Open in new window / Try shogun cloud
--- Log opened Mon Mar 26 00:00:19 2012
makokalhey are they any mentors online?00:06
makokalesp for GPs00:07
blackburnI am afraid GPs mentor could be reached only by email00:07
blackburnhave never seen him there00:07
makokalkk, mail on the way00:08
blackburnmakokal: I can't be considered as mentor but I am responsible for any development00:09
makokaloh, I just want to ask about clarification on imeplementing GPs for gsoc this year00:09
blackburnyeah ask00:10
makokalon the desc, on regression is mentioned, I was wondering if using GPs for classification will be part of the project00:11
makokalfor dev it should not mean much, but just a clarification00:12
makokalas I am working on finalizing my document for applying for this00:12
blackburnit is up to you generally00:13
blackburnI mean if you want to do both regression and classification - it would be better for sure00:13
makokalok, I will just add it on the final bits(so depending on the progress, add it or not)00:14
blackburnbut this should be clarified with Oliver as well00:14
makokalya am mailing him at the moment00:14
blackburnand consider to come with small patch! it is a requirement00:15
-!- makokal [] has left #shogun []00:49
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]01:09
-!- wiking_ [~wiking@huwico/staff/wiking] has quit [Quit: wiking_]01:40
-!- n4nd0 [] has quit [Ping timeout: 245 seconds]01:48
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun02:27
-!- PhilTillet [~Philippe@] has quit [Remote host closed the connection]02:39
-!- vikram360 [~vikram360@] has quit [Ping timeout: 252 seconds]04:24
-!- gsomix [~gsomix@] has joined #shogun05:23
-!- gsomix [~gsomix@] has quit [Read error: Operation timed out]06:11
-!- n4nd0 [] has joined #shogun07:12
-!- wiking [] has joined #shogun07:32
-!- wiking [] has quit [Changing host]07:32
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun07:32
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]08:00
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun08:01
-!- wiking [~wiking@huwico/staff/wiking] has quit [Ping timeout: 264 seconds]08:05
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun08:11
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has joined #shogun08:18
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]08:34
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun08:39
-!- wiking [~wiking@huwico/staff/wiking] has quit [Client Quit]08:40
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds]08:50
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun09:03
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds]09:13
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun09:26
blackburnmonday is a hard day lol09:31
n4nd0blackburn: feeling sleepy? :P09:36
blackburnn4nd0: strange but not at all - funny thing is only 9 here are idling09:37
blackburnbut before it was ~20 :)09:37
n4nd0that's actually right, funny09:37
blackburnn4nd0: what time is it there you are?09:38
blackburnI am curious whether you practice to sleep sometimes :D09:38
n4nd0blackburn: haha I do, I do09:39
n4nd0but not that much actually, I've always been ok not sleeping that much09:39
n4nd0it is 9:4009:39
blackburnfor me it is ok as it is 11.39 here09:39
blackburnbut it is 8.39 in Sweden?09:40
n4nd0mmm no09:40
blackburn9? :)09:40
n4nd0it is 9:40 in Sweden09:40
n4nd0yesterday we changed the time in here, summer time! :)09:40
blackburnah yeah09:40
n4nd0did you change it too there?09:41
blackburnthing our gov't decided to abandon09:41
blackburnno, we have static const time now09:41
-!- harshit_ [~harshit@] has joined #shogun09:41
harshit_hey blackburn09:41
harshit_did you hav  a look at my pull request ?09:41
blackburnnot really09:42
blackburnI think Soeren should take a look as well09:42
harshit_okay .. So what do you think how long will it take to merge into shogun09:43
harshit_Also wanted to discuss Deep learning idea !09:44
blackburnI will check the code again a little bit later09:46
harshit_okay, with whom do you think i should discuss my deep learning idea !09:47
harshit_soren ?09:47
blackburnyeah probably09:48
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun10:09
-!- harshit_ [~harshit@] has quit [Ping timeout: 276 seconds]10:15
-!- harshit_ [~harshit@] has joined #shogun10:16
-!- harshit_ [~harshit@] has quit [Ping timeout: 244 seconds]10:23
CIA-64shogun: Soeren Sonnenburg master * r001c17c / (8 files in 5 dirs):10:30
CIA-64shogun: Merge pull request #384 from gsomix/python3_interface10:30
CIA-64shogun: Partial support for python3 interface -
blackburnsonney2k: but he reported a problem on strings?10:31
CIA-64shogun: Soeren Sonnenburg master * r97e8e21 / (13 files in 7 dirs):10:31
CIA-64shogun: Merge pull request #386 from iglesias/qda10:31
CIA-64shogun: QDA -
blackburnmerge machine10:32
n4nd0oh QDA :)10:34
@sonney2kblackburn, btw I tried this director stuff10:47
@sonney2kgot only crashes though10:47
blackburnsonney2k: was it working? I am very curious with it10:47
blackburnbad bad10:47
@sonney2kI got some infinite recursions10:47
blackburnwell lets extend SWIG then10:47
@sonney2ki will add this class nevertheless10:47
@sonney2kbut ifdef'd10:47
blackburnI see10:48
-!- pluskid [~chatzilla@] has joined #shogun10:51
shogun-buildbotbuild #549 of python_static is complete: Failure [failed compile]  Build details are at  blamelist: gsomix@gmail.com10:56
shogun-buildbotbuild #620 of libshogun is complete: Failure [failed compile]  Build details are at  blamelist:, fernando.iglesiasg@gmail.com10:58
