AutoAugment image augmentation

[innat]

Paper: https://arxiv.org/abs/1805.09501 Cited by 832 Ref code. https://github.com/tensorflow/models/blob/ded32f0500604928e52e27fd3f678e694e5133b7/official/vision/image_classification/augment.py#L731

[LukeWood]

I've been weighing adding this layer.

Thoughts on what it adds over RandAugment?

It looks as though it's much more customizable, which could be good to have.

@innat

[bhack]

I think that we could also evaluate differentiable augmentations like:

https://arxiv.org/abs/2104.04282 https://arxiv.org/abs/2109.15273 https://openreview.net/forum?id=St-53J9ZARf

[innat]

(AFAIK), the advantages of auto-augment over rand-augment also bring computational costs. I think both of them are used and maintained in research, so it's ok to have them both.

• AutoAugment - Code.
• RandAugment - Code
• Review.

[bhack]

I think that we could also evaluate differentiable augmentations like: https://arxiv.org/abs/2104.04282 https://arxiv.org/abs/2109.15273 https://openreview.net/forum?id=St-53J9ZARf

/cc @haifeng-jin

[LukeWood]

this should probably use RandomAugmentationPipeline.

[bhack]

I think that we could also evaluate differentiable augmentations like: ..... https://openreview.net/forum?id=St-53J9ZARf

https://github.com/MSU-MLSys-Lab/DeepAA

[LukeWood]

I've received numerous inquiries about AutoAugment, so I definitely think we want this component.

[bhack]

Has https://github.com/keras-team/keras-cv/issues/900 was cloud I will unify the references on this ticket.

[LukeWood]

Lets deprioritize in favor of RandAugment; which serves a similar purpose but is much easier to tune.