search

xiaoachen98 / Open-LLaVA-NeXT

An open-source implementation for training LLaVA-NeXT.
325 stars 16 forks source link

vit not saved #5

Closed RifleZhang closed 5 months ago

RifleZhang commented 5 months ago

It seems that vit is not saved when unfreeze_mm_vision_tower is True.

xiaoachen98 commented 5 months ago

Actually, they will be saved in the final chenckpoint. And when you load from the checkpoint, the fine tuned vit will be re-loaded after initialization from the OpenAI version. Hope this will help you! If you like, please consider to give a star of our repo and give a like to our huggingface models (open-llava-next-8b).

---- Replied Message ---- | From | @.> | | Date | 06/01/2024 04:34 | | To | xiaoachen98/Open-LLaVA-NeXT @.> | | Cc | Subscribed @.***> | | Subject | [xiaoachen98/Open-LLaVA-NeXT] vit not saved (Issue #5) |

It seems that vit is not saved when unfreeze_mm_vision_tower is True.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.***>

RifleZhang commented 5 months ago

Yes. It is in saved tensors. Thanks!