`MaybeApply` high order abstraction

We should consider an implementation for a MaybeApply layer:

maybe_solarize = MaybeApply(Solarization(), rate=0.5)
# maybe_solarize solarizes half of the images passed to it

https://github.com/chjort/chambers/blob/b796f507b7353fb141b5c8b026b811a023d303e6/chambers/augmentations/image_augmentations.py#L514

Notes:

• we can rely on this in RandomAugmentationPipeline
• this must be element-wise, not batch wise!

[LukeWood]

Got it, please let me know if you decide.

Personally I wouldn’t mind an option to create custom random layers.

It seems to me this is such a common pattern that we should include it.

I’ve seen:

if tf.random.uniform() > 0.5:
    # apply some augmentation

many times in my life, and this is actually incorrect if you are performing batched augmentation. By providing a proper implementation that does this on an element-wise basis, we are providing some value. I’m fairly sure we want this, let me confirm one more time tomorrow, but I am 90% sure this is an abstraction we want.

[LukeWood]

FYI @sebastian-sz I’ve tagged this contribution welcome.

These high order components that take other components are a nice pattern.