n4nd0blackburn: why does shogun-buildbot usually fail to compile the first time?10:58
blackburnn4nd0: he doesn't actually10:58
blackburnthis time he failed because of erroneous code :)10:58
n4nd0blackburn: really? our fault then?10:59
blackburnn4nd0: it seems you forgot to import cblas10:59
n4nd0blackburn: mmm why does it get to compile without problems in my machine then?11:00
@sonney2kwell it build *every* commit11:00
@sonney2kso if someone commits something crappy11:00
blackburnn4nd0: different configuration11:00
blackburnthat is usual ^11:00
@sonney2kand fixes it afterwards it will first fail and then succeed11:00
blackburnsonney2k: it seems that python stuff should be elaborated more11:00
@sonney2kif people were using git amend to merge fixes together this wouldn't happen - but hey I don't complain11:01
blackburnn4nd0: have you secured QDA with HAVE_LAPACK?11:01
blackburnn4nd0: no include for lapack.h11:02
blackburnprobably that is the reason11:02
n4nd0I see11:02
CIA-64shogun: Soeren Sonnenburg master * rdfbb2fd / (9 files in 3 dirs): add support for director kernel -
blackburnsonney2k: I think if we manage to add these directors it should be an option11:06
blackburn--enable-directors or so11:07
pluskidDo you have procedure to evaluate the performance of clustering result in shogun?11:07
blackburnsonney2k: ah yes I did implement a proof-of-concept for libedrt11:07
blackburnworks even faster11:07
blackburnand totally flexible11:07
blackburnpluskid: no you are welcome to add any to shogun/evaluation ;)11:08
@sonney2kblackburn, yeah - it is currently on but I certainly will add this ifdef stuff11:08
pluskidblackburn: OK, I'll do that for my GSoC sample patch11:08
pluskidBTW: I get a much much more simple patch11:09
pluskidin your example code11:09
pluskidfor the matlab and octave module11:09
pluskidthe content in clustering_hierarchical.m is for kmeans clustering11:10
pluskidwhile the content for  clustering_kmeans.m is for hierarchical clustering11:10
pluskidIt should be a simple renaming11:10
pluskidbut I think this should be fixed11:11
blackburnfeel free to do that using pull requests11:11
pluskidsure :)11:11
blackburnhaha we have one task that would scary any student from participating in GSoC11:14
blackburn there is a race condition somewhere LOL11:14
n4nd0blackburn: oh, any idea to reproduce it?11:15
blackburnn4nd0: oh better do not do that11:16
shogun-buildbotbuild #550 of python_static is complete: Success [build successful]  Build details are at
n4nd0blackburn: haha already tried?11:16
blackburnn4nd0: well easy to reproduce but 5K source lines there11:17
blackburnwe figured it out and nothing mor11:17
blackburnresults with 2 threads are totally different11:17
blackburnthan with 111:17
n4nd0it could be useful anyway to take a look to that code for SO11:19
shogun-buildbotbuild #390 of java_modular is complete: Failure [failed compile]  Build details are at  blamelist: gsomix@gmail.com11:23
blackburndidn't I fix it??11:23
-!- romi_ [~mizobe@] has quit [Ping timeout: 246 seconds]11:25
-!- pluskid [~chatzilla@] has quit [Quit: ChatZilla [Firefox 11.0/20120312181643]]11:29
-!- wiking [~wiking@huwico/staff/wiking] has quit [Remote host closed the connection]11:36
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun11:37
-!- romi_ [~mizobe@] has joined #shogun11:37
-!- gsomix [~gsomix@] has joined #shogun11:48
blackburngsomix: notwithstanding that you've got some troubles with strings your patch is in11:51
blackburngsomix: what is the exact problem with strings here?11:51
n4nd0blackburn: should I fix something in QDA then? this include lapack?12:04
-!- romi_ [~mizobe@] has quit [Ping timeout: 246 seconds]12:04
-!- romi_ [~mizobe@] has joined #shogun12:04
shogun-buildbotbuild #391 of java_modular is complete: Success [build successful]  Build details are at
blackburnn4nd0: yes12:17
n4nd0n4nd0: any to configure locally like in buildbot so I'll be sure what compiles here will do it there?12:18
n4nd0blackburn: or any way to test it at least12:18
blackburnshould be a way, but I don't know exactly12:19
n4nd0mmm ok12:19
n4nd0so include lapack.h in QDA.h should get rid of that12:19
blackburnif my guess is right12:20
blackburngsomix: I see..12:32
gsomixblackburn, I think problem is in the swig_typemaps file. I didn't realized something.12:34
n4nd0blackburn: I'll open now a pull request with it, can you try to merge it? apart from that I've added do_free member to SGNDArray similar to the one in SGMatrix and SGVector12:35
blackburnn4nd0: I'll be able to merge a little bit later12:36
n4nd0blackburn: perfect ;)12:37
blackburnactually when I get to my notebook :)12:37
gsomixblackburn, O_o12:50
blackburngsomix: what's up?12:52
gsomixblackburn, it's work, but...12:52
gsomixhowever, for some reason installing libraries is not working properly.12:52
blackburnwhat kind of libraries?12:53
gsomixblackburn, _modshogun.12:53
blackburnwhat is the error then?12:53
blackburnit is pretty clear that typemaps get broken with python312:54
blackburnwhile string list is not being recognized12:56
gsomixblackburn, it does not matter. I just processed incorrectly new unicode strings.12:56
blackburnI can suggest you to take a look at static int is_pystring_list(PyObject* obj, int typecode)12:56
gsomixblackburn, even last night, as it turned out.12:56
gsomixblackburn, nevermind12:56
blackburnI am pretty sure problem is related to such detection12:57
gsomixI just could not check, because libraries are not installed.12:57
gsomixbut now everything is working12:57
