[SciPy-user] SVN: sandbox/pyloess

Pierre GM pgmdevlist@gmail....
Mon Mar 12 15:13:31 CDT 2007


All,
A few weeks ago, an implementation of loess was requested on this list. I just 
uploaded a beta version in the 'pyloess' directory of the sandbox. There are 
actually one interface for two modules: lowess, the univariate  locally 
weighted linear estimation, and STL, the seasonal/trend decomposition. The 
modules are just f2py wrappers of the initial fortran sources (converted to 
double precision). I'd be happy to get your feedback, as usual.

I've been also trying to port the loess algorithm (a more generic version of 
lowess that can accept up to 8 variables). However, I've been hitting a wall 
on this one. The initial code is in C, with regular calls to fortran 
subroutines. The base object is a nested structure, and the operations are 
performed on this nested structure. 

Initially, I tried to code the nested structure in python directly, so that I 
would just have to call the C/fortran routines, but that didn't really work 
(the results from some tests didn't match the ones I was getting from the C 
code directly). I then tried to use SWIG, before realizing I was spending 
most of my time writing extensions. Eventually, I decided to stick to Pyrex. 
I wrote as many classes as sub-structure to get access to the underlying 
objects, plus an extra class for the combination of the substructures. The 
result is so far fairly disappointing, as the final object doesn't even get 
initialized properly. After one week of this "going-nowhere-fast", I'm giving 
up temporarily. 
Nevertheless, I uploaded this prototype in a 'sandbox' subfolder of the 
package, hoping that a kind soul would provide me with some pointers about 
the generic approach and some ideas about why I segfault all the time (I'm 
thinking about the pyrex specialists among us...). 

Thanks a lot in advance for your comments and feedback.

Pierre



More information about the SciPy-user mailing list