Re: [Dclib-devel] about dlib C++ library

[Davis K. <dav...@us...>]

Glad to be of help :)

You might want to take a look at this code specifically:
http://dclib.sourceforge.net/dlib/svm/svm.h.html#do_train.  That is the main
block of code that does SMO.  The main action is inside the while loop where
it alternates selecting the two best alphas and performing the optimization
step Platt talks about.   The code inside the find_working_group() function
is different from what Platt talks about but still similar.  For the details
behind why that code is the way it is you need to read the paper "Chang and
Lin, Training {nu}-Support Vector Classifiers: Theory and Algorithms"

Cheers,
Davis

On Sun, Jul 5, 2009 at 10:17 PM, Lazaro Bustio Martinez
<lb...@gm...>wrote:

> Hello Davis... yes, it help me a lot. You'll see, I'm trying to find a
> standard SMO algotithm implementation like Platt states in his paper. I'm
> studying SMO complexity, basically in how many time SMO spent calculating
> kernel function, updating caches and so on; but unfortunately I could not
> find any straight SMO implementatios as Platt states in his paper.
> Your suggestions help me a lot, thanks again. If you could give some more
> information, I´ll appreciate a lot.
>
> Best regards..
>
> Lazaro
>
>
>
>
> 2009/7/5 Davis King <dav...@us...>
>
> If you are looking at the svm_ex.cpp (
>> http://dclib.sourceforge.net/svm_ex.cpp.html) example program then you
>> are seeing the svm_nu_trainer object (see
>> http://dclib.sourceforge.net/ml.html#svm_nu_trainer).  It uses the
>> standard SMO algorithm and there are some more relevant papers sited in the
>> link above.
>>
>> If you want to know a lot more about SVMs and SMO then the best resource
>> available is probably the book Learning with Kernels: Support Vector
>> Machines, Regularization, Optimization, and Beyond by Schlkopf and Smola.
>>
>> Hope that helps.
>>
>> -Davis
>>
>>
>> On Sun, Jul 5, 2009 at 2:38 AM, Lazaro Bustio Martinez <lb...@gm...
>> > wrote:
>>
>>> Hello... I was looking for an SVM implementation and I found yours... but
>>> I can't see wich algorithm for SVM training did you used. Can you give me
>>> more information about it?
>>>
>>> Excuse me my english...
>>>
>>> Thanks a lot.
>>>
>>> Lazaro Bustio
>>>
>>
>>
>