blackburnI don't get it - is it working now?12:58
gsomixblackburn, there are not problems with strings.12:59
blackburnis there any problem then?12:59
gsomixa little13:00
-!- romi_ [~mizobe@] has quit [Ping timeout: 246 seconds]13:00
-!- romi_ [~mizobe@] has joined #shogun13:00
gsomixblackburn, now running over 90 percent of the tests.13:02
-!- pluskid [~chatzilla@] has joined #shogun13:06
-!- pluskid_ [] has joined #shogun13:24
-!- pluskid__ [~chatzilla@] has joined #shogun13:25
-!- pluskid [~chatzilla@] has quit [Ping timeout: 252 seconds]13:25
-!- pluskid__ is now known as pluskid13:25
-!- pluskid_ [] has quit [Ping timeout: 264 seconds]13:29
gsomixsonney2k, hey13:30
CIA-64shogun: iglesias master * r388d734 / src/shogun/lib/DataType.h : + dofree member in SGNDArray -
CIA-64shogun: iglesias master * r8d16f45 / src/shogun/classifier/QDA.cpp : + QDA include to avoid failure in buildbot and ~ ndarray destroy -> free -
CIA-64shogun: Soeren Sonnenburg master * rdefc6b5 / (src/shogun/classifier/QDA.cpp src/shogun/lib/DataType.h):13:48
CIA-64shogun: Merge pull request #393 from iglesias/qda13:48
CIA-64shogun: QDA -
CIA-64shogun: Soeren Sonnenburg master * r16be749 / (2 files):13:49
CIA-64shogun: Merge pull request #392 from pluskid/master13:49
CIA-64shogun: The filenames for two example code are wrong -
shogun-buildbotbuild #622 of libshogun is complete: Success [build successful]  Build details are at
blackburnsonney2k: feel like being a merge machine? :D13:59
pluskidwhy java modular failed to build?14:07
blackburnit was built succesfully later14:08
pluskidbut I failed on my local box14:09
blackburnwhat is the error?14:09
pluskidsomething related to SWIG?14:11
blackburnwell yes but could you please try to clean current build and try to build it again?14:12
n4nd0blackburn: I think that maybe we have a bug in MulticlassMachine14:19
blackburnn4nd0: could be14:19
n4nd0blackburn: if apply is passed without arguments, it doesn't use the features that were used for training14:19
n4nd0blackburn: if perfectly ok if apply is called again with apply(feats_train), the same ones used for training for example14:20
blackburnehm.. why?14:20
n4nd0blackburn: I don't know yet ... I think I was messing up with one_vs_one but the problem is the same in one_vs_rest14:20
n4nd0blackburn: I have to check whether the MulticlassMachine loses reference to CFeatures somewhere or it happens in m_machines14:22
blackburnwhen you train()14:22
blackburnand apply() it14:22
blackburnwhat is going on?14:22
n4nd0seg fault14:23
blackburnoh fine14:23
n4nd0features is null when we do m_machines[i]->apply()14:24
n4nd0so the outputs[i] are all null14:24
blackburnI see14:24
n4nd0and when we want to count the votes ... boom! :O14:25
n4nd0I can work on it14:26
blackburnyeaah please do14:27
blackburnn4nd0: probably once it trained for some features it should init_machines_for_apply14:28
n4nd0blackburn: I will take a look to it later today, have to do some other stuff now14:30
blackburnok I'll add this once I get home14:31
n4nd0maybe I should push then my changes in my branch and you modify it there?14:32
n4nd0or are you pushing to master and later I can rebase?14:32
blackburnyes I can push to master14:33
blackburnso just rebase, do not fix it in your branch14:33
-!- pluskid [~chatzilla@] has quit [Quit: ChatZilla [Firefox 11.0/20120312181643]]14:34
-!- pluskid [~chatzilla@] has joined #shogun14:40
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has quit [Quit: Page closed]14:40
shogun-buildbotbuild #401 of csharp_modular is complete: Failure [failed compile]  Build details are at  blamelist: fernando.iglesiasg@gmail.com14:49
n4nd0shogun-buildbot: what have I done now? :(14:56
n4nd0shogun-buildbot: work!14:56
shogun-buildbotWhat you say!14:56
-!- vikram360 [~vikram360@] has joined #shogun15:04
shogun-buildbotbuild #389 of python_modular is complete: Success [build successful]  Build details are at
-!- PhilTillet [~Philippe@] has joined #shogun15:40
-!- PhilTillet is now known as RagnarX15:41
-!- RagnarX [~Philippe@] has quit [Client Quit]15:41
-!- romi_ [~mizobe@] has quit [Ping timeout: 246 seconds]15:43
shogun-buildbotbuild #393 of java_modular is complete: Failure [failed compile]  Build details are at  blamelist:,, sonne@debian.org15:52
-!- pluskid_ [~chatzilla@] has joined #shogun15:53
-!- pluskid [~chatzilla@] has quit [Ping timeout: 240 seconds]15:54
-!- pluskid_ is now known as pluskid15:54
-!- romi_ [] has joined #shogun15:56
-!- pluskid_ [] has joined #shogun15:58
-!- pluskid_ [] has quit [Client Quit]16:00
-!- pluskid [~chatzilla@] has quit [Ping timeout: 260 seconds]16:02
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]16:16
shogun-buildbotbuild #390 of python_modular is complete: Failure [failed test_1]  Build details are at  blamelist:,, sonne@debian.org16:19
-!- pluskid [] has joined #shogun16:24
pluskidis there a recent API change to rename classify() to apply() in SVM?16:24
pluskidI find I could not run the script16:25
pluskidunless I change its calling of svm.classify() to svm.apply()16:25
shogun-buildbotbuild #386 of ruby_modular is complete: Failure [failed compile]  Build details are at  blamelist:,, sonne@debian.org16:31
n4nd0sonney2k: any idea what's going on? all the builds are failing to compile :(16:32
-!- pluskid [] has quit [Quit: ChatZilla [Firefox 11.0/20120312181643]]17:07
-!- blackburn [~qdrgsm@] has joined #shogun17:30
-!- PhilTillet [~Philippe@] has joined #shogun17:32
blackburnhah all is failing lol17:32
n4nd0blackburn: what did I do man?17:34
