LFD Book Forum

LFD Book Forum (http://book.caltech.edu/bookforum/index.php)
-   Homework 8 (http://book.caltech.edu/bookforum/forumdisplay.php?f=137)
-   -   Why is CVXOPT slower than LIBSVM? (http://book.caltech.edu/bookforum/showthread.php?t=1288)

Willem 09-02-2012 10:50 AM

Why is CVXOPT slower than LIBSVM?
 
libsvm runs all problems in a few seconds or less, while cvxopt takes quite a bit longer. I wonder what's the reason for this, since using cvxopt is quite a basic routine and optimized for quadratic programming. The only reason I can think of is that for SVMs many solutions are equal to 0 and somehow this is built into the libsvm solver since it's targeted for SVMs. Does anybody know if this is the case or are there other tricks too? Thanks!

fgpancorbo 09-02-2012 11:12 AM

Re: Why is CVXOPT slower than LIBSVM?
 
I have never used CVXOPT but libsvm implements this algorithm http://en.wikipedia.org/wiki/Sequent...l_optimization which is specific to the type of quadratic programming that is used in SVM.

htlin 09-02-2012 04:00 PM

Re: Why is CVXOPT slower than LIBSVM?
 
Quote:

Originally Posted by Willem (Post 4796)
libsvm runs all problems in a few seconds or less, while cvxopt takes quite a bit longer. I wonder what's the reason for this, since using cvxopt is quite a basic routine and optimized for quadratic programming. The only reason I can think of is that for SVMs many solutions are equal to 0 and somehow this is built into the libsvm solver since it's targeted for SVMs. Does anybody know if this is the case or are there other tricks too? Thanks!

Designing efficient SVM solvers has been an important and non-trivial research problem after SVM has been proposed. Modern, specific purpose SVM-solvers carefully takes into account the special optimization and machine learning properties of the formulation. For instance, the "sparsity" (not too many support vectors) is one thing often used to shrink the optimization problem, the "simplicity" (only one equality constraint in the dual problem and other constraints are "box" constraints) is another, the fact that the kernel matrix can be computed online when needed instead of offline is yet another. Hope this helps.


All times are GMT -7. The time now is 08:39 PM.

Powered by vBulletin® Version 3.8.3
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.
The contents of this forum are to be used ONLY by readers of the Learning From Data book by Yaser S. Abu-Mostafa, Malik Magdon-Ismail, and Hsuan-Tien Lin, and participants in the Learning From Data MOOC by Yaser S. Abu-Mostafa. No part of these contents is to be communicated or made accessible to ANY other person or entity.