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]

No description provided.

[ZhenbangDu]

应该是这样的，看源代码是这样

[KingSpencer]

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.