Fix stats.scoreatpercentile

[rgommers]

Fix ndarray input for per parameter and empty input. Closes gh-1897 and gh-2861.

[coveralls]

Coverage remained the same when pulling 07bd6eb on rgommers:fix-scoreatpercentile into fd99d3f on scipy:master.

[josef-pkt]

I'm not sure if the _equal is always guaranteed.
Floating point calculations could still introduce a noise, but maybe not in this case (1. * x + 0. * y == x) ? IIUC

[rgommers]

Oops, copying test case without thinking. Will change to assert_allclose.

[josef-pkt]

looks good to me

[coveralls]

Coverage remained the same when pulling 7e72d97 on rgommers:fix-scoreatpercentile into fd99d3f on scipy:master.

[juliantaylor]

scoreatpercentile should be implementable in terms of numpy.percentile in numpy 1.9. That supports all the usual numpy features including extended axis.

[rgommers]

@juliantaylor that should be done as soon as numpy 1.9 is the minimum supported version I think. I don't see a reason to now do

if np.__version__ < '1.9':
    ....
else:
    # use np.percentile
    ....

[juliantaylor]

its significantly faster as it uses partition instead of sort. but it can still wait a while 1.9 isn't even out yet :)

[josef-pkt]

My general opinion is that we should point users to numpy.percentile as soon as 1.9 is out and tell them it's much faster, and deprecate this function as soon as possible.

(I haven't checked whether there could be anything specific "stats" that would make it worth keeping. maybe nan handling or masked arrays?)

[rgommers]

Point out that users should use np.percentile as soon as 1.9 is out: definitely. Deprecating this function: only once we require >=1.9. That'll take a few years probably.

[ev-br]

What's wrong with wrapping it with a deprecationwarning as soon as 1.9 is out? (given that DeprecationWarning is off by default with py2.7 anyway)

[rgommers]

We should only generate DeprecationWarnings for things that are fixable by users once they see them. And many of those users can't upgrade numpy easily (or at all, if their numpy is installed by their admin, or they ship an older numpy version with some product, or ...).

In this case there's also no hurry, given that scoreatpercentile is a perfectly fine function.

[josef-pkt]

I don't think we need to add a DeprecationWarning already, but we can mention it in the docstring so users/developers can start to switch, and benefit from the faster numpy version if they have it, and don't have to change all their code when we actually deprecate.

"This function is dooomed."

[rgommers]

Agreed, will do.

[rgommers]

docstring updated

[rgommers]

Can we merge this one now?

[ev-br]

Thanks, merged.