EHN: add nanmin/nanmax in scipy.sparse

[glemaitre]

NumPy allows for finding np.min and np.max by ignoring NaN values.
Currently, there is no implementation in the sparse module while np.fmin and np.fmax could be used similarly as np.minimum and np.maximum ufunc.

Our use case is in scikit-learn: we have some pre-processing methods for which we would like to discard NaN to compute those extrema.

[rgommers]

Makes sense to have nanmin/nanmax. I'm just not sure that these should be methods rather than functions (they're functions in numpy).

[glemaitre]

Makes sense to have nanmin/nanmax. I'm just not sure that these should be methods rather than functions (they're functions in numpy).

It was one of my concern as well. Would you make them accessible by scipy.sparse.nanmin and scipy.sparse.nanmax.

[rgommers]

For symmetry with numpy, that would be my preference I think. Since this is a new feature, it would be useful if you could mention this on the scipy-dev mailing list to see if others have opinions.

[glemaitre]

OK

…

On 9 June 2018 at 20:15, Ralf Gommers ***@***.***> wrote: For symmetry with numpy, that would be my preference I think. Since this is a new feature, it would be useful if you could mention this on the scipy-dev mailing list to see if others have opinions. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#8902 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AHG9P4v3LYNKs5Psl66raIAwTHdPGH8Cks5t7BCkgaJpZM4UZxgP> .

-- Guillaume Lemaitre INRIA Saclay - Parietal team Center for Data Science Paris-Saclay https://glemaitre.github.io/

[charris]

The current nep for numpy __array_function__ might be relevant here. @shoyer Comment?

[shoyer]

If __array_function__ becomes a thing, then it would certainly make sense for scipy to use it here. But there's no reason to hold up adding separate functions/methods for it.

[rth]

If array_function becomes a thing, then it would certainly make sense for scipy to use it here. But there's no reason to hold up adding separate functions/methods for it.

Even if one makes np.nanmin work on sparse arrays, I imagine having a method or a function in scipy.sparse would still be useful for discoverability?

[rth]

There is a merge conflict that would need resolving.

[perimosocordiae]

@glemaitre this PR is looking close to mergeable, just a few minor issues to address. If you don't think you'll have time, let us know and someone else can take over (your existing commits will remain).

[glemaitre]

@perimosocordiae Be aware that right now, nanmin nanmax are just methods. I am unsure what's best between methods and functions then.

[perimosocordiae]

I just opened gh-18542 which preserves these commits and resolves the merge conflict from scipy/sparse/data.py, so I'll close this PR now.