[SciPy-User] Graph connect components and sparse matrices

Robert Cimrman cimrman3@ntc.zcu...
Wed Nov 18 08:44:16 CST 2009


Gael Varoquaux wrote:
> On Wed, Nov 18, 2009 at 03:24:37PM +0100, Robert Cimrman wrote:
>>> Well, getting the largest eigenvector of the transition matrix is in
>>> o(n), using arpack, AFAIK. So the cost is similar, and on one side we
>>> have optimized C code, and on the other side I only had Python code (or C
>>> code that I don't want to maintain). In addition, as I am doing diffusion
>>> maps, I needed to call arpack anyhow.
> 
>> I see. BTW. putting a code into scipy somewhat alleviates the
>> maintenance burden ;)
> 
> I'd love to, but the code I am talking about is not something you want to
> see. I inherited it from the lab, and its been a horrible burden. Not
> that there are not good part in it (there are a lot of excellent
> alogrithms), but the problem is that it uses home grown vector
> abstractions, and graph structures, which makes it really hard to split
> out the good part. In the long run, I hope I will be able to trim out the
> bad parts and the vector library, and replace this by scipy components,
> and the work that David Cournapeau has been doing to expose numpy
> internals to C libraries. Once this is doing, we can think of moving
> things out to other libraries: scipy, networkx, or the machine learning
> scikit (we have an engineer hired to work on that, beginning in January).

Now this is an interesting shift in attitude that I experienced myself - 
instead of putting all the cool stuff into own code, distribute it over 
well-known and maintained packages ;)

cheers,
r.


More information about the SciPy-User mailing list