Open in new window / Try shogun cloud
--- Log opened Thu Nov 15 00:00:17 2012
-!- n4nd0 [] has quit [Quit: leaving]00:29
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]00:29
shogun-buildbotbuild #171 of nightly_default is complete: Failure [failed tutorial]  Build details are at
wikingdont die01:46
shogun-buildbotbuild #172 of nightly_default is complete: Success [build successful]  Build details are at
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has joined #shogun09:12
blackburnwiking: what can we do with your oppan gangnam style solver?09:32
wikingsolveit :)09:41
blackburnheeey sexy solver09:41
blackburnoop oop09:41
wikingblackburn: new job?10:17
blackburnwiking: yes10:17
wikingblackburn: where to?10:17
wikingand congratz10:18
wikingis this the java company?10:20
blackburnwiking: yeah they do java and C#10:24
blackburnwiking: well they fucking write code10:24
blackburnnot fixing legacy bugs10:24
blackburnhere I fix legacy bugs10:24
blackburnI've been fscking fixing bugs for an year10:24
blackburnI just want to
wikingthis ascii serialization format is not the nicest parsable format10:31
blackburnwiking: ah btw could you please take a look at when you have time10:32
wikingah yeah a friend asked10:32
blackburnmay be you could have some thoughts what is wrong10:32
wikingdo you know a very efficient method to diagonilze small matrices10:32
wikingbut a lot of it10:32
wikingsmall matrix10:33
wikingso yeah dense matrix10:33
wikingok what could be wrong with this?10:33
wikingi mean it's just like design wise i should look10:33
wikingor what exactly you would like to talk about?10:33
blackburnwiking: yes, design wise10:33
blackburnsure I don't ask you to find a lowlevel mistake somewhere :)10:33
wikingok i'll check it10:34
blackburnokay actually what I know about diagonalization10:34
blackburneigen has bad and mkl has good10:34
wikingwhy we cannot have a serializer like10:37
wikingnumpy has10:37
blackburnwiking: what's the difference?10:39
wikingwell the format? :D10:43
wikinghave u seen the output of SerializableAsciiFile10:43
sonne|workwiking: we have simple load / save routines10:46
sonne|workserializable* stuff works for much more complex data10:47
sonne|worklike saving whole svm incl. kernel & features etc10:47
wikingsonne|work: for sgvector?10:50
sonne|workwiking: for *features10:50
sonne|workbut yes this should be moved into sgvector and friends10:50
wikingi just want to save a simple sgvector10:51
sonne|workbut we didn't have that at the time10:51
sonne|workbasically features were what SGVector etc were before10:51
wikingi'm chechking the code10:51
sonne|workwiking: btw I asked again10:51
sonne|work(not sure if you have seen)10:52
sonne|workanswer 'tonight'10:52
sonne|workbut that was yesterday10:52
wikingsonne|work: yeah i've seen it10:52
sonne|workno idea10:52
wikingi've even answered10:52
sonne|workheh I went to bed10:52
wikingso yeah i know that around 1pm CET10:52
wikingor somewhere there we should get some mail10:52
sonne|workor not10:54
sonne|workwhatever strange handling10:54
wikingsonne|work: why do you set/reset locale?11:00
sonne|work0,1 <-> 0.111:03
sonne|workin germany it is 0,111:03
sonne|workin us 0.1`11:03
sonne|workso better use C always11:03
wikingah got it11:03
wiking../shogun/io/File.h:33:22: error: expected class name11:06
wikingclass CFile : public CSGObject11:06
wikingok this is fucking hilarious11:09
wikinghahah i cannot include File.h in SGVector11:10
wikingit's like a snake biting his own tail11:14
wikingsomewhere we need a class CSGObject; ;)11:14
wikingbut it's hilarious11:15
wikingwhatever i start in the last 3 days nothing works11:15
wikingso i think i shouldnt do anything11:15
sonne|workthat all needs to be in a .cpp file11:21
sonne|workthen you can use forward declarations11:21
sonne|workI hate this dependency hell in C++11:22
wikingwhat all needs to be in a cpp file11:31
