Closed dkozlov closed 3 weeks ago
From my perspective, most probably not. DALI has they own augmentation pipeline which is not compatible with albumentations API. Therefore support of DALI would require to introduce breaking changes in albumentations with doubtful outcome.
I want to try it and figure it out if we can simplify DALI usage because it's really cool but really weird
any progress @albu
sorry, no progress here
We have some interesting code to port into data augs libs like albumentations
and
https://github.com/pytorch/ignite/pull/493#issuecomment-581778218
Maybe we can try to build something like that here:
ComposeOps
TransformPipeline
Certainly it wont support all your augs, only DALI ones, but this could help user to avoid copying common stuff.
@albu @ternaus what do you think ?
I can try to provide a draft PR if you are interested.
I have an additional example of the usage here - it really brings DALI usage inline with Albumentations/PyTorch Compose usage, so would be very cool to have integrated
Usability wise this code is a super handy way of dropping GPU acceleration as a nearly 'drop in' replacement to existing PyTorch/Albumentations workloads - we just need to find a home for it.
Take a look! Here is another Colab:
https://colab.research.google.com/drive/1XDRzDeIeteTPyzYbZp2XWjKtUWMLXqcg
No plans to support DALI
Hi, any plans to add https://github.com/NVIDIA/DALI support?