Batched (random) augmentations about vision HOT 3 OPEN

Thanks for the feature request @Guitaricet . This has been a reccurent feature request for a while, but there is no current obvious path to support that just yet, as most kernel implementations are just not easily parallelizable over RNG. vmap is a potential avenue, but last time I tried I ended up quickly blocked.

You'll see torchaug mentioned in #2929 (comment) which supports what you need. Some kernels are natively implemented with support for per-sample RNG, but for most transforms the batches are instead split into sub-batches, transformed and then re-concatenated, so it's more of a per-subbatch-RNG AFAIU. If you try it and it works for your use-case I'd love to hear it, and perhaps consider something simliar in torchvision (CC @juliendenize who's the author of torchaug).

from vision.

Hi thanks for the ping @NicolasHug , you're right about how Torchaug works it offers the flexibility to do per-subbatch-RNG when a per-sample RNG is not possible in Torch.

I'd like to also hearing from you @Guitaricet if you try it and if it works, I try to maintain it as much as I can and follow Torchvision advances but I'd need more feedback from users to make it better. I'd love it if eventually Torchvision could natively support such features 😄.

from vision.

Thank you, I will take a look!

from vision.