wikingi mean yes sure but in order to be able to declare the load/save function for SGVEctor i have to include the header11:32
wikingand now it's extremely fucked up11:32
wikingbecause no matter where i put the class CSGObject; declaration11:32
wikingi mean in which header11:32
wikingthe includes are first11:32
wikingso i have to put it before any include11:33
wikingok solved it11:33
-shogungit:#shogun- [shogun] vigsterkr opened pull request #825: Add load/save functionality for SGVector (master...latent)
wikingso here it goes... a bit of a fixing as well of precisions...12:31
wikingoh sorry some shit got left in sgvector...12:32
sonne|workwhat is the +GET_VECTOR(get_vector, get_int8_matrix, int8_t)13:04
sonne|work68 6913:04
sonne|work GET_VECTOR(get_vector, get_matrix, uint8_t) ?13:04
wikingline 68/69 was there13:06
wikingi've just added GET_VECTOR(get_vector, get_int8_matrix, int8_t)13:06
wikingto support int8_t13:06
sonne|workahh I see. another transition needed :/13:39
sonne|workwe decided to use overloading this GSoC13:39
blackburnsonne|work: in germany you also have comma instead of point?14:04
blackburnI thought it is used only in ussr :)14:06
sonne|workyou have it too :)(14:07
blackburnI don't like that pretty much14:07
blackburnand I don't like cyrillic letters actually14:08
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has quit [Quit: Page closed]14:09
sonne|workthat's the reason russians drink lot of vodka14:35
-!- n4nd0 [] has joined #shogun16:59
-!- n4nd0 [] has quit [Quit: leaving]17:35
wikingsonne|work: i guess there's still no news :D17:53
-!- koen2 [9232343c@gateway/web/freenode/ip.] has joined #shogun17:58
-!- blackburn [~blackburn@] has joined #shogun18:02
-!- koen2 [9232343c@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]18:02
wikingblackburn: yey18:16
wikingalthough i haven't been looking for it18:16
wikingbut i found a solution for solving nonconvex function with bundiling methods :)18:17
wikingyo anybody18:25
wiking(Fenchel Dual) is only true for convex functions?18:26
wikingwell i'll assume this :DD18:32
wikingand let's see where it takes me18:32
blackburnwiking: hey18:47
wikingi'm just checking if i can apply the dual form18:52
wikingas for that we have libqp as a solver18:52
wikingor do we have a primal solver for min_w \lambda/2 ||w||^2+max_{j=1..t}{c_j(w)}18:54
wikingi guess this is a no19:08
wiking :)19:08
blackburnwiking: wat ist da c_j?19:17
wikingc_j(w) = <a_j, w> + b_j19:19
wikingit's a cutting plane basically19:19
blackburnwiking: can't libqp solve this?19:27
wikingwell the dual form of it yes19:29
wikingi'm just not sure that the dual that i get with Fenchel Dual theorem stands if that function is non-convex19:30
blackburnI feel stupid19:32
wikingyeah me too19:33
wikingbut i'm just bruteforcing it19:33
wikingblackburn: what's gonna happen with that 2 PR? :)19:43
blackburnwiking: no idea I didn't take a look19:45
blackburndid soeren?19:45
blackburnwiking: why to load sgvector from file?19:46
wikingblackburn: why not19:46
wikingi needed now19:47
blackburnI see19:50
blackburnhow do you use it though?19:50
-shogungit:#shogun- [shogun] lisitsyn pushed 3 new commits to master:
-shogungit:#shogun- shogun/master df7368c Viktor Gal: Add load/save functionality for SGVector19:56
-shogungit:#shogun- shogun/master 887e00d Viktor Gal: Remove commented code from SGVector19:56
-shogungit:#shogun- shogun/master 2cd18bc Sergey Lisitsyn: Merge pull request #825 from vigsterkr/latent...19:56
wikingif ya wanna save a w of an svm20:10
wikingthere a new one20:12
-shogungit:#shogun- [shogun] vigsterkr opened pull request #826: Fix dual structure solvers to use CMath::sqrt (master...latent)
-shogungit:#shogun- [shogun] lisitsyn pushed 2 new commits to master:
-shogungit:#shogun- shogun/master 98b85f6 Viktor Gal: Fix dual structure solvers to use CMath::sqrt20:13
-shogungit:#shogun- shogun/master 097f9c8 Sergey Lisitsyn: Merge pull request #826 from vigsterkr/latent...20:13
