I’m trying to implement a kind of custom sequence class in Python:
from typing import Sequence, TypeVar, List
T = TypeVar('T')
class MySequence(Sequence[T]):
def __init__(self):
self._container: Sequence[T] = []
def __getitem__(self, idx):
return self._container[idx]
def __len__(self):
return len(self._container)
Now I want to check that mypy is aware that elements of MySequence are items of type T:
foo: MySequence[str] = MySequence() reveal_type(foo[0]) # Revealed type is 'Any'
So it fails: mypy knows nothing about items of foo. The same example for ordinary Sequence works:
bar: Sequence[str] = [] reveal_type(bar[0]) # Revealed type is 'builtins.str*'
If I’m trying to add type annotations to __getitem__ implementation, I have another error:
def __getitem__(self, idx) -> T: # Signature of "__getitem__" incompatible with supertype "Sequence"
I also tried
def __getitem__(self, idx) -> Union[T, Sequence[T]]:
as idx can be a slice and in that case my code will return a sequence instead of one element. It fails either with the same message.
As discussed in my previous question, there is an open discussion on issues like that.
However, I still wonder, is it possible to create custom sequence types that allow mypy to extract information about the type of its items, like in my example?
Advertisement
Answer
In this case, the correct thing to do is to correctly override the exact signature for __getitem__, including the overloads.
from typing import Sequence, TypeVar, List, overload, Union
T = TypeVar('T', covariant=True)
class MySequence(Sequence[T]):
def __init__(self):
self._container: Sequence[T] = []
@overload
def __getitem__(self, idx: int) -> T: ...
@overload
def __getitem__(self, s: slice) -> Sequence[T]: ...
def __getitem__(self, item):
if isinstance(item, slice):
raise Exception("Subclass disallows slicing")
return self._container[item]
def __len__(self) -> int:
return len(self._container)
foo: MySequence[str] = MySequence()
reveal_type(foo[0])
(Note that I made the typevar covariant. This isn’t, strictly-speaking, required, but if the container is effectively meant to represent a “read-only” kind of structure, we might as well for maximum flexibility.)
Note: the fact that mypy decides that the return type is Any in the first example is expected behavior. According to PEP 484, any method or signature without type annotations is treated as if the argument and return types are all Any.
This is a mechanism designed so that untyped Python code is treated as fully dynamic by default.
Mypy comes built-in with a variety of command line arguments you can use to try and coerce it to check the contents of untyped functions anyways (I believe it’s --check-untyped-defs?), though it won’t attempt to infer what the return type is.