Open in new window / Try shogun cloud
--- Log opened Thu Feb 09 00:00:19 2012
-!- n4nd0 [] has quit [Quit: Leaving]00:08
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]02:39
-!- dfrx [] has joined #shogun05:46
-!- n4nd0 [] has joined #shogun08:01
-!- n4nd0 [] has quit [Remote host closed the connection]08:14
-!- wiking [] has joined #shogun08:32
-!- wiking [] has quit [Changing host]08:32
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun08:32
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]08:44
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun08:48
-!- wiking [~wiking@huwico/staff/wiking] has quit [Remote host closed the connection]08:49
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun08:49
-!- n4nd0 [] has joined #shogun09:21
-!- n4nd0 [] has quit [Quit: AndroIRC - Android IRC Client ( )]09:54
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]09:59
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun10:30
-!- naywhayare [] has quit [Read error: Operation timed out]11:39
-!- naywhayare [] has joined #shogun11:42
-!- dfrx [] has quit [Quit: Leaving.]12:09
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]12:56
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun13:23
-!- n4nd0 [] has joined #shogun13:25
-!- wiking [~wiking@huwico/staff/wiking] has quit [Read error: Connection reset by peer]13:45
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun13:46
wikingmm those are 2011 ideas that have not been implemented?14:31
-!- axitkhurana [~akshit@] has joined #shogun14:44
-!- n4nd0 [] has quit [Read error: Connection reset by peer]14:50
-!- akshit_ [~akshit@] has joined #shogun14:58
-!- axitkhurana [~akshit@] has quit [Ping timeout: 248 seconds]15:02
-!- iglesias [] has joined #shogun16:50
-!- iglesias [] has quit [Ping timeout: 252 seconds]16:55
wikingwhy is that that when i create a kernel with the constructor with the left and right hand features shogun runs parallel, i.e. all 4 of my cores are used by shogun, while if i create the kernel with the default constructor and after that i call init function with the left and right-hand features, shogun runs only on one cpu17:01
wikingtalking here about multiclasssvm train/classification17:01
-!- akshit_ [~akshit@] has quit [Ping timeout: 240 seconds]17:56
-!- blackburn [~qdrgsm@] has joined #shogun18:43
blackburnsonney2k: we've got Cheng Soon as Action Editor :)18:44
-!- n4nd0 [] has joined #shogun19:16
wikingblackburn: heeey19:18
wikingblackburn: got another question... it's more about the internals of shogun...19:18
wikingblackburn: why is that that when i create a kernel with the constructor with the left and right hand features shogun runs parallel, i.e. all 4 of my cores are used by shogun, while if i create the kernel with the default constructor and after that i call init function with the left and right-hand features, shogun runs only on one cpu19:19
n4nd0hi there19:30
n4nd0yesterday I installed shogun with support for the matlab static interface19:30
n4nd0but I get some troubles trying out some of the examples19:31
n4nd0do I have to install matlab_modular or any other for them to work or is it another problem?19:31
n4nd0has someone installed before cplex in ubuntu?19:56
-!- n4nd0 [] has quit [Ping timeout: 252 seconds]20:44
blackburnwiking: no idea, will have to check later20:53
blackburnwiking: but you can use kernel.parallel.set_num_threads(4)20:53
wikingno help20:53
blackburnto set it explicitly20:53
wikingit seems there's something else20:54
wikingbecause during training time it's always sequential, i.e. never use threads20:54
wikingbut for some reason in case of histogramintersectionkernel there's a different speed depending on which constructor i'm using20:54
wikingalthough i know it doesn't make much sense, but this is what i get :S20:54
blackburnwhat exactly do you do?20:54
blackburnfunny I use HIK too now :)20:55
wikingwell i'm trying to test my features with different kernels20:55
wikingso my usual thing is to create a kernel and then do the init20:56
wikingi mean do it later the init, after i've created the right kernel20:57
