Fix video processing channel format

[zucchini-nlp]

What does this PR do?

FIxes failing VJEPA test. The test was failing because the video was not permuted to the correct format that torchvision expects (T, C, H, W)

[zucchini-nlp]

run-slow: vjepa

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ydshieh]

Works, thank you. Do you know what it worked before you PR? (or your PR change something regarding this channel stuff?)

[zucchini-nlp]

Do you know what it worked before you PR? (or your PR change something regarding this channel stuff?)

My PR stopped casting all inputs to a numpy format, so when a video is already a torch.Tensor it started failing. It's an assumption that channel format should not be adjuested if input is in torch

[vasqu]

Just some thoughts since I'm also exploring the vid processor myself in ernie vl :D

[zucchini-nlp]

run-slow: vjepa

[vasqu]

Thx, makes sense

[zucchini-nlp]

slow CI is not running anymore 🥲 Locally it's green, so I believe it's really fixed

[vasqu]

Did my comment break up things? 👀

[zucchini-nlp]

Lol, no, looks like slow CI isn't being triggered since yesterday night. I've had similar issues in other PRs