wikingblackburn: here?20:50
wikingblackburn: why we only have gmnp_imdm ?20:51
blackburnmalozemov demyanov mitchell?20:51
wikingImproved Mitchell-Demyanov-Malozemov20:51
blackburnjust random guess - it is the best20:51
blackburnokay actually I asked myself why20:51
blackburnbut I didn't ask vojtech20:52
blackburnmay be he was lazy20:52
blackburnmay be it is the best approach20:52
wikingbecause he has other ones as well20:52
wikingin his other20:52
wiking    mdm  ... Mitchell-Demyanov-Malozemov20:52
wiking%    imdm  ... Improved Mitchell-Demyanov-Malozemov (default).20:52
wiking%    iimdm  ... Improved (version 2) Mitchell-Demyanov-Malozemov.20:52
wiking%    kozinec ... Kozinec algorithm.20:52
wiking%    keerthi ... Derived from NPA algorithm by Keerthi et al.20:52
wiking%    kowalczyk ... Based on Kowalczyk's maximal margin perceptron.20:52
blackburnwhat do you expect from say kozinec?20:52
blackburnI mean they are all nearly the same I believe20:53
wikingoh nothing20:53
wikingi was just wondering why20:53
blackburnI just mean20:53
blackburnnot really sure we can speed it up20:53
blackburnor anything20:53
blackburnbut yes would be nice to have all of them20:53
blackburnI wanted to implement that even20:53
blackburncopy and paste code I mean20:53
wikingit's just porting basicaly20:54
wikingat first step20:54
blackburnI am afraid it is vojtech's phd related20:54
blackburnif you know what I mean20:54
blackburnit looks more powerful when you have 6 methods20:54
blackburnand his phd is all about multiclass20:55
wikingok so now the last question20:58
wikingwhich is better to use20:58
wikinggnmp or libqp?20:58
wikingas i can see20:58
wikinggnmplib solves almost the same (actually exactly what i need)20:59
wikinglibqp is more general20:59
wikingah ok20:59
wikingno i dont get it20:59
blackburnyeah libqp is more general20:59
blackburndo you think gmnplib solves the same problem?21:00
wikingwell not the same problem21:00
wikingbut the non-convex thing21:00
blackburnis MC problem non-convex?21:00
wikingi have this problem21:00
wiking1/2 ||w||^2 + R(w)21:01
wikingusual bundle problem21:01
wikingand the dual form of this21:01
wikingis actually exactly what gnmplib solves21:01
blackburnhmm I see21:01
blackburnthen use it :)21:01
wikingu mean gnmplib?21:01
blackburnbut I didn't like gmnp when I use it21:01
wikingi wonder why uricamic used libqp and not gnmplib21:01
blackburnconverges for hours21:02
wikingi guess vojtec had a good reason for telling him to use libqp instead of gnmp21:02
wikingor? :)21:02
blackburnyeah he probably knows better21:02
blackburnthe prob is 1.021:02
wikingwell actually i'll write both21:02
wikingand one will be able to choose which QP solver to use21:03
blackburnwhat are you developing?21:03
wikingand we can even compare21:03
blackburnis that latent?21:03
wikingthis nonconvex bundle method21:03
blackburnoppan gangnam style?21:03
wikinglatent model's formulation is usually nonconvex21:03
blackburnI see21:03
wikingfor solving this21:03
wikingusually one uses somekind of an approx21:03
wikinglike CCCP21:03
wikingbut this solver would fucking solve it in your face :)21:04
wikingwithout approxx21:04
wikinggangnam style shit bro21:04
blackburnhow can they name method like that21:04
wikingblackburn: well yeah21:04
blackburnI have seen that before21:04
blackburnbut always laugh21:04
wikingme too21:04
wikingbut that's how it is21:04
blackburnI will call my next method ?????21:05
blackburnor ??????21:05
wikingstalin is cool21:05
wikingalmost fonetic of the cyrilic21:05
wikingso that u dont have to use utf821:05
wikingyeah indeed21:07
wikingmmm shitfuck21:07
wikingi think i had an enlightment moment now :)21:07
wikingwhere the fuck was this paper i was looking for now21:09
blackburnin journal21:09
blackburnhope that helps!21:10
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]23:36
--- Log closed Fri Nov 16 00:00:17 2012