ENH: Add annotations for np.lib.arrayterator
#18545
Conversation
numpy/lib/arrayterator.py
Outdated
| @@ -124,14 +124,17 @@ def __getitem__(self, index): | |||
| out.stop[i] = min(stop, out.stop[i]) | |||
| return out | |||
|
|
|||
| def __array__(self): | |||
| def __array__(self, dtype=None): | |||
There was a problem hiding this comment.
Looks to me like dtype should be a positional argument:
>>> x = np.arange(5)
>>> x.__array__(np.float32)
array([0., 1., 2., 3., 4.], dtype=float32)
>>> x.__array__(dtype=np.float32)
Traceback (most recent call last):
File "<ipython-input-43-6adb0d04469e>", line 1, in <module>
x.__array__(dtype=np.float32)
TypeError: __array__() takes no keyword arguments
There was a problem hiding this comment.
Based on some trial and error it seems that the dtype parameter must have a default value but can otherwise be either position-only or positional-or-keyword (ndarray being an example of the former).
Secondly, I'm not convinced anymore that this change is even necessary in the first place,
as I was initially under the (incorrect) impression that a dtype parameter would be required here
if one wants to use it with the likes of, e.g., np.array(x, dtype=<insert dtype different from x's dtype>).
I think it might be best to revert this change as well and then make the npt._SupportsArray protocol a bit more flexible, such that it can accomodate __array__ without the dtype parameter.
There was a problem hiding this comment.
Yes, adding the dtype keyword is not needed, reverting sounds good.
Implementations of `__array__` do not require the `dtype` parameter
The likes of `np.array(..., dtype=dtype)` will always pass `np.dtype` instances to `__array__(self, dtype=...)`, so that latter should not have to take arbitrary dtype-like objects but only actual `np.dtype` instances.
|
rebased to avoid the lint-check early CI termination. |
| @@ -14,4 +14,4 @@ | |||
| reveal_type(a[...]) # E: numpy.ndarray[Any, numpy.dtype[numpy.str_]] | |||
| reveal_type(a[:]) # E: numpy.ndarray[Any, numpy.dtype[numpy.str_]] | |||
| reveal_type(a.__array__()) # E: numpy.ndarray[Any, numpy.dtype[numpy.str_]] | |||
| reveal_type(a.__array__(np.float64)) # E: numpy.ndarray[Any, numpy.dtype[Any]] | |||
| reveal_type(a.__array__(np.dtype(np.float64))) # E: numpy.ndarray[Any, numpy.dtype[{float64}]] | |||
There was a problem hiding this comment.
One follow-up question: numpy.dtype[{float64}] is a little verbose. How hard would it be to make numpy.ndarray[Any, numpy.float64] work?
This PR adds annotations for the
Arrayteratorclass innp.lib.arrayterator.Two things of note:
__getattr__method ofArrayteratorwraps around the underlyingndarrays attributes.To simulate this behavior the former has been typed such that it subclasses the latter
(even though, strictly speaking, this isn't true at runtime).
dtypeparameter was missing from theArrayterator.__array__method. It has now been added.