wikingand so with HIK it's weird20:57
wikingwith poly kernel it's fine for instance20:57
blackburnwell there is no difference between ^ and ^20:57
blackburnI mean CKernel class is parallelized20:58
blackburnnot particular ones20:58
wikingi've checked the c++ code20:58
wikingso it's really weird20:58
wikingbut now it's not even that20:58
wikingbecause as far as i can see now20:58
wikingthe training time is never parallel20:58
wikingso it's always running on one cpu20:59
wikingbut somehow when i use one or the other way of constructing the HIK kernel20:59
wikingthe training time is very different20:59
wikingso let's say one way is like 10-20 seconds20:59
wikingthe other way it's at least 2-3 minutes20:59
blackburncould u please paste the code somewhere?20:59
wikingok sure20:59
wikingletme pastebin it21:00
wikingbut note that i'm using java api21:00
blackburnwell should not make any difference21:00
blackburnbut who knows :D21:00
wikinghere's the snippet of the code:
wikingand depending on line 35 boolean you'll get slow or 'norma'/fast training time21:04
blackburnhmm I see21:05
blackburnthe same cache size!21:05
blackburnstrange it could act in a different way21:06
wikingso i've tried to set everything the same21:06
wikingjust the way i construct the kernel is different21:06
wikingi've checked the cpp code21:06
wikingi cannot see the difference21:06
blackburnme neither21:07
wikingso now the only thing is somewhere in the java modular stuff21:07
wikingsomething happens21:07
wikingi've tried to turn on debugging mode21:07
blackburnno, can't be true21:07
wikingbut i couldn't get there21:07
wikingfor some reason the inheritance thingy is not working in java21:08
wikingi.e. i cannot call kernel.parallel.get_num_cpus ()21:08
wikingi have to do first a kernel.get_global_parallel21:08
wikingand then do the thingy there21:08
blackburndo you think it is parallel issue?21:09
wikingnah not really21:09
wikingas said21:09
wikingi've tried to monitor the number of threads21:09
wikingit's always 4, i.e. as many core i have21:09
wikingand in train time i never get more than 1 threads running21:10
blackburnLaRank is not parallelized so it is ok21:10
wikingso it's not really the parallelization problem21:10
wikingit's more like, why does it run slower21:10
blackburnin fact we have no svm with parallel training21:10
blackburnthat's really strange21:11
blackburnsorry I have to go now, but I'll check a little bit later21:11
wikingi think some flag is not turned on in some if branch21:11
wikingbut i'll try to do first the same thing with a silly21:11
wikingc++ code21:11
wikingif i can get the same thing21:12
wikingok no worries thnx!21:12
blackburnare you subscribed to the mailing list?21:12
blackburnI would suggest you to create github issue or write to the mailing list21:12
blackburnif you want21:12
wikingah ok i'll do the github issue then21:16
blackburnit would be easier then to take care :)21:17
-!- n4nd0 [] has joined #shogun21:25
blackburnn4nd0: what is the problem with matlab?21:30
n4nd0blackburn, so there were a couple of issues21:30
n4nd0blackburn, the one that I was concerned about was that the lpboost example does not run21:31
n4nd0blackburn, it says sth like classifier LPBOOST not found21:31
blackburnn4nd0: matlab static?21:32
n4nd0blackburn, I have looked it up on google and it seems that cplex is required for it to work so I am with that right now21:32
n4nd0blackburn, yes, I am using matlab static this far21:32
blackburnah yes21:32
n4nd0blackburn, should it be an issue static or modular just to try the example out?21:32
n4nd0blackburn, I found that example under matlab_static so I assumed that it should work there also21:32
n4nd0there too*21:33
blackburnwell should work indeed21:33
blackburnbut cplex is required I guess21:33
blackburnlemme check21:33
n4nd0let's see then if I managed to get cplex to work21:33
n4nd0yeah I found that in this forum21:34
blackburnyou could check it by yourself too :) just take a look on classifier/LPBoost.h21:34
blackburnthere is a #ifdef HAVE_CPLEX21:34
