What is expected Kinetics400 dataset directory structure?

[vadimkantorov]

Given that the dataset does not come with official downloader scripts and that most roll their own or hack some third-party scripts, it would be much clearer if https://pytorch.org/docs/stable/torchvision/datasets.html#kinetics-400 explained what directory structure is expected by torchvision.datasets.Kinetics400

What is the expected dataset size? and the video file extensions?

Thanks!

cc @pmeier

[vfdev-5]

@vadimkantorov maybe notes from this ref example could help: https://github.com/pytorch/vision/tree/master/references/video_classification#data-preparation

cc @bjuncek

[vadimkantorov]

This is not very helpful, since it does not mention explicitly the class-label directory organisation. It is also not clear if root_dir should contain the splits or if the root_dir is the path to a chosen split.

[vfdev-5]

Unfortunately, I removed the dataset from my storage where I tried a minimal training of the ref example.
Maybe, Bruno could share more info about that...

[vadimkantorov]

The problem is that the dataset does not have official download or dataset stats scripts or recommended directory structure. From what I found, the root_dir should be the path to the split, the split directory should contain directories for every class labels with videos inside them. But this should better be mentioned explicitly in the docs

[vfdev-5]

I agree. I had a similar question in my case... If you'd like to improve the docs, feel free to send a PR.