blackburndirector kernel fails17:34
n4nd0shogun-buildbot blames me all the time17:34
blackburnnevermind, rather check logs17:34
n4nd0how do we fix that anyway?17:38
blackburnno idea I'd rather ifdef that kernel for now17:41
blackburnhaha 614 warnings17:42
n4nd0did I hear record?17:44
blackburnn4nd0: we had >2.5k once17:44
blackburnwhen Soeren turned on doc warnings :D17:45
n4nd0blackburn: shit, we have to make more efforts then :P17:45
blackburnwell it is easy - just insert a couple of useless variables somewhere in template class17:46
-!- harshit_ [~harshit@] has joined #shogun17:47
harshit_hey n4nd017:49
n4nd0harshit_: hey man17:49
harshit_whats going on .. are you coding any thing for shogun for now ?17:50
harshit_Is there any thing you knw i can work on17:50
n4nd0harshit_: so blackburn had an idea about some scripting work17:51
n4nd0harshit_: to integrate tests with some datasets17:52
harshit_in which language ?17:52
n4nd0harshit_: idk if you would like that or you prefer we look for another thing more for C++17:52
n4nd0that would probably depend more on how you want to do it but I'd that the most appropriate is to use a shell script probably17:53
blackburnnice plots with iris dataset would be a nice example17:53
harshit_not much experience with shell script .. but i was thinking to incorporate any nice math library other than cblas and lapack17:54
harshit_will that example work, if i code it in C++17:54
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun17:54
n4nd0what do you think blackburn?17:55
n4nd0do we need other math libraries?17:55
blackburnn4nd0: no idea what can be useful17:55
harshit_while porting code from MATLAB i found it really difficult to find functions in shogun17:56
harshit_so i think that would be useful17:56
n4nd0harshit_: but those functions existed right?17:56
harshit_not really17:57
harshit_like for finding solutions for Ax=B17:57
harshit_i had to use pinv(A)*B17:57
harshit_and i am working on the which part, blackburn17:57
n4nd0I think that maybe cblas and lapack are a bit difficult to use17:58
n4nd0but probably the stuff is there17:58
blackburnno, lapack is wonderful balanced for me18:00
harshit_yeah, cblas and lapack are the most powerful ones but still variety is always good18:00
n4nd0yeah might be18:01
n4nd0harshit_: wiking and I were talking some days ago about QP solvers18:02
n4nd0now we depend on private solvers (mosek or CPLEX)18:02
n4nd0would be nice to have something of that18:02
n4nd0but I don't know how difficult could that be18:02
blackburnto integrate shouldn't be difficult, but to use..18:02
harshit_havn't had much experience with optimization, but still nice idea .i'll think about it18:04
-!- PhilTillet [~Philippe@] has quit [Ping timeout: 248 seconds]18:05
CIA-64shogun: Sergey Lisitsyn master * r904696a / src/shogun/lib/DataType.h : Warnings removal -
CIA-64shogun: Sergey Lisitsyn master * rc2b5c34 / src/shogun/machine/MulticlassMachine.cpp : Fixed mc machine training -
n4nd0wiking: do you remember the name of the QP library we talked about? I think that one was open source18:07
n4nd0yeah that one18:09
blackburnvojtech really rocks18:10
blackburnhe is an author of libqp, libocas and a lot of other things18:12
harshit_i think there is something related to libqp already in shogun ..18:12
blackburnyes libqp itself18:12
n4nd0but can it be used?18:14
blackburnwhy not18:14
n4nd0or is it left to do wrappers and stuff?18:14
n4nd0mmm I don't understand then why some parts need CPLEX18:14
blackburnwell it is a general library :)18:14
n4nd0maybe CPLEX does more than that18:15
blackburnI have no expertise with it :(18:16
blackburnn4nd0: actually I'm moving towards to the way libqp and libocas are implemented18:19
n4nd0blackburn: what do you mean with moving towards their way?18:19
blackburncallbacks and other such stuff18:19
n4nd0is it a kind of event-driven programming?18:20
blackburnyeah I'll show you in a min18:20
blackburnok have to go18:28
blackburnsee you18:28
n4nd0any special interesting part with callbacks?18:29
n4nd0ok nvm18:29
n4nd0will check it ;)18:29
harshit_n4nd0 , started with gsoc application ?18:32
n4nd0not yet18:33
harshit_was thinking to start today18:33
harshit_there are 6-8 days18:33
-!- blackburn [~qdrgsm@] has quit [Ping timeout: 276 seconds]18:33
harshit_but i heard somewhere that its good to submit early18:33
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]18:34
n4nd0I think they will wait to check the applications18:35
n4nd0at least that makes sense for me18:35
n4nd0but who knows18:35
harshit_feeling depressed  :(18:36
harshit_didnt find anyone who could help me with my deep learning idea18:36
harshit_i think its too young too be incorporated into shogun18:37
n4nd0maybe it could be interesting to do it after GSoC or sth18:38
harshit_yeah, but i thought my research and gsoc could go along with each other..18:39
harshit_never mind18:39
n4nd0look at this way18:40
n4nd0you have the oportunity of learning other stuff18:41
harshit_yeah, right ..18:42
harshit_n btw do you ppl speak spanish in sweden18:43
harshit_i remember you saying that you use spanish keyboard18:44
n4nd0my laptop is Spanish18:44
n4nd0I'm Spanish :)18:44
n4nd0I speak Spanish with my Spanish friends in Sweden :P18:44
n4nd0but people speaks Swedish here normally18:45
harshit_okay actually i was also learning spanish :18:45
harshit_its a really nice language18:45
harshit_must say18:45
harshit_not better than C++ though18:45
n4nd0I also prefer C++18:47
n4nd0I don't like that much to speak Spanish actually18:47
