Randomizable constructor

[rijobro]

Adds randomizable constructor.

Description

This allows the parent class to store prob and _do_transform.

I'll need this for the inverse transformation PR.
#1515

Status

Ready

[rijobro]

@wyli any chance you could have a look at this? Hoping to get it merged so I can move onto the next set of PRs.

[wyli]

@wyli any chance you could have a look at this? Hoping to get it merged so I can move onto the next set of PRs.

thanks I'll need more time to check the details as there's no unit test to cover the changes...

Randomizable is a general one, for example it's also used for shuffling the dataset https://github.com/Project-MONAI/MONAI/blob/master/monai/apps/datasets.py#L136 perhaps we could subclass it for the randomised transforms?

[rijobro]

thanks I'll need more time to check the details as there's no unit test to cover the changes...

Sure, my metric was just that none of the previous unit tests were broken. Slightly worrying that the code format wasn't picking up any warnings with this one... https://github.com/Project-MONAI/MONAI/blob/master/monai/apps/datasets.py#L136

EDIT: oh, it's because it doesn't call the base class.

[rijobro]

@wyli updated to add a RandomizableTransform class.

[wyli]

/integration-test

[wyli]

thanks, it looks good to me

[wyli]

shall we support None here and in the RandomizableTransform? (if it doesn't break anything...)

Suggested change

	def randomize(self, data: Any) -> None:
	def randomize(self, data: Optional[Any] = None) -> None:

[rijobro]

I shouldn't have enabled automatic merging!

I would quite like the default of data for Randomizable and RandomizableTransform to default to None. Currently I have to use a lot of super().randomize(None), but I'd rather have super().randomize().

However, this is incompatible with certain child classes. E.g., RandFlip takes np.ndarray as argument, which is incompatible with None.

The alternative is to switch all child classes from e.g., np.ndarray to Optional[np.ndarray], but that doesn't make sense either, because the np.ndarray in this case isn't optional.

Another alternative is just to ignore the mypy errors for those classes.

What do you think?

[wyli]

thanks! it's just a very minor issue, I think the current one is also good, I forgot the mypy errors...