE_AT==0 in MaskFormer. I checked out the swin-transformer in detectron2, looks like the arg cfg.BACKBONE.FREEZE_AT is not correctly fed into the swin-transformer initialization. I believe you are right, there is a bug. However, I conjecture (not verified) whether freezing or not would only have slight differences for the final performance.

I close this issue. Feel free to reopen it if it doesn't solve your question.