n4nd0just sometimes18:47
-!- harshit_ [~harshit@] has quit [Quit: Leaving]18:56
-!- n4nd0 [] has quit [Quit: leaving]19:40
CIA-64shogun: Soeren Sonnenburg master * r0f26327 / src/shogun/kernel/DirectorKernel.h : fix directorkernel init warning -
shogun-buildbotbuild #190 of nightly_none is complete: Failure [failed compile]  Build details are at
shogun-buildbotbuild #189 of nightly_all is complete: Failure [failed compile]  Build details are at
-!- vikram360 [~vikram360@] has quit [Ping timeout: 252 seconds]21:26
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has joined #shogun21:28
blackburnsonney2k: re to flexibility of libedrt21:36
blackburncheck compute_kernel and LocallyLinearEmbedding21:36
blackburnit can be easily changed with custom kernel or some other kernel21:37
CIA-64shogun: Soeren Sonnenburg master * re6d427d / src/shogun/kernel/DirectorKernel.h : copy some virtual function calls to director kernel to fix some errors -
@sonney2kblackburn, not now - I have to look at newtonsvm21:39
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds]21:40
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]21:41
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun21:41
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has joined #shogun21:44
-!- n4nd0 [] has joined #shogun21:49
n4nd0blackburn: I have just seen you pushed a change on mc :)21:58
blackburnn4nd0: yes21:58
blackburnand changed my name on github lol21:58
n4nd0blackburn: I see, surname style21:59
blackburnn4nd0: I wonder do you know how to spell it? ;)21:59
n4nd0let me try22:00
n4nd0blackburn: shit, I think I am going to do it wrong22:00
blackburntry ;)22:01
n4nd0blackburn: so I rather not to do it :P22:01
@sonney2kblackburn, did anyone here talk to harshit (newton svm ...)22:02
blackburnsonney2k: yes22:02
@sonney2khis code is not yet perfect but not too bad for first attempt22:02
blackburnsonney2k: why do you ask?22:02
@sonney2kit is very inefficient at the moment (computing the whole feature matrix)22:02
@sonney2kbut that can be fixed...22:03
@sonney2khe seems to know what he is doing22:03
blackburnn4nd0: let me try to describe - lee-see-ts-yh-n22:03
@sonney2kblackburn, wasnt' he the one wanting to integrate NNs?22:03
n4nd0blackburn: lisitsyn22:03
blackburnsonney2k: yes deep learning22:03
blackburnsonney2k: he wants still22:04
n4nd0sonney2k: he was asking about it today22:04
n4nd0sonney2k: do you think it could be a possible project?22:04
blackburnI have doubts22:05
blackburnnobody asked me yet but I have doubts :D22:05
n4nd0about what?22:05
blackburndeep learning22:05
n4nd0about the project you mean in deep learning?22:05
n4nd0what doubts?22:05
@sonney2kwe have no mentor for that ...22:06
n4nd0aham I see22:06
blackburnI am not an expert in anything22:06
blackburnand sonney2k is probably not an expert in deep learning22:06
blackburnso hard to decide22:06
n4nd0I remember that someone suggested in the mailing list that it was not required to have a mentor to carry out the project22:07
n4nd0a mentor specialized in the topic22:07
@sonney2kn4nd0, yeah gunnar did22:07
@sonney2k(the shoGUN gunnar )22:07
blackburnsonney2k: do you agree?22:07
@sonney2kproblem is that this is only true if we know a high quality C/C++ impl. of deep learning stuff22:08
@sonney2kwhich would merely need porting/shogun integration22:08
@sonney2kI know that real experts are the people doing lush and torch522:09
blackburnI can hardly imagine yann lecun mentoring for us22:09
@sonney2kbut lush is some lisp dialect stuff22:09
@sonney2kand ronan collobert (torch main author) comes up with torch6 soon I guess :)22:10
blackburnsonney2k: btw we have no lisp typemaps! ;)22:10
@sonney2kluckily :D22:10
blackburngsomix: let you implement it ;)22:11
blackburnto get sonney2k feeling not so lucky22:11
gsomixblackburn, L??P?22:11
@sonney2kif we want deep learning in shogun I think I should talk to some big shots and find someone who wants to help us22:11
blackburngsomix: yeah!22:11
@sonney2kLUCP :)22:11
blackburnsonney2k: you got it right!22:12
blackburnsonney2k: try if you have time for that22:12
@sonney2kI cannot write the pppp - i-i22:12
gsomixsonney2k, check mail, please22:13
@sonney2kshogun seriously lacks boosting algortihsm, decision trees and deep learning (even when all projects a done after this gsoc)22:13
blackburnsonney2k: decision trees is painful thing22:14
n4nd0sonney2k: what about SO, isn't that important?22:14
@sonney2kn4nd0, yeah but you said you will do it this gsoc - right ?22:15
@sonney2kblackburn, yeah but people *love* trees22:15
n4nd0sonney2k: :) that's what I wish22:15
blackburnsonney2k: most painful is to design it properly22:15
blackburnI would rather jump out of the window :D22:16
@sonney2kblackburn, anyway what I am saying is that harshit could do a good job - not sure if he wants to do sth else (like trees :D22:16
blackburnsonney2k: lets see it is unclear for me yet22:16
@sonney2kblackburn, yeah for me it is also not clear - lets see how he reacts to my (many!) comments22:17
@sonney2kgsomix, no worries about incomplete patches (as long as they don't break python2 support)22:17
blackburnhe asked to review his code22:17
blackburnmany times actually22:17
n4nd0sonney2k: this afternoon he asked for something to work on so I think it could be ok another topic22:18
gsomixsonney2k, ok22:18
blackburnI just did not want to do it hahah22:18
n4nd0we better suggest him next time in any case22:18
@sonney2kblackburn, I only saw his pull request 2 days ago? anyway I did now more or less22:18
