Bowen Yuan

GPU memory rises continuously

1

When I finetune the ALBEF on VQA task, I find that the GPU memory rises continuously. It ends up with 1.5 times as much GPU memory as it started with....

How to generate instructions automatically using GPT-4?

3

I noticed that the video dataset such as NextQA has many different instructions for each sample, how can I create this? (in the provided files in `Ask-Anything/video_chat2/DATA.md`)

stage 3 cpkt for zero-shot NextQA

1

Hello, could you please release the stage 3 checkpoint for zero-shot NextQA, which in the paper is obtained by performing stage 3 instruction tuning without the NextQA dataset?

Some files for VQA v2 may be missing..

1

When I trian on VQA v2 by default: CUDA_VISIBLE_DEVICES=$1 python src/vqa.py \ --train train \ --valid train \ --test minival,nominival \ It reported: No such file or directory: './train.json' I...

llava-video使用llms-eval测试出错

5

按照evaluation部分，目前的llava好像已经没有llava_vid， 在lmms-eval下面好像也有类似的错误 [#242 in lmms-eval](https://github.com/EvolvingLMMs-Lab/lmms-eval/issues/242) 想问一下有没有什么快速的解决方案。。如果自己进行适配的话该怎么操作。。

我注意到llava-video 7B上用到的video表示方式是(64, 679, 1, 2)，每个frame是679个token，但是在ov上是729个，我想问一下这其中有什么细节上的修改吗

Do not support for finetuning on only one dataset..

1

It seems `./dataset/__init__.py` can not support for enabling only one dataset in `instruction_data.py`... How can I fix this ...

Your paper has inspired me a lot! After reading the paper, I have some questions and hope you can reply to me! 1. Will the situation observed in Fig.2 appear...