Predictor attr not properly cleared when saving

[ascillitoe]

When saving explainers, the predictor attribute is set to None to avoid attempting to save it, and it is then given to load_explainer at load time:

alibi/alibi/saving.py

Lines 110 to 112 in c2da012

	def _simple_save(explainer: 'Explainer', path: Union[str, os.PathLike]) -> None:
	predictor = explainer.predictor # type: ignore[attr-defined] # TODO: declare this in the Explainer interface
	explainer.predictor = None # type: ignore[attr-defined]

However, some explainers, such as AnchorTabular, set other protected attributes just as _predictor when setting the predictor property:

alibi/alibi/explainers/anchors/anchor_tabular.py

Lines 979 to 999 in c2da012

	@predictor.setter
	def predictor(self, predictor: Optional[Callable]) -> None: # Optional here because in saving.py we set it to None
	# deal with the case from saving.py first
	# TODO: how do we prevent users from passing predictor=None? Probably beartype.
	if predictor is None:
	self._predictor = None
	if self.ohe:
	self._ohe_predictor = None
	else:
	# if input is one-hot encoded
	if self.ohe:
	# this predictor expects ordinal/labels encoded categorical variables
	ord_predictor = lambda x: predictor(ord_to_ohe(x, self.cat_vars_ord)[0]) # noqa: E731
	self._predictor = self._transform_predictor(ord_predictor)
	# this predictor expects one-hot encoded categorical variable
	self._ohe_predictor = self._transform_ohe_predictor(predictor)
	else:
	# set the predictor
	self._predictor = self._transform_predictor(predictor)

These protected attributes are not cleared when saving, and are included in the explainer.dill file. I'm don't have a good example of where this is actually a problem at load time, but we should investigate, and if needed implement a more robust approach to clearing self.predictor. This might be as simple as something like:

@predictor.deleter
def predictor(self):
	del self._predictor
        if self.ohe:
             del self._ohe_predictor

and then del explainer.predictor in _simple_save.

[jklaise]

Related issue: #830

We probably shouldn't straight-up delete or set attributes to None as when serialization fails then the explainer is left in a broken state. Instead, the attributes should be cleared whilst maintaining a local variable with a reference to the object and if for some reason serialization fails we can reset the references to restore the explainer to its state pre-calling .save.

[ascillitoe]

Good point, didn't think about that!