google-research/l2p

请问您的模型是否是一个freeze的vit提取query feature,然后再用相同的Vit进行继续训练?

LemonQC opened this issue · 2 comments

请问您的模型是否是一个freeze的vit提取query feature,然后再用相同的Vit进行继续训练?

应该是这样的,看源代码是这样

Thanks for Yuxian's answer. Exactly, we use the same pretrained ViT copy for query feature extraction, and then we use another copy (starting from the same initialized weights) for training.