Open in new window / Try shogun cloud
--- Log opened Tue Nov 27 00:00:15 2012
-!- zxtx [] has quit [Ping timeout: 256 seconds]00:19
-!- zxtx [] has joined #shogun00:24
-!- gsomix [~gsomix@] has quit [Ping timeout: 276 seconds]00:34
-!- ptizoom [] has quit [Quit: Ex-Chat]03:47
-!- ptizoom [] has joined #shogun03:47
-!- gsomix [~gsomix@] has joined #shogun04:55
-!- gsomix [~gsomix@] has quit [Ping timeout: 248 seconds]05:16
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]08:18
-!- gsomix [~gsomix@] has joined #shogun10:07
-!- gsomix [~gsomix@] has quit [Ping timeout: 255 seconds]10:53
-!- gsomix [~gsomix@] has joined #shogun12:31
-!- K0stIa [~kostia@2001:718:2:1634:4a5b:39ff:fe99:cc49] has joined #shogun13:36
K0stIahello, everybody!13:40
K0stIathe example stopped working. It does segmentation fault if you run the script for several times without exiting from python environment. The same problem appears if I am trying to extend DirectorStructuredModel, it does also segmentation fault if you run the same script for several times.13:40
-!- K0stIa [~kostia@2001:718:2:1634:4a5b:39ff:fe99:cc49] has left #shogun []14:01
-!- K0stIa [~kostia@2001:718:2:1634:4a5b:39ff:fe99:cc49] has joined #shogun14:02
wikingK0stIa: mmm14:25
wikingK0stIa: which version of shogun r u  usinf14:25
K0stIalast version from github14:47
wikingK0stIa: can u share the script?14:48
K0stIayou made make file ?14:49
K0stIasorry :)14:49
K0stIayou want me to show make file ?14:50
wikingi'd like to test it but i need an example to test this problem with14:50
K0stIaI see :)14:50
wikingso if u can share me your script that does segfault14:50
K0stIayes :)14:50
K0stIawait a minute please14:50
wikingthat'd help me to debug the problem14:50
K0stIaI run this file several times in ipython14:51
K0stIaafter third it throws exeption14:51
wikingok i'll try that14:52
wikingbut why r u running this several times? :)14:52
K0stIawell, the same problem appear if you want to run it for different lambda14:52
K0stIato make cross-validation14:52
wikingok i'll try it now14:53
K0stIathis is just example which simulates my problem14:53
K0stIabtw: did you implement the director class for structure model ?14:53
K0stIado you know who did it ?14:53
wikingbut i've rewritten lately couple of parts of bmrm14:54
K0stIaI need now bmrm :)14:54
K0stIaand it was ok before...14:54
wikingwell maybe i've introduced a bug :(14:54
K0stIaso maybe the problem is there ?14:54
wikingi'll check it for you14:54
K0stIathanks a lot :)14:54
wikingi've done tests with it14:54
K0stIaok :)14:54
wikingbut i haven't done it with neither python14:54
wikingnor trying to run it several times14:55
wikingalthough i run it several times14:55
wikingin a different setup14:55
wikingand i haven't noticed any problems14:55
K0stIaplease, check this script :)14:55
wikingyeah i'll do14:55
wikingi use bmrm via c++14:55
wikingbut that shouldn't matter in this case anyways14:56
K0stIawell, maybe it could14:56
wikingsoon we'll see.. i'm just recompiling shogun with python support14:56
K0stIaplease, recompile it also with director classes support14:57
wikingu just start ipython and start fomr ther the script?14:59
K0stIai started ipythin, then did %run several times15:00
K0stIaand it did segmentation fault15:00
K0stIathe same things appears in my working script15:00
K0stIawhich I can not share :(15:00
wikingsecret? :)15:01
K0stIakind of :P15:01
wikinghow come? :)15:01
K0stIano :) it will take a time to make it independent :)15:02
K0stIaand I don't have it know15:02
K0stIaI detected the problem15:02
K0stIayep :)15:02
wikingdont understand :)15:02
K0stIawell, i would have to share some files in addition for you :) and that I don't want :)15:03
wikingi see15:04
wikingwhat context are u using this for?15:04
wikingi mean shogun15:04
