ENH: __array_function__ for multiarray functions
#12175
Conversation
|
Notes/questions:
|
There was a problem hiding this comment.
Only a comment about use of None in where - not sure how best to solve that... Revenge of the object dtype!
More generally, except perhaps for where, I think the extra overhead for these functions would generally be acceptable, as the work they do is fairly complex (for empty_like, in filling the newly created array).
In any case, it seems worth extending this especially to concatenate, even if right now it makes things a little slower.
The only disadvantage that I can see is related to the arguments: if we now allow keyword arguments, we're stuck with them, and in C that is so slow that we are effectively stuck with the python wrapper. But, again, in practice I think the benefits well outweigh these disadvantages.
|
I also tend to think the tradeoffs are worth it here. Even if |
|
I finished implementing multiarray overloads with NumPy wrappers. The most awkward part was for a few datetime functions, because their docstrings lie about their function signatures -- they claim to accept |
|
The Azure build failure looks spurious -- is there an easy way for me to reset that? |
|
@shoyer I restarted the build manually. I added your google email address to Azure devops admin group--not sure if you have a different Microsoft account though. |
|
Thanks Tyler! If you could add my personal Gmail address (my GitHub username at Gmail.com) that would be great. |
|
|
|
@tylerjereddy You can rerun the failing test by itself; click on the failing test so that it shows, pull down the three dot menu up top, and hit rebuild. |
|
@charris I thought I did that, but maybe I failed to properly select so the whole matrix was re-triggered by accident |
|
Since my comment about |
__array_function__ for multiarray functions
The original motivation for the style of these wrapper functions, introduced in numpygh-12175, was to preserve introspection. But it turns out NumPy's functions defined in C don't support introspection anyways, so the extra wrapper functions are entirely pointless. This version reverts the additional wrapper functions, which put default arguments in two places and introduced slow-down due to the overhead of another function call. I've retained docstrings in multiarray.py, since it's definitely more readable to keep docstrings and dispatchers together rather than leaving docstrings in _add_newdocs.py. One bonus of this approach is that dispatcher functions have the same name as their implementations, so `np.concatenate(unknown=True)` gives an error message mentioning "concatenate" rather than "_concatenate_dispatcher": `TypeError: concatenate() got an unexpected keyword argument 'unknown'`
Testing out the Python wrapper approach.
xref #12028