Have you tuned the vision model in sft stage?
xmy0916 opened this issue · 2 comments
xmy0916 commented
I found code here: https://github.com/X-PLUG/mPLUG-Owl/blob/main/mPLUG-Owl2/scripts/finetune.sh#L32
you have freezed the vision backbone, but according to your paper, it seems all the parrams have been trained?
MAGAer13 commented
Yes, this is just reimplementation of our model. You can enable it during training and you should implement layer-wise learning rate by yourself.
xmy0916 commented
Yes, this is just reimplementation of our model. You can enable it during training and you should implement layer-wise learning rate by yourself.
Thanks for your great job and answer~