wikingor that's unshareable information15:04
K0stIaI use it in my work on paper now :)15:05
K0stIafor learning :)15:05
wikingah so u r using it for research15:05
K0stIaI also have some suggestion for some improvements for structure output learning15:07
wikingwe are listening15:07
wikingthe reason i'm running it several times is because i use these algos to do latent structural classification...15:08
wikingmmm i'm getting this error: SystemError: [ERROR] FATAL: Parameter::add_type(): Double parameter `m_TolRel'!15:08
wikingi'm having some patches already prepared to for a bit better API of SO learning15:08
K0stIait would be nice to provide in director class possibility to override the the risk funtion15:08
wikingwell the risk function is soon to be changed15:09
wikingas i've just commited a non-convex risk minimizer15:09
wikingso that u don't have to do funky things if u want to minimize an concave risk function15:09
K0stIawill it be possible to override the risk function ?15:09
wikingwell yeah15:10
K0stIathat's nice15:10
wikingwhat's your risk function now?15:10
wikingor that's secret as well?15:10
K0stIano :)15:10
wikingok i have totally different problem now: SystemError: [ERROR] FATAL: Parameter::add_type(): Double parameter `m_TolRel'!15:10
K0stIait's risk approximation for supervised learning15:11
K0stIawhich is convex15:11
wikingi see it's cool15:11
wikinghope you'll share it15:11
wikingwhen u have the paper done15:11
K0stIabtw: which algorithm is implemented in your non-convex ?15:11
wikingit's simply solving15:11
wiking1/2 \lambda ||w||^2 + R(w)15:12
wikingwhere R(w) could be anything15:12
wikingit's based on libqp15:12
K0stIait's known approximation... you can find it in SO related papers15:12
K0stIais there some paper about it ?15:12
wikingnot yet15:12
wikingi'm preparing one15:12
K0stIaI see :)15:12
K0stIawhat is the class of functions that algorithm works for ?15:13
wikingit's in structure15:14
K0stIaI mean for with type of risk functions it could be applied ?15:14
wikingno python api for it15:14
wikingany risk function15:14
wikingcurrently it's bundled to structured model15:15
wikingso you would need to give a structured model with your desired risk function15:15
K0stIareally ? for any nonconvex function it will find a global optima ?15:15
wikingworked for me15:15
wikingat least for 2 problems15:16
wikingthat are non-convex15:16
K0stIacan you say what they problems are ?15:16
wikingwell i'm applying this in computer vision15:16
wikingso the problems are from that field15:17
wikingok i think ishould switch to master branch15:18
K0stIado you have any proof about guarantees ?15:18
wikingas i already have some other patches in my local branch and i get totally different15:18
wikingK0stIa: you mean about the convergence15:18
K0stIait would be interesting to see you r paper15:21
wikingwell it's not published15:21
wikingit's not even submitted15:21
K0stIaand now can you do something with that bugs ?15:21
wikingwell i'm trying to reproduce it first15:21
wikingbut i'm getting very different errors15:22
wikingbut i'm guessing it's because of the local patches i've done15:22
K0stIawell, could you try to install version from github  ?15:24
K0stIawithout patches ?15:24
K0stIamaybe there are some another bugs ?15:24
wikingyeah that's what i'm trying15:25
K0stIagood :)15:25
wikingK0stIa: well i'm running now for the 5th time15:42
wikingno segfault yet15:42
K0stIareally ?15:43
K0stIayou tried the same example ?15:43
K0stIaseveral times ?15:44
wiking%run so_multiclass_BMRM.py15:46
K0stIashould be %run so_multiclass_director_BMRM.py15:47
wikingthis one segfaulted15:48
wikingi'll gdb it15:48
K0stIawiking: how it goes ?16:43
@sonney2kwiking, is there any reason for this ::vec1_plus_scalar_times_vec2 function in SGVector17:16
