[SciPy-dev] scipy.distance

Charles R Harris charlesr.harris@gmail....
Mon Nov 3 01:02:29 CST 2008


On Sun, Nov 2, 2008 at 10:03 PM, Anne Archibald
<aarchiba@physics.mcgill.ca>wrote:

> 2008/11/2 Jarrod Millman <millman@berkeley.edu>:
> > On Sun, Nov 2, 2008 at 4:47 PM, Damian Eads <damian.eads.lists@gmail.com>
> wrote:
> >> I had a chance to peruse the code in the spatial branch, and it looks
> >> like there are a considerable number of tests for the kd-tree code. It
> >> seems mature enough for it to be incorporated into the 0.7 release as
> >> a technology preview. I think we should move that code over into the
> >> trunk too, since this will give us more feedback from early adopters
> >> and testers so we can better refine. It should also be noted, we are
> >> changing the module index in the documentation to indicate those
> >> modules/packages that are part of the technology preview.
> >
> > +1
> > I think that it would be better to get the kd-tree code out in this
> > release, but clearly mark it as a technology preview.  This will give
> > us the opportunity to get more feedback and testing without any major
> > restrictions on improving the API.  I know that there was some
> > concerns that including new code could slow down the release of 0.7,
> > but I don't think that has to be the case.  Since it is a technology
> > preview and isn't changing current behavior; as long as it compiles
> > and doesn't create a lot of noise for testing, I don't think I will be
> > necessary to delay the release for it.
> >
> > Thoughts?
>
> I think for computing nearest neighbors of points, the code is
> basically done. For other applications, well, it's not yet clear what
> other applications people want. One aspect I do expect will change is
> the API: there are many applications where a kd-tree is a useful data
> structure (beyond simple point nearest neighbor), but for most of them
> one needs to write a custom tree traversal and possibly annotate the
> tree. I'm not sure how to set the kd-tree code up so that it can
> easily (and efficiently) be used for that, but I think the only way to
> find out is for users to actually go ahead and try to do it.


I usually want a complete list of points in some neighborhood. I looked
through your cython code and I think the loops can be improved a bit to make
better use of low level C code.

I'm working on some cover tree code with an eye to future inclusion unless
someone beats me to it.

Chuck
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://projects.scipy.org/pipermail/scipy-dev/attachments/20081103/b812f9b6/attachment-0001.html 


More information about the Scipy-dev mailing list