locuslab/FLYP

ImageNet few shot experiment

Closed this issue · 2 comments

May I ask how do you initialize the classification head in ImageNet few-shot experiments when finetuning? Do you use the zero-shot head or from random initialization?

Thanks for your reply!