K0stIasonney2k: hi, who is responsible for doing director class for structured model ?17:20
K0stIacould he change oracle for structured model in the following way, the oracle should be represented as an interface  and user should provide function which takes point and returns function value and subgradient ?17:20
@sonney2kK0stIa, n4nd0 did this.17:21
@sonney2kK0stIa, and gsomix is the master who made it possible...17:22
K0stIasonney2k: now it is a bit confusing... one has to pass some loss function to the svm, however loss is not needed, which is not that crucial for me, but I guess it would be better to get rid of it.17:23
@sonney2kK0stIa, well ask on the mailinglist if he doesn't appear here before you leave.17:27
K0stIasonney2k: for structured learning now it is not that convenient as it could be,17:29
K0stIa1. the summation in case of learning more than one example is done inside shogun, I think it would be better if user do this summation, because this is something with is hard to compute and could be easily paralyzed. One can do it now of course, however is not convenient as it could be. that's why I think this director class should be as follows, just an interface with risk function that takes point from R^n space and returns function va17:29
-!- blackburn [~blackburn@] has joined #shogun17:41
gsomixsonney2k, K0stIa hey17:55
K0stIagsomix: hey17:58
-!- K0stIa [~kostia@2001:718:2:1634:4a5b:39ff:fe99:cc49] has quit [Remote host closed the connection]18:43
-!- K0stIa [] has joined #shogun19:16
wikingsonney2k: well just that because there's a cblas optimized code for that19:25
wikingK0stIa: u know that u can do it and submit a PR, right? :19:38
wikingK0stIa: pr = pull request19:39
wikingsonney2k: i think ccsosvm broke... aint' working anymore for me :(19:39
K0stIawiking:  I haven't been working on it. It could take time for me which I don't have now.19:40
wikingK0stIa: well yeah that's a usual problem for everybody :(19:40
wikingtime i mean19:40
blackburnwiking: I was reading logs and don't you find strange you can find global maxima of non-convex function?19:40
wikingblackburn: why? :)19:41
wikingand not global maxima19:42
wikingbut minima19:42
blackburnit is the same19:42
wikingyeah indeed19:42
wikingnow i have to cherry pick this stuff :(19:47
wikingbisect is my friend19:55
wikingK0stIa: maybe you could use bisect to find out which commit introduced this bug if it used to work before...19:56
wikingwould be great if we would have unittests for this :)20:00
wikingcould run bisect automatically20:00
-!- gsomix [~gsomix@] has quit [Ping timeout: 246 seconds]20:28
wikingsonney2k: that patch of yours for std::vector is not working20:29
wikingso there's something wrong with DynArray afaik20:29
naywhayaresonney2k: curious question; was there a particular reason you settled on buildbot as a CI tool?20:49
naywhayareI don't necessarily have a better suggestion or anything, just, curious, is all20:50
-shogungit:#shogun- [shogun] lisitsyn pushed 1 new commit to master:
-shogungit:#shogun- shogun/master 315b058 Sergey Lisitsyn: Flexible director risk21:01
@sonney2knaywhayare, no I just tried it out first and it was python so I could easily extend it to my needs.21:04
@sonney2kwiking, what happens?21:05
@sonney2kwiking, re SGVector I meant why not do template specializations instead of overloading? I am currently changing the code to do exactly that21:05
@sonney2kwiking, I am a bit fed up with the get_uint_matrix etc functions and try to do overloading for that stuff now too...21:06
wikingsonney2k: dunno what happens currently... i'm just saying that std::vector -> DynArray introduced a bug that basically the whole CCSOSVM doesn't work21:06
@sonney2kwiking, well if you had I test or at least an example then I could check21:06
