chenjoya
commented
3 days ago Hi kffeng, sorry for the delayed response.
So you can just
- download the Ego4D goalstep videos
- get the video ids of Ego4D goalstep videos
- select the conversations under the video id in https://huggingface.co/datasets/chenjoya/videollm-online-chat-ego4d-134k/blob/main/goalstep_livechat_trainval_filtered_21k.json
- follow the frame preprocessing & encoding for Ego4D goalstep videos
- now you should have .pt files and .json files. It is okay that you can start to train.
As Ego4D goalstep videos are just a few videos, I estimated that you can finish the above in one day.
Please feel free to email me joyachen@u.nus.edu if you encounter some problems and I did not reply in Github. Many thanks for your patience!
Thank you for your work, because the dataset is so big, I would like to make my own small dataset to run through your code to run videollm-online/scripts/ego4d/live1+.sh. I'm only using one dataset ego4d_goalstep_livechat_trainval I am wondering what is under this directory? Is it the processed .pt files and .json files? Or is it something else