Can't reproduce the results of lisa-7b

[L1uShuai]

Has anyone managed to replicate the outcomes of lisa-7b? I'm puzzled as to why, after ten epochs of training, my test results for giou are 0.05 and for ciou are 0.03.

[L1uShuai]

I've employed 6 GPUs, and the llava version I'm using is LLaVA-Lightning-7B-delta-v1-1. And the rest of the settings have been left as is.

[wxliii]

Have you solved this problem?

[L1uShuai]

Have you solved this problem?

Up until now, I haven't been able to resolve this issue.

[chenqi1126]

Maybe you should add the delta (LLaVA-Lightning-7B-delta-v1-1) to the original LLaMA weights to obtain the LLaVA weights (LLaVA-Lightning-7B-v1-1). Refer to the legacy models part in https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md