blackburnsonney2k: well you are an svm expert so let you review such things22:19
@sonney2kI have too many comments so I need him to answer first before digging deeper22:19
@sonney2kblackburn, well he said that the matlab code and his impl. of newton svm give the same result22:19
blackburnfor 5 feature vectors22:19
blackburnthat can be a problem..22:19
@sonney2kgood enough for me :)22:20
blackburnthanks google for grouping emails22:20
@sonney2kanyway the C++ code is too memory demanding / slow currently22:20
blackburnI just received a bunch of your comments22:20
@sonney2kn4nd0, ok - I have no idea what timezone he is in - lets see if we find sth we can mentor...22:21
blackburngsomix: check directorkernel, that's the thing I have described you earlier22:21
blackburnsonney2k: I am pretty sure he +4-5 from you22:21
n4nd0n4nd0: he is Indian, I normally speak with him in the mornings ... I quite not remember meeting him at this time22:22
blackburnn4nd0: you always highlight yourself ;)22:23
blackburnhow can it be?22:23
n4nd0fuck!! I have a problem with that22:23
n4nd0I am an egocentric :P22:23
n4nd0sonney2k: he is Indian, I normally speak with him in the mornings ... I quite not remember meeting him at this time22:23
@sonney2kn4nd0, we all are22:23
n4nd0ok, you got it :)22:24
n4nd0sonney2k: you think so? I shouldn't worry about that then22:24
* sonney2k cares most (after shogun code quality) about his hairstyle, taint, muscles :D22:25
blackburnsonney2k: btw are you well again?22:25
@sonney2kI would wish22:25
blackburnbut not?22:25
@sonney2kmust be the code quality I was seeing :D22:25
@sonney2k...and am producing22:26
@sonney2kblackburn, we really need the tests we have in python also in other languages22:26
@sonney2kthis legacy test suite is a pain in the ....22:27
blackburnsonney2k: yeah but it turned to be useful22:27
@sonney2kblackburn, yeah but the example based tests have higher coverage... but are not available in all languages yet22:28
@sonney2konly python_{static,modular}22:28
@sonney2kso much easier to maintain22:28
blackburnsonney2k: I have doubts it has high coverage..22:29
@sonney2kn4nd0, btw not sure if you have seem it - the comments you wrote in your example - I moved them to descriptions (examples have no documentation in the file - these are all separate)22:30
@sonney2kn4nd0, the rationale is: if someone converts your example to e.g. octave_modular then the same description will be prepended22:30
@sonney2kblackburn, 170 tests much higher then we have in the old at least22:31
@sonney2k(for python modular only...)22:31
blackburnsonney2k: oh March, 27 is the day I talked to you for the first time :D22:31
gsomixblackburn, do you mean director classes?22:31
blackburngsomix: yeaahh I hope they would work someday22:32
n4nd0sonney2k: I understand, ok22:32
@sonney2kgsomix, yeah22:32
n4nd0I should do the doc for QDA by the way, I think I still have a TODO there :O22:32
gsomixblackburn, ok22:32
blackburnsonney2k: the problem is they should be updated and cleaned22:33
@sonney2kn4nd0, yes please22:33
@sonney2kblackburn, the problem is we don't run them by default22:33
blackburnin means of quantity it covers more things for sure22:33
@sonney2kif we would we would see any breakage when it appears and they would be kept up to date22:33
@sonney2kand it is sooo easy to update them22:33
@sonney2kjust run <example_name>22:33
blackburnwe will see a lot of breakages next months22:34
@sonney2kand commti22:34
@sonney2kblackburn, do you think so? I am not aware of changes like we had with SGVector etc22:34
@sonney2kfor this gsoc at least22:35
blackburnsonney2k: yeah I am pretty sure, with labels subclassing at least22:35
@sonney2konly additions...22:35
@sonney2kahh yes22:35
@sonney2kand actually SGVector refcounting :)22:35
blackburnfeatures subclassing22:35
@sonney2kfeatures subclassign?22:35
blackburnsonney2k: ah I had one idea before22:36
blackburndo you think it can be useful to add list of strings into the features22:36
blackburnto store feature vector names22:36
blackburnsonney2k: new features will come with latent models22:36
@sonney2kblackburn, it is probably useful but the overhead...22:37
blackburnif it is needed22:37
blackburnsonney2k: what is the overhead?22:37
@sonney2kwe used to have 0 byte overhead22:37
blackburnit can be optional22:37
@sonney2kI mean SGVector used to be ptr + len22:37
blackburnI can't agree with you22:37
blackburnserialization and migrations stuff22:37
blackburnhave a lot of overhead22:37
n4nd0btw what about this guy who asked about reference counting for SGVector?22:37
@sonney2kserailization /migration are not in sgvector/matrix22:38
blackburnsonney2k: but in features22:38
blackburnand other classes22:38
@sonney2kblackburn, yeah but that is just once for whole feature matrix22:38
@sonney2knot per feature / feature vector22:38
@sonney2kn4nd0, I am just browsing the list...22:39
blackburnI can't see extra overhead there22:39
n4nd0sonney2k: cool, sorry I rushed22:39
blackburnif list of names is provided - it stores this list22:39
blackburnif not - no overhead at all22:39
@sonney2kblackburn, if you have an SGVector of bytes and each dim is described there?22:39
blackburnone pointer22:39
@sonney2khuge overhead22:40
blackburnI did not suggest this crazy way of describing features22:40
blackburnjust optional list of names in CFeatures*22:40
blackburn"1.png" "2.png" "3.png" "4.png"22:41
blackburnsonney2k: did you suggest to Evan to implement it?22:43
@sonney2kblackburn, yeah sure no problem with that (even for dims then - but not sure if this is so useful...)22:43
