[Numpy-discussion] Vectorized percentile function in Numpy (PR #2970)

[Sebastian Berg]

On Tue, 2013-04-23 at 23:33 -0400, josef.pktd at gmail.com wrote:
> On Tue, Apr 23, 2013 at 6:16 PM, Sebastian Berg
> <sebastian at sipsolutions.net> wrote:
> > On Tue, 2013-04-23 at 12:13 -0500, Jonathan Helmus wrote:
> >>      Back in December it was pointed out on the scipy-user list[1] that
> >> numpy has a percentile function which has similar functionality to
> >> scipy's stats.scoreatpercentile.  I've been trying to harmonize these
> >> two functions into a single version which has the features of both.
> >>      Scipy PR 374[2] introduced a version which look the parameters from
> >> both the scipy and numpy percentile function and was accepted into Scipy
> >> with the plan that it would be depreciated when a similar function was
> >> introduced into Numpy.  Then I moved to enhancing the Numpy version with
> >> Pull Request 2970 [3].  With some input from Sebastian Berg the
> >> percentile function was rewritten with further vectorization, but
> >> neither of us felt fully comfortable with the final product.  Can
> >> someone look at implementation in the PR and suggest what should be done
> >> from here?
> >>
> >
> > Thanks! For me the main question is the vectorized usage when both
> > haystack (`a`) and needle (`q`) are vectorized. What I mean is for:
> >
> > np.percentile(np.random.randn(n1, n2, N), [25., 50., 75.], axis=-1)
> >
> > I would probably expect an output shape of (n1, n2, 3), but currently
> > you will get the needle dimensions first, because it is roughly the same
> > as
> >
> > [np.percentile(np.random.randn(n1, n2, N), q, axis=-1) for q in [25., 50., 75.]]
> >
> > so for the (probably rare) vectorization of both `a` and `q`, would it
> > be preferable to do some kind of long term behaviour change, or just put
> > the dimensions in `q` first, which should be compatible to the current
> > list?
> 
> I don't have much of a preference either way, but I'm glad this is
> going into numpy.
> We can work with it either way.
> 
> In stats, the most common case will be axis=0, and then the two are
> the same, aren't they?
> 
> What I like about the second version is unrolling (with 2 or 3
> quantiles), which I think will work
> 
> u, l = np.random.randn(2,5)
> or
> res = np.percentile(...)
> func(*res)
> 
> The first case will be nicer when there are lots of percentiles, but I
> guess I won't need it much except for axis=0.
> 
> Actually, I would prefer the second version, because it might be a bit
> more cumbersome to get the individual percentiles out if the axis is
> somewhere in the middle, however I don't think I have a case like
> that.
> 

I never thought about the axis being where to insert the dimensions of
the quantiles. That would be a third option. It feels simpler to me to
just always use the end (or the start) though.

- Sebastian

> The first version would be consistent with reduceat, and that would be
> more numpythonic. I would go for that in numpy.
> 
> my 2.5c
> 
> Josef
> 
> >
> > Regards,
> >
> > Sebastian
> >
> >>   Cheers,
> >>
> >>      - Jonathan Helmus
> >>
> >>
> >> [1] http://thread.gmane.org/gmane.comp.python.scientific.user/33331
> >> [2] https://github.com/scipy/scipy/pull/374
> >> [3] https://github.com/numpy/numpy/pull/2970
> >> _______________________________________________
> >> NumPy-Discussion mailing list
> >> NumPy-Discussion at scipy.org
> >> http://mail.scipy.org/mailman/listinfo/numpy-discussion
> >>
> >
> >
> > _______________________________________________
> > NumPy-Discussion mailing list
> > NumPy-Discussion at scipy.org
> > http://mail.scipy.org/mailman/listinfo/numpy-discussion
> _______________________________________________
> NumPy-Discussion mailing list
> NumPy-Discussion at scipy.org
> http://mail.scipy.org/mailman/listinfo/numpy-discussion
>