BUG: Default to stat axis for SparseDataFrame when axis=None

[gfyoung]

Title is self-explanatory. Closes #13048.

[gfyoung]

Surprisingly, there aren't a lot of places where axis is used when calling a cumulative function except for SparseDataFrame because that is the only (non-deprecated) multi-dimensional sparse object.

[jreback]

say axis=0, users have no idea what self._stat_axis_number is

[gfyoung]

Done.

[jreback]

add kurt, skew, sem, max, min, idxmax, idxmin

[gfyoung]

np.kurt, np.skew, np.sem, np.idxmin, and np.idxmax do not exist. Will add for np.max and np.min though.

[jreback]

right, so do for np.argsort/argmin as well then.

[gfyoung]

np.argmin or np.argmax breaks because __array_wrap__ for SparseDataFrame expects a 2-D input, which is not what you get when you call either of those functions. I think this is a bug (or lacking feature) on the pandas side because argmin and argmax should return scalars, meaning __array_wrap__ should have to accept them. @jreback thoughts?

[jreback]

arg mom and max are aliases for idxmin/max which are correct

[gfyoung]

Really? I think that's true only for Series. Doesn't exist for DataFrame. Unless you want to add them for compat purposes.

[jreback]

oh u r testing frames

need same tests for series

[gfyoung]

Ah, okay. But what about the aliasing for DataFrame?

[jreback]

what is aliasing?

[gfyoung]

For Series, you alias idxmin to argmin and likewise idxmax to argmax for ndarray compat. However, that aliasing doesn't exist for DataFrame.

[gfyoung]

@jreback : would be great to get some feedback regarding the argmin/argmax failure (due to the SparseDataFrame not accepting scalar input for __array_wrap__) as well the potential aliasing of idxmin to argmin (and idxmax to argmax) for SparseDataFrame

[jreback]

@gfyoung can you a copy-pastable example here of issue with argmax/min we generally don't support them AFAICT on non Series.

[gfyoung]

>>> from pandas import SparseDataFrame
>>> import numpy as np
>>> sdf = SparseDataFrame([[1, 2, 3], [4, 5, 6]])
>>> np.argmin(sdf)
...
ValueError: Must pass 2-d input

I am not advocating that it be added to the list of supported methods, but I see no reason to test it per se in the list of functions for test_numpy_func_call (for SparseDataFrame that is)...that's all. Or rather, we should test it that raises like this? Not sure...

[jreback]

doesn't work on dense either. I suppose you could make a separate issue about this. Its not high on priority list though.

In [5]: np.argmin(sdf.to_dense())
ValueError: Must pass 2-d input

[jreback]

so let's take out argmin/max

[gfyoung]

Done.

[gfyoung]

@jreback : I don't consider to be an issue, as the documentation makes it clear that you should call the DataFrame.idxmin or DataFrame.idxmax instead of the np.argmin or np.argmax. I just wanted to double check that I could take it out of the list.

[jreback]

@gfyoung yep, take em out. ping when green.

[gfyoung]

@jreback : Travis is all green. Ready to merge if there is nothing else.

[jreback]

thanks!