blackburnso just install cplex and ./configure, etc21:34
n4nd0didn't think of looking it up in the code :-O21:35
n4nd0the other problem I found is that all the data directories are empty21:36
n4nd0I made first a defauult installation: ./configure, make, make install21:36
n4nd0and later another one with ./configure interfaces=matlab_static21:37
n4nd0but those directories are empty in any case21:37
blackburngit submodule update --init21:39
blackburnor git submodule --init update21:39
blackburnthat's a typical problem for sure21:40
blackburnit is because shogun and shogun-data are two different repos21:40
n4nd0mmm I installed it from the compressed directory so I must first activate the repo21:41
blackburnuh oh21:41
blackburnthen I would suggest you to git clone shogun :)21:41
blackburnno idea how to activate it from compressed one21:41
n4nd0all right21:42
n4nd0I will take a look to that once I am done with the cplex issue21:42
n4nd0thank you very much :D21:42
blackburnyou're welcome :)21:43
blackburnwiking: did you get any progress with this strange issue? :)21:43
wikingblackburn: none yet... postponed it... now i'm using the lil hack which is faster... since i would really like to get to the point where i can try out MKL on multi class :) that was the main reason i've started to write my shogun wrapper in my code... ;)21:44
blackburnwiking: okay sure21:44
wikingaaha coool21:45
wikingi've just realized that i can iterate through the features i've appended to the CombinedFeatures21:46
blackburnI feel butthurt remembering that there was a leak in java interface :D21:46
blackburnit was reproduced with HUGE mkl script21:46
blackburnokay heading to bed right now21:47
blackburnsee you21:47
wikinglet's see what happens21:47
wikingi really wouldn't like to rewrite now the whole thing in python or c++ :>21:48
wikinganyhow gnite21:48
blackburnbug is not here anymore21:48
blackburnI hope21:48
blackburnit was CustomKernel issue21:48
blackburnbut consider using python anyway, it is much cooler for shoguning :)21:48
wikingi'm using mahout as well21:49
wikingso i need a common ground :>21:49
blackburnI've never used it21:49
wikingit's cool21:49
blackburnah yeah makes sense21:49
wikingbut it's mainly developed for text processing21:49
blackburnis it for distributed things?21:49
wikingbased on hadoop21:50
wikingbut for me that's just for acquiring the features21:50
blackburnI see21:50
wikingso it's good otherwise with shogun21:50
blackburnbut they don't have many svms :D21:50
wikinguntil now i was using weka :>21:50
wikingbut it's kind of like sucks as well with their limited kernels...21:51
blackburnwhat do you mean saying limited?21:51
wikingwell like i really wanted HIK21:51
wikingand now i really hope i'm getting some boost with MKL21:51
blackburnwith which features do you use HIK?21:51
wikingah i'm doing image classification21:51
blackburncolor histograms?21:52
wikingsift histogram21:52
wikingso basically bag of visual words...21:52
blackburnI used to do similar thing21:52
wikingah and what's now? :)21:52
blackburnactually I'm going to implement some IKSVM a little bit later21:52
wikingmmm paper?21:53
blackburnthis one21:53
wikingash a cvpr one :)21:54
wikingcool i'll give it a read21:54
wikingany plan for latent svm21:54
blackburnnot really, I don't know much about latent svms yet21:56
blackburnI used to implement dim reduction things so I'm pretty new to svms still hah21:56
blackburnokay now going to sleep really :)21:56
blackburnsee you later21:56
blackburnwe should exchange experience later :)21:57
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]21:57
-!- wiking_ [~wiking@huwico/staff/wiking] has joined #shogun22:19
-!- wiking [~wiking@huwico/staff/wiking] has quit [Ping timeout: 246 seconds]22:23
-!- wiking_ [~wiking@huwico/staff/wiking] has quit [Ping timeout: 248 seconds]22:24
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun22:31
-!- wiking [~wiking@huwico/staff/wiking] has quit [Client Quit]22:31
-!- n4nd0 [] has quit [Ping timeout: 265 seconds]22:58
-!- n4nd0 [] has joined #shogun23:10
--- Log closed Fri Feb 10 00:00:19 2012