blackburnsee you22:43
blackburnn4nd0: is o-v-o stuff ready?22:45
n4nd0blackburn: I have not tested it yet with your last commit22:46
n4nd0blackburn: but classification is ready, yes22:46
n4nd0blackburn: the set_subset part for kernel machines is still missing though22:47
blackburnah ok I'll implement it22:47
n4nd0as you wish, I will test the classify o-v-o and commit it soon22:48
blackburnn4nd0: you already have contibuted with 1.7k LoC22:51
blackburnand gsomix with 1.9k LoC22:52
n4nd0blackburn: oh, is that good?22:52
n4nd0I use lot of spacing :P22:52
blackburnn4nd0: no idea I am just clicking buttons on github22:52
@sonney2kn4nd0, comments are even better :D22:53
n4nd0sonney2k: yeah, I realized I have not commented the code that much :(22:53
n4nd0I feel bad about that22:53
-!- prm [~prm@] has joined #shogun22:54
blackburnsonney2k: are you back? ;)22:54
@sonney2kn4nd0, well good code doesn't need comments...22:55
blackburnsonney2k: have you seen list of scikits gsoc topics?22:55
@sonney2kblackburn, the situation relaxed22:55
n4nd0sonney2k: I don't dare to say that mine's such a good code22:55
blackburnsonney2k: that's fine ;)22:56
blackburnsonney2k: most interesting thing is they keep 'possible candidate' there22:57
@sonney2kblackburn, I didn't see these before - not a lot though?22:57
@sonney2kwell every highly interested student knows that n4nd0 is possible candidate for SO, gsomix for accessibility, wiking for latent svm22:58
@sonney2kno idea about others22:58
n4nd0blackburn for multitask! :)22:58
@sonney2kahh yes ;)22:59
* sonney2k goes back answering that email22:59
@sonney2kn4nd0, btw regarding students - it also depends on how many slots we get. last year we had 5 but I would not expect we get much more than that if at all...)23:03
blackburnsonney2k: I hope more..23:04
blackburnsonney2k: remember we have heiko23:04
n4nd0sonney2k: yeah ... and lot of people are starting to be around here ... is going to be difficult to get a place there23:04
blackburnI really hope they would give 6-723:05
@sonney2kit is not that difficult ... just contribute a lot before gsoc even starts (potentially even start with `your' project) and you are in.23:06
@sonney2klast year heiko and blackburn where students like this (submitting patches like crazy before gsoc did start)...23:06
@sonney2kthe other 3 submitted convincing patches too - but by far not so many23:07
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds]23:07
n4nd0we'll have to keep up with the hard work then :)23:07
-!- blackburn_ [5bde8018@gateway/web/freenode/ip.] has joined #shogun23:08
@sonney2kn4nd0, blackburn certainly always has ideas :)23:08
n4nd0that is very good actually :)23:09
blackburn_had you been talking about me by the moment I came? ;)23:09
blackburn_yes I have a lot of crazy ideas23:09
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]23:10
blackburn_n4nd0: do you mind constructing a big submarine?23:11
n4nd0blackburn_: eing?!? what do you mean? lol23:11
gsomixblackburn_, yellow submarine :]23:11
blackburn_sonney2k: and yes, SGObject -> CCCP transition is still ongoing23:11
blackburn_gsomix: why not, yellow works well23:12
blackburn_sounds pretty beatly23:12
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun23:14
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun23:23
@sonney2kn4nd0, blackburn_ - can I get your opinion on that reference counting thing and SGVector  etc business?23:23
@sonney2kwe basically need reference counts even for the simple storage container SGVector / SGMatrix etc23:24
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has joined #shogun23:24
@sonney2kwe realized that we are running into trouble without23:24
n4nd0sonney2k: I think that it would be very interesting to have it23:24
@sonney2k<sonney2k> n4nd0, blackburn - can I get your opinion on that reference counting thing and SGVector  etc business?23:24
blackburndamn are you talking only when I get disconnected?23:24
@sonney2kn4nd0, so we have it currently for SGObject23:25
@sonney2kone can do SG_REF / SG_UNREF and all good23:25
-!- blackburn_ [5bde8018@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]23:25
@sonney2know with SGVector we were always passing around copies23:25
@sonney2ke.g. return SGVector<flaot64_t>(xxx)23:25
@sonney2kthat is ptr + associated length23:25
@sonney2know if we add more fields overhead will kill performance23:26
@sonney2kafter all we do this *everywhere*23:26
@sonney2kin learning algorithms etc23:26
n4nd0I understand23:26
@sonney2kso even a inc_ref() is significant overhead23:26
@sonney2kand having a common base class for reference counts will create even more overhead (vtable...)23:27
@sonney2kso the question is what we can do about it (or should)23:27
n4nd0but even if this inc_ref() and the reference counts will increase overhead23:28
n4nd0isn't better than working with the copies?23:28
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]23:29
-!- blackburn [~qdrgsm@] has joined #shogun23:30
blackburnok stable connection finally23:30
blackburnchanged room :D23:30
blackburnsonney2k: n4nd0: still here?23:30
@sonney2kn4nd0, ok so you are +1 for using references?23:30
@sonney2kso we return SGVector<float64_t>& ?23:31
n4nd0I think so23:31
n4nd0from my point of view it is also more intuitive23:31
@sonney2kand then add whatever we want, blackburns dimnames, derivative class etc23:31
blackburnI do not think we should count references of vectors..23:31
n4nd0but that is something more personal I guess23:32
