S3D feature request

[sophiazhi]

🚀 The feature

S3D model for video classification (paper, paperswithcode, implementation in torchmultimodal).

Motivation, pitch

I've onboarded a model to TorchMultimodal that depends on S3D. Right now the S3D implementation is in torchmultimodal, but as a purely vision model (and a fairly popular model) it would fit nicely in torchvision. There is some discussion on this PR in torchmultimodal about the changes needed to onboard S3D to torchvision.

Alternatives

Right now, S3D is implemented in torchmultimodal.

Additional context

I started working on the suggested refactoring to move S3D from multimodal to vision. However, the process of onboarding myself to the torchvision library and implementing the shared base with the inception model will likely take more time than I have, and my torchvision POC is on PTO. Are there any steps I can take within the next week (my internship is ending), or can the model be onboarded by the torchvision team?

[datumbox]

@sophiazhi Happy to sync with you on how we can unblock this. Since you are an intern with Meta, ping me internally to speed this up.