[video] exposing _backend in datasets

[bjuncek]

A simple PR that exposes the new experimental backend for training purposes. Using it allows for an overall speedup in multiGPU setups.

Specifically, I found the new backend to be:

1. on average 0.2s faster per iteration for equivalent model
2. on average 7min (stdev: 3) minutes faster per training epoch

The tests were done using 64 GPU setup across 8 machines, for 4 different models.

[fmassa]

Thanks for the PR @bjuncek !

Can you mention how much relative speed up we got from changing backends? Like, training went from 1.2s / iter to 1.0 s / iter?

Also, I'm thinking that I'd prefer to avoid having to pass a _backend to each function that uses a video_reader, but my ideal preference would be to just let VideoClips use the torchvision.set_video_backend internally.

Thoughts?

[bjuncek]

Can you mention how much relative speed up we got from changing backends? Like, training went from 1.2s / iter to 1.0 s / iter?

For the smallest difference: training on average went from 1.17s/iter to 0.94s/iter. For the largest one it went from 1.20s/iter to 0.88s/iter.

The differences depend on a) gpu/cpu ratio, b) model c) num_workers hyperparameter.

---

but my ideal preference would be to just let VideoClips use the torchvision.set_video_backend internally

I agree, but wouldn't that cause a rather big overhaul?

[fmassa]

Subsumed by #1376

I agree, but wouldn't that cause a rather big overhaul?

I don't think this is going to require that many changes, and IMO seems better. Let me know if you disagree