Question regarding the dataset layout

[meifish]

Hi!

I tried to train the model, and just a tiny question regarding the structure of the --root_frames directory. As instructed in README, I created a folder frames and inside the directory, each subfolder is named by the video id, i.e., the integer video id, and contain the break-down frames of the video, i.e., 0001.jpg. Then this directory ./frames/ is used for the root_frames argument.

However, when I started training, it doesn't seem to be able to find the frames. The error is like: [WinError 3] The system cannot find the path specified: './toy_dataset/frames/frames/67698'

It seems coming from data_loader_frames.py line 115, where a literal frames/ is hardcoded in the path that causes the double naming. I wonder what is the dataset structure of the root_frames should be like?

Thanks!

[tjinjin95]

hi, I have the same problem. Have you solved it? @meifish @joaanna @xiaolonw

[meifish]

Line 115 in data_loader_frames.py, there is a literal `frames/' in the path, remove that, and also check if other path has similar issues.

[haozheng-visioner]

hi @meifish @xiaolonw @joaanna , I am confused how to create the root_frames directory. Should the --root_frames directory be the original something-something v2 video dataset?

[meifish]

@HaoShawnZheng :

Not exactly, you need to extract the frames from the original videos first, so each videos will be broken down into N frames. The directory will look like:

| Root project folder -->| Frames --------> | 67698 -----------------> | 0001.jpg -----------------> 0002.jpg -----------------> 0003.jpg

Your --root_frame argument should be "/Root_project_folder/frames/". If this doesn't work, try "/Root_project_folder/". In my last attempt around a year ago, the "/frames/" part was harcoded in the data_loader_frames.py.

Just try run it, and see what you get from the error.