README.rst

pyDistances

Fast cython distance computations in a variety of metrics

This is the outline of a framework that will allow general distance metrics to be incorporated into scikit-learn BallTree. The idea is that we need a fast way to compute the distance between two points under a given metric. In the basic framework here, this involves creating an object which exposes C-pointers to a function and a parameter structure so that the distance function can be called from either python or directly from cython with no python overhead.

Framework

I am trying to mimic as much as possible the metrics available in scipy.spatial.distance, matching the speed of these computations, while maintaining a c function which can be exposed for fast distance computation within cython loops. Once this is complete, we will be able to re-write BallTree to use any of these distance metrics, and then possibly modify scipy.spatial.cKDtree as well.

Benchmarks

run bench.py for a comparison of runtimes between pyDistance and scipy.spatial for the available metrics. So far, we're doing pretty well: very close in most cases. One exception is boolean values: I need to figure out how to do fast boolean computations within cython.

TODO

Search TODO within distmetrics.pyx to see a list. One big one (which should be straightforward) is to make the distance metrics work with CSR matrices. This will involve writing an alternate version of each core distance function.