Closed 2U1 closed 2 months ago
It would be better if freezing the vision backbone were the default option.
I thought it could be better for the adding --freeze_vision at the train.sh can be easier to see the option for users using this repo.
I've added the freeze_vision option at the train.sh
The lora modules are not added to the vision tower anyway though, unfreezing the vision tower will have no effect.
Will implement the settings a bit later, thanks
Unfreezing the vision backbone while fine-tuning could be helpful for when finetuning according to the following tech report. https://arxiv.org/pdf/2403.06199