This indeed looks deceptively simple

Agreed.

The axis=None, keepdims=True case doesn't seem tested though

Ah! Yes. Missed it. Thanks for pointing out. I am on it.

Please let me know if there are any cases where the current approach fails or is inefficient.

Ah, other issue: this introduces a mismatch between np.argmax and np.ndarray.argmax. That is undesirable. I think that means the changes can't be Python-only after all.

Yeah. I was thinking so, while making the change. Noticed that, np.nadarray.max has keepdims argument. So, for overall consistency np.ndarray.argmax should also have keepdims. I will investigate where the changes might be needed to make things work.

I have made some updates to the interface of argmax and argmin. Let me know if any changes are required.

I have made some updates to the interface of argmax and argmin. Let me know if any changes are required.

Did you mean to touch MaskedArray methods? For ndarray methods I think you need to change array_argmax and PyArray_ArgMax.

Did you mean to touch MaskedArray methods?

I thought that, MaskedArray.max has keepdims, so added it to MaskedArray.argmax. Do we have to avoid making changes in it? I just looked for places where max has keepdims but argmax doesn't and for consistency made the changes to argmax.

Do we need to use, np._NoValue for keepdims in MaskedArray methods?

Ah no, if it was intentional then it's all good! Updating numpy.ma at the same time as the functionality in the main namespace is great.

Do we need to use, np._NoValue for keepdims in MaskedArray methods?

Yes, it looks like that's necessary. It looks odd; the reason for it is to support (partial) subclasses. Say .method1(..., keepdims=False) would internally call self.method2(keepdims=keepdims) and the subclass overrides method2 but not method1, then an update to the newer numpy version which adds the keepdims keyword would break the subclass.

@BvB93 I have added a release notes file to the PR. Are there any improvements to be made to that?

Why were all the annotations removed?

keepdims is now a keyword only argument (refer #19211 (comment)) and hence I removed annotations for it. I checked some other functions (like, any) which are using keyword only arguments and saw that they were not having annotations for keyword arguments. Let me know the format to add annotations for these, and I will add them. Thanks.

Edit - Apologies for resolving the conversations here. Actually, I do it just keep track of stuff which is yet to be addressed. Let me know if I should un-resolve them.

keepdims is now a keyword only argument (refer #19211 (comment)) and hence I removed annotations for it. I checked some other functions (like, any) which are using keyword only arguments and saw that they were not having annotations for keyword arguments. Let me know the format to add annotations for these, and I will add them. Thanks.

I suspect you might be confusing keyword-only (func(a, *, b=...)) with variadic keyword arguments (func(a, **kwargs)). The latter is often annotated as Any, simply due to a lack of suitable and/or convenient alternatives. On the contrary, the former is (arguably) even easier to annotate than normal keyword-or-positional arguments.

In any case, compared to the annotations prior to 9260d40 only a single change will have to be made:
a * must be added preceding the keepdims keyword in order to mark the argument as keyword-only.

@overload
def argmin(
    a: ArrayLike,
    axis: None = ...,
    out: Optional[ndarray] = ...,
    out: Optional[ndarray] = ...,
+    *,
+    keepdims: Literal[False] = ...,
) -> intp: ...
@overload
def argmin(
    a: ArrayLike,
    axis: Optional[int] = ...,
    out: Optional[ndarray] = ...,
+    *,
+    keepdims: bool = ...,
) -> Any: ...

I have restored the annotations. Let me know if anything else is required. Thanks.

One awesome thing (happy with followup or skipping), would be to unify the two functions completely. They seem 100% copy-paste of each other with 2 variations (the method, and the "argmax" vs. "argmin" in error messages). And with PyErr_Format these days, both of those should be straight forward to unify.

I am thinking to do bit of refactoring (in C and test files) after this PR gets in. Let me know if that works.

It might be good to ping the mailing list briefly, also about the C-API addition (which I honestly don't care about personally). Those should maybe also have a very brief (just a bullet point is enough) additional release note under the C-API section.

I have added a release notes entry for C-API too. Let me know if it is okay.

Thanks for bearing with me trying to find a minimal solution!

In my opinion the solution is really elegant, simple and abstract. Thanks for guiding me to finally arrive at it. :-)

I have removed the C-API label as well. Thanks.

OK, lets put this in. Thanks Gagandeep! @czgdp1807, I think you did not apply all style fixes in both places always. I assume you will followup with a PR to merge the two functions into a single one that will clean up the styles where it is nicer?

I assume you will followup with a PR to merge the two functions into a single one that will clean up the styles where it is nicer?

Yes sure. On the way. I will try to factor out the common parts in argmin, argmax and any other functions in the main code and tests. Thanks for merging.