@sonney2kblackburn, we have to23:32
n4nd0blackburn: why?23:32
blackburnwhat is the motivation?23:32
@sonney2kblackburn, it is messy with this do_free business - one never knows if this vector is still in use by someone23:32
blackburnwhere in code it becomes a problem?23:33
@sonney2kblackburn, several algos fetch vectors and we never know who should free it23:33
@sonney2kblackburn, I don't have a concrete example but any place where we don't have a sole owner of the vector23:33
@sonney2k(or matrix / string list)23:33
blackburnit is pretty intrusive change23:34
blackburnwe shall to think about its pros and cons..23:34
@sonney2kso I like Evan Shelhamer's suggestion with the base class23:34
@sonney2kblackburn, exactly - that is why I ask you both23:34
blackburnfor me access/free approach is ok23:34
blackburnbut if there is any example of something not possible with it - yeah we should change it then23:35
@sonney2kfor me no longer - it is too messy23:35
@sonney2kI never know if I can SG_FREE the vector or not (and I wrote that code...)23:35
n4nd0I think that it gets a bit difficult to know that, exactly23:36
n4nd0I also assumed that it was lack of practice though ...23:36
@sonney2kblackburn, for example consider CLabels having an SGVector internally and passing it to the outside23:36
blackburnactually we can avoid reference counting and just use it as object23:36
blackburnand add free stuff to destructor23:36
@sonney2kblackburn, but who frees it then?23:36
@sonney2kblackburn, who calls delete then :D23:37
n4nd0wouldn't that increase the calls to delete around the code?23:37
blackburnSGVector<float64_t> vector();23:38
n4nd0isn't that a limitation?23:38
blackburnwell it won't be deleted if we return it23:38
blackburnor pass a reference to other function23:38
@sonney2kblackburn, except that if vecto is passed to the outside it might get deleted because the original object is gone23:38
n4nd0it may be interesting to destroy the objects before the end of the scope23:39
n4nd0yeah, that is even a bigger limitation I guess23:39
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]23:39
blackburnsonney2k: that shouldn't happen, are you sure?23:39
@sonney2kblackburn, CLabels23:39
@sonney2ksomeone calls get_labels and gets the vector23:39
@sonney2kthen the CLabels object is deleted23:40
blackburnah in that mean23:40
blackburnI forgot that is shared23:40
blackburnok then23:40
@sonney2kquestion is how expensive this is23:41
-!- prm [~prm@] has quit [Quit: Leaving]23:41
@sonney2knot sure if we can do it the SGObject way23:41
blackburnI actually wish we could map memory at least for python23:41
n4nd0I don't really see that point, why the reference counting can increase overhead that much?23:41
blackburnwithout copying it in in/out typemaps23:41
@sonney2kblackburn, different issue23:42
blackburnyes but much more important (for me)23:42
-!- prm [~prm@] has joined #shogun23:42
blackburnI mean SGVector stuff works OK for now :)23:42
@sonney2kn4nd0, at least currently we check for NULL ptrs etc and deriving from some class with potentially virtual functions will create a vtable for that class23:42
blackburnsonney2k: I believe you care about it too much23:43
n4nd0sonney2k: oh I see, I clearly understimated this problem23:43
blackburnpassing vectors is occured only a few times23:44
blackburnthere can't be such huge overhead23:45
@sonney2kblackburn, well what we could do for 0 overhead cases is pass around ptr / len directly (if needed)23:47
-!- prm [~prm@] has quit [Client Quit]23:47
@sonney2kso I am very much in favor of having that ref counting across SGObject/SGVector/SG*23:48
@sonney2kwith some common base class23:48
blackburnI'm ok with it23:48
@sonney2k*but* we need a common way of passing SGVector / SGObject around23:48
@sonney2kSGObject's are passed around as ptr's23:49
@sonney2kso I we should pass SGVectors as pointers too23:49
-!- makokal [~makokal@] has joined #shogun23:49
blackburnthat would mean a lot of work possibly23:49
@sonney2kyeah - the alternative is to have it inconsistent for now and maybe later even change SGObject's to not pass pointers but references?23:50
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun23:51
gsomixsonney2k, blackburn latest build. I continue to work on it.23:52
blackburnsonney2k: it seems we would have to try23:52
@sonney2kgsomix, that is python3 already?23:52
gsomixsonney2k, yep. there are some small problems.23:53
@sonney2kblackburn, so with references first?23:53
blackburngsomix: please install pickle and scipy23:53
blackburnsonney2k: I like references mostly23:53
gsomixblackburn, ok23:53
blackburnbut it is obvious that any experiments should be done in a separate branch23:54
@sonney2kgsomix, line 178 - python3 converts int's to float when x/2 returns a float23:55
@sonney2kso you just need to cast to int23:55
@sonney2kblackburn, yeah one could do a+b for SGVector objects23:55
blackburnsonney2k: have you ever seen accurate derivation of crammer-singer formulation?23:55
@sonney2kso that is nice23:56
blackburndual task23:56
blackburndual task of C-S23:56
@sonney2kblackburn, derivation of dual starting from primal?23:56
@sonney2kno - but I guess it is easy23:56
gsomixsonney2k, thx23:56
gsomixsonney2k, also I know how to fix the problem "TypeError" in kernel_weighted_degree_position_string_modular.py23:57
blackburnsonney2k: have time to help? ;)23:57
@sonney2kgsomix, you don't have lzo installed23:57
@sonney2kblackburn, haha23:58
blackburnsonney2k: I am for real actually :)23:59
--- Log closed Tue Mar 27 00:00:19 2012