Yuxuan Wang

bhys，这个proj是一年前的工作，改了好几次，有好多文件我自己也不太记得清了╥﹏╥，所以数据也没整理。 我暂时先把 train/val/test.json 传到aliyunpan，image和flow等我有空再打包，inputs下有两个文件夹，一个videoinstruct：来自于videochatgpt，val/test 就是从train里面分出来便于观察实验结果的，你可以用别的。另个一是ivinstruct：结合了videochatgpt和llava1.5的数据，val/test同上。如果还缺别的你再说，我给你找。 https://www.alipan.com/s/Az5R3sSPxZj

伪标签可以参考 https://github.com/bigai-nlco/VideoTGB/issues/8#issuecomment-2527109019 光流脚本参考 https://github.com/patrick-tssn/FlowExtractor/blob/main/extract_of.sh 至于光流文件，实在是有点大，最近应该不会搞。

Thank you for your interest! The current repository is the chat version of our work, where we've integrated the pseudo-label generation code into the framework to enable one-step execution (although...

It might take some time to organize these scripts due to my current bandwidth, sorry for any inconvenience.

it's a problem of flash-attention, wait a minute or build from the source: https://github.com/Dao-AILab/flash-attention/releases

The VideoChat2 data is a visual instruction dataset, all of which is used for training.

modify here: --model_name_or_path

Thank you for your interest. These processed links actually come from the author of VideoChat2. See this [issue](https://github.com/OpenGVLab/Ask-Anything/issues/176) for reference. If they are not available, you are advised to follow...

Sorry for any inconvenience. Following are the version details of the related packages : ``` pytorchvideo 0.1.5 pypi_0 pypi torch 2.2.1+cu118 pypi_0 pypi torchaudio 2.2.1+cu118 pypi_0 pypi torchvision 0.17.1+cu118 pypi_0...