invertible Lambdad transform?

[wyli]

not sure if this should become a feature request,
do we need an invertible LambdaD with a customisable .inverse()?

MONAI/monai/transforms/utility/dictionary.py

Lines 703 to 707 in 583e35b

	class Lambdad(MapTransform):
	"""
	Dictionary-based wrapper of :py:class:`monai.transforms.Lambda`.
	For example:

[rijobro]

Could do. Are you thinking (untested):

class InvertibleLambdad(Lambdad, InvertibleTransform):
    def __init__(
        self,
        keys: KeysCollection,
        func: Union[Sequence[Callable], Callable],
        inv_func: Union[Sequence[Callable], Callable],
        overwrite: Union[Sequence[bool], bool] = True,
        allow_missing_keys: bool = False,
    ) -> None:
        super().__init__(keys, func, overwrite, allow_missing_keys)
        self.inv_func = ensure_tuple_rep(inv_func, len(self.keys))

    def __call__(self, data):
        d = dict(data)
        for key, func, overwrite in self.key_iterator(d, self.func, self.overwrite):
            ret = self._lambd(d[key], func=func)
            if overwrite:
                d[key] = ret
            self.push_transform(d, key)
        return d

    def inverse(self, data):
        d = deepcopy(dict(data))
        for key, inv_func, overwrite in self.key_iterator(d, self.inv_func, self.overwrite):
            transform = self.get_most_recent_transform(d, key)
            ret = self._lambd(d[key], func=inv_func)
            if overwrite:
                d[key] = ret
            self.pop_transform(d, key)
        return d

N.B.: Doesn't allow the user to define extra_info or access other information, such as the image's original size. If a user wanted that I guess they'd have to write a whole class themselves. Same for RandLambdad.

[Nic-Ma]

Addressed by PR: #2667.
Thanks.