Repository navigation
BUG: np.ndarray.__array_function__ cannot be handle NEP 35 creation functions #27500
Description
Describe the issue:
NEP 35 introduced the ability to pass in an instance via like= for array creation functions such as zeros and arange, which are passed through __array_function__. But ndarray subclasses cannot use super().__array_function__: one gets
a = np.array([])
a.__array_function__(np.asarray, (np.ndarray,), [1., 2., 3.], {})
AttributeError: 'builtin_function_or_method' object has no attribute '_implementation'
Reproduce the code example:
import numpy as np
a = np.arange(3)
# Passing through reshape works fine
a.__array_function__(np.reshape, (np.ndarray,), (a, (3,1)), {})
# array([[0],
# [1],
# [2]])
# But not asarray
a.__array_function__(np.asarray, (np.ndarray,), [1., 2., 3.], {})
# AttributeError: 'builtin_function_or_method' object has no attribute '_implementation'Error message:
No response
Python and NumPy Versions:
2.1.1
3.11.2 (main, Aug 26 2024, 07:20:54) [GCC 12.2.0]
Runtime Environment:
No response
Context for the issue:
Found while trying to add support for NEP 35 functions for astropy's Quantity.
Related: #27451, about these creation functions not being present in the output of np.testing.overrides.get_overridable_numpy_array_functions
Reacted by Clément Robert
Activity
seberg commentedon Oct 4, 2024
Hmmm, yeah. Not that I am sure
super()is very useful... Should we have an_implementationthat just ignores thelike=!? Or uses it for.view()on the result?(Just thinking that might be easier for symmetry then some dropping of
like=itself.)mhvk commentedon Oct 4, 2024
It is definitely not super useful, but I liked how for a subclass one does not have to know about
_implementationgenerally, but can just callsuper(). Currently, the implementation ofndarray.__array_function__explicitly looks for_implementation:numpy/numpy/_core/src/multiarray/arrayfunction_override.c
Lines 159 to 161 in 5c8f540
Ideally, it would just call functions that do not have this directly. Note that
__array_function__is not passed thelikeargument, though it does have itstype, so I guess in principle one could do a view. I don't know that that is really necessary, just not failing would be good enough!Note that I'm happy to make the change, just not quite sure what it should be: just call the function directly if there is no
_implementation? That should be OK in principle...mhvk commentedon Oct 4, 2024
p.s. Looking more at the code, I see that if a subclass does not define
__array_function__one is just passed on to the regular implementation, so it seems reasonable to do the same for asuper()call.mhvk commentedon Oct 4, 2024
@seberg - see #27504 for a possible fix (built on #27503, since I found I could cause segfaults with bad calls to
ndarray.__array_function__...)seberg commentedon Oct 4, 2024
Thanks, I have to think about it abit and look at all of the open PRs on this in the next day or two!