TYP: Correct type annotation for to_dict. #55130
Conversation
The `into` argument of DataFrame.to_dict and Series.to_dict can be either a class or instance of a class of dict; this is covariant - subclasses of dict can also be used. The argument was annotated as `type[dict]` though, so type checkers marked passing initialized objects (required for collections.defaultdict) as an incorrect argument type. Fix by annotating `into` to take either a subclass of dict or an initialized instance of a subclass of dict.
jsspencer
force-pushed
the
to_dict_into_type_annotate
branch
from
afbc799
to
449eb46
Compare
jsspencer
changed the title
Unfortunately a generic type annotation with a default triggers an existing mypy limitation (python/mypy#3737). The current workaround is to use overloads and then not annotate the implementation containing the default parameter; this still enables mypy to deduce correct return types. Two overloads are added for Series.to_dict, even though they could be combined using a Union type, as at least two overloads are required for a single method.
pandas/core/frame.py
Outdated
| index: bool = True, | ||
| ) -> dict | list[dict]: | ||
| ): |
There was a problem hiding this comment.
Can you add the annotations for the non-overload definition? If it triggers a mypy error within the function, it is okay to add an ignore comment.
edit: having the ignore comment will function as a TODO enforced by mypy (when a future mypy allows TypeVars+default )
There was a problem hiding this comment.
@jsspencer I pushed to your branch so that pd.DataFrame([]).to_dict("dict") works correctly
There was a problem hiding this comment.
Should now work for all cases:
import pandas as pd
from typing import MutableMapping
class MyDict(dict): ...
def test(into: MutableMapping):
# MutableMapping
reveal_type(pd.DataFrame([]).to_dict("dict", into=into))
reveal_type(pd.Series([]).to_dict(into=into))
# dict
reveal_type(pd.DataFrame([]).to_dict("dict"))
reveal_type(pd.Series([]).to_dict())
# MyDict
reveal_type(pd.DataFrame([]).to_dict("dict", into=MyDict))
reveal_type(pd.Series([]).to_dict(into=MyDict))
There was a problem hiding this comment.
I created pandas-dev/pandas-stubs#784 so we make a similar change in the stubs
| # error: Incompatible default for argument "into" (default has type "type[ | ||
| # dict[Any, Any]]", argument has type "type[MutableMappingT] | MutableMappingT") | ||
| @deprecate_nonkeyword_arguments( | ||
| version="3.0", allowed_args=["self"], name="to_dict" |
There was a problem hiding this comment.
Is this okay @mroeschke? It would make it consistent with DataFrame.to_dict (and allow the dict default case for typing) but it might be a bit "odd" to require keyword-only for a function that takes exactly one argument
(need to also fix the failing doc test for this change)
There was a problem hiding this comment.
Yup the consistency argument makes sense to go forward with this deprecation
|
Thanks @jsspencer |
The
intoargument of DataFrame.to_dict and Series.to_dict can be either a class or instance of a class of dict; this is covariant - subclasses of dict can also be used. The argument was annotated astype[dict]though, so type checkers marked passing initialized objects (required for collections.defaultdict) as an incorrect argument type.Fix by annotating
intoto take either a subclass of dict or an initialized instance of a subclass of dict.