-
Good deal, glad its working :)
For optimizing parameters you should definitely also check out the BOBYQA algorithm (http://dclib.sourceforge.net/algorithms.html#find_min_bobyqa) and the simpler find_min_single_variable() function for problems with just one variable. These are two deterministic algorithms for optimizing functions in the absence of derivatives. So they apply to the same...
2009-11-24 12:02:19 UTC by davisking
-
Wow, I didn't think the gamma had such a huge impact on the training time. Setting it from 0.005 to 0.0005 brought the training time for 365 samples from 4 hours to less than a second.
I thought I 'd use the default value from the example to start with and later optimize it as part of the evolutionary algorithm that will select good sets of features.
Haven't yet used the KRLS algorithm, but I...
2009-11-24 07:32:13 UTC by ulrichard
-
Are you trying to solve a classification or regression problem? By that I mean, are your labels just two classes or are they continuously varying numbers? The KRLS algorithm is usually a good regression algorithm and you can upper bound the number of "support vectors" it uses which tends to make it fast. Also, have you tried changing the gamma argument of the radial basis kernel to a smaller...
2009-11-23 12:25:03 UTC by davisking
-
Training with 98 features on my low powered netbook takes:
92 samples -> 1 minute
184 samples -> 25 minutes
365 samples -> 4 hours
1095 samples -> I stopped the process after 3 days and 48 hours of CPU
RAM never went above 50%, so I was not worried about swapping causing any troubles.
I use my netbook because later I want to deploy to an even lower powered embedded board...
2009-11-23 08:14:59 UTC by ulrichard
-
How long does it take to train with the rvm\_regression\_trainer? I'm a little surprised that it is taking too long on only 1000 samples. Usually, for me, it is quite fast on data sets of that size.
Also, how are you generating the support vectors? Dlib doesn't include a support vector *regression* class so either you are using some other tool to obtain them or something isn't quite...
2009-11-22 21:35:15 UTC by davisking
-
Hi,
I want to train a RVM with 100 features on 1000 samples. I quickly realized that it takes way too long to train.
Somewhere on the internet I read that a good approach is to partition the samples, and use the support vectors of these partitions as samples for the next step. Maybe repeat a few times until you reach the tip of the iceberg.
I found out that for my problem with 100...
2009-11-22 20:36:33 UTC by ulrichard
-
davisking committed revision 3271 to the dlib C++ Library SVN repository, changing 1 files.
2009-11-14 17:36:48 UTC by davisking
-
davisking committed revision 3270 to the dlib C++ Library SVN repository, changing 3 files.
2009-11-14 02:40:09 UTC by davisking
-
davisking committed revision 3269 to the dlib C++ Library SVN repository, changing 2 files.
2009-11-14 01:17:14 UTC by davisking
-
davisking committed revision 3268 to the dlib C++ Library SVN repository, changing 1 files.
2009-11-11 01:15:07 UTC by davisking
