THUDM / CogVLM

a state-of-the-art-level open visual language model | 多模态预训练模型
Apache License 2.0
5.68k stars 390 forks source link

Eva CLIP model: trained or freezed? #81

Closed shkarupa-alex closed 6 months ago

shkarupa-alex commented 8 months ago

Did you trained eva clip model weights or freezed them and trained just mlp adapter?

mactavish91 commented 8 months ago

It was frozen during the pre training phase, but we trained its parameters during the instruction tuning phase.

xmy0916 commented 8 months ago

@mactavish91 hello,看你们的论文这部分 image 想请教下,这四个阶段的训练,那个阶段训练了vit的参数呢?

paulpaul91 commented 1 month ago

It was frozen during the pre training phase, but we trained its parameters during the instruction tuning phase.

为什么在pretrain阶段freeze图像参数,有没有做过freeze和unfreeze的消融实验