Question: about how to fine-tune ulip to realize modelnet40 classification (not zero-shot classification)
Closed this issue · 2 comments
xzExertion commented
944104439 commented
i guess just a simple full finetune
Tycho-Xue commented
Yeah, just initialize the model with ULIP pre-trained weights and possibly add the classification head. empirically I found that you need to lower the lr to get the optimal performance compared training from scratch, which makes sense since the model has been pre-trained and gained some knowledge.