issue_comments

Since this is basically a one line call to apply_ufunc, I propose that we add it as an example of using apply_ufunc in the docs. Maybe even emphasize by adding a section on "wrapping numpy functions"?

@erbian can you change this PR to improve the documention this way?

OK, thanks. Being able to call np.pinv(da, dim='x') would be awesome, but ack the downsides you outlined in the PEP

Yes, we'll want to add our own wrapper functions. (Or possibly a new apply() method could do the dim -> axis renaming)

On Fri, Feb 15, 2019 at 4:01 PM Maximilian Roos notifications@github.com wrote:

Yes, np.pinv will raise if passed a dim kwargs. This was an intentional choice: http://www.numpy.org/neps/nep-0018-array-function-protocol.html#support-for-implementation-specific-arguments

Thanks.

With the current state, am I correct in thinking we still need wrappers for any numpy.linalg function that requires a dim kwarg? (I guess we could use the axis kwarg instead, but that might be confusing things)

— You are receiving this because you are on a team that was mentioned. Reply to this email directly, view it on GitHub https://github.com/pydata/xarray/pull/2766#issuecomment-464254541, or mute the thread https://github.com/notifications/unsubscribe-auth/ABKS1gNtPViHaVR_wyopKAktKCHgztfUks5vN0phgaJpZM4a4kfe .

Yes, np.pinv will raise if passed a dim kwargs. This was an intentional choice: http://www.numpy.org/neps/nep-0018-array-function-protocol.html#support-for-implementation-specific-arguments

Thanks.

With the current state, am I correct in thinking we still need wrappers for any numpy.linalg function that requires a dim kwarg? (I guess we could use the axis kwarg instead, but that might be confusing things)

Or would np.pinv raise if passed a dim kwarg? (if so, we'd still need to declare n wrappers, IIUC?)

Yes, np.pinv will raise if passed a dim kwargs. This was an intentional choice: http://www.numpy.org/neps/nep-0018-array-function-protocol.html#support-for-implementation-specific-arguments

No, that would give an error message about passing in a bad argument name. We'll need our own wrappers to switch axis to dim.

I understand we'd need to define __array_function__ on xarray objects managed the conversion.

Conditional on that, could someone call np.pinv(da, dim=['x', 'y']? And then DataArray.__array_function__ would receive the dim kwarg.

Or would np.pinv raise if passed a dim kwarg? (if so, we'd still need to declare n wrappers, IIUC?)

Since this is basically a one line call to apply_ufunc, I propose that we add it as an example of using apply_ufunc in the docs. Maybe even emphasize by adding a section on "wrapping numpy functions"?

Reading the NEP, I think you could literally pass dim=['x','y'] to the numpy function, and it'll pass that onto xarray. Lmk if I'm mistaken.

No, that would give an error message about passing in a bad argument name. We'll need our own wrappers to switch axis to dim.

Note that with NEP-18 we'll be able to directly everything in numpy.linalg on xarray objects, without requiring the use of a separate wrapper.

It would be awesome for those numpy functions to just work! Could we start on this now, making it conditional on numpy 1.16 installed?

Reading the NEP, I think you could literally pass dim=['x','y'] to the numpy function, and it'll pass that onto xarray. Lmk if I'm mistaken.

Speaking for @erbian (my colleague) , no rush to merge this - we thought it would be helpful but have it internally. Even better if we can get a general solution.

It's not entirely clear to me where we should draw the line between included wrappers in xarray and leaving this to external packages, e.g., see xr-scipy.readthedocs.io

If we do include this in xarray, I would prefer to have both (1) a clear policy around what we include (e.g., only direct wrappers for numpy/scipy functions) and (2) a separate namespace for wrapped functions, e.g., xarray.linalg or something more generic like xarray.wrapped.

Note that with NEP-18 we'll be able to directly everything in numpy.linalg on xarray objects, without requiring the use of a separate wrapper.

cc @pydata/xarray for discussion

Unfortunately this requires numpy 1.14. That's a year old - too soon to bump the minimum version?