thuml / CLIPood

About Code Release for "CLIPood: Generalizing CLIP to Out-of-Distributions" (ICML 2023), https://arxiv.org/abs/2302.00864
MIT License
61 stars 4 forks source link

model.eval() #3

Closed Wallace-222 closed 1 year ago

Wallace-222 commented 1 year ago

if the clip architecture uses transformer, do we still need to use model.eval() to fix all the norm layers in the training code? Thanks a lot.

GXZlegend commented 1 year ago

Yes, we still use model.eval() in all experiments even if the backbone architecture is Transformer.