@sonney2kthis way you have to figure out what is wrong21:07
naywhayareok.  thanks for the answer :)21:07
wikingsonney2k: do u have mosek?21:08
@sonney2knaywhayare, he - well and all issues I had were resolved in #buildbot - IIRC tomprince helped a lot and basically immediately21:10
@sonney2kor at the GSoC mentors summit FWIW21:10
naywhayareah, that's convenient21:10
naywhayareI use Jenkins and have had decent success with it, but it's one of those things where I haven't had the time to set it up properly21:11
naywhayareI'll keep buildbot in consideration for the future when I have infinite free time21:11
naywhayarewhich unfortunately seems infinitely far away...21:11
wikingsonney2k: i've just changed back to std::vector and it works perfectly... :S21:14
wikingso i dont know what happens with DynArray actually21:14
wikingas i really just changed 3 things. definition of the variable21:15
wikingand resize_array -> resize21:15
wikingso most probably there's a bug in DynArray :(21:16
@sonney2knaywhayare, I also have good experience with jenkins21:16
@sonney2kwiking, what does valgrind say?21:17
wikingsonney2k: i cannot valgrind it, as valgrind is broken currently on my os :(21:17
@sonney2kwiking, this makes us a perfect team :-/21:18
shogun-buildbotbuild #443 of cyg1 - libshogun is complete: Failure [failed compile]  Build details are at  blamelist: Sergey Lisitsyn <>21:19
@sonney2kwiking, the only idea I have is that we add some small example in examples/undocumented/library_dynarray.cpp that uses SGVectors...21:20
@sonney2kwhich I guess might be the issue21:20
wikingok i've pushed a pr21:22
-shogungit:#shogun- [shogun] vigsterkr opened pull request #839: CCSOSVM refactoring (
wikingit's a code refactor anyways of CCSOSVM21:22
wikingand currently switches back to std::vector21:22
wikingso don't merge it until we don't find out the dynarray stuff21:22
wikingand then i'll do a little commit for this PR21:23
wikingthat switches back to dynarray21:23
@sonney2kI am currently deep in refactoring hell21:23
wikingthat's ok21:23
wikingthis can sit there for some time21:23
@sonney2khah it looks like it compiles21:24
@sonney2kwiking, so now we can just use SGVector.load21:25
wikingwhat happened to travis?21:25
@sonney2kSGMatrix.load etc21:25
@sonney2kand no more need for tons of defines21:25
shogun-buildbotbuild #680 of deb3 - modular_interfaces is complete: Failure [failed compile csharp_modular]  Build details are at  blamelist: Sergey Lisitsyn <>21:29
-shogungit:#shogun- [shogun] sonney2k pushed 2 new commits to master:
-shogungit:#shogun- shogun/master b4322ff Soeren Sonnenburg: Refactoring...21:32
-shogungit:#shogun- shogun/master b7a6786 Soeren Sonnenburg: compile fix for csharp (directors & SO)21:32
shogun-buildbotbuild #444 of cyg1 - libshogun is complete: Failure [failed compile]  Build details are at  blamelist: Soeren Sonnenburg <>21:46
naywhayaresonney2k: does buildbot do something like jenkins' "configuration matrix"?21:56
naywhayarewhere you can build your project against a combination of different compilers, libraries, and so forth21:57
wikingnaywhayare: no21:57
shogun-buildbotbuild #445 of cyg1 - libshogun is complete: Failure [failed compile]  Build details are at  blamelist: Soeren Sonnenburg <>21:57
naywhayareI couldn't seem to find anything about that in the documentation I was perusing21:57
-!- heiko [~heiko@] has joined #shogun21:59
shogun-buildbotbuild #681 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Soeren Sonnenburg <>22:12
shogun-buildbotbuild #682 of deb3 - modular_interfaces is complete: Failure [failed compile csharp_modular]  Build details are at  blamelist: Soeren Sonnenburg <>22:12
@sonney2kblackburn, you broke it all :)22:26
blackburnsonney2k: can't really understand the error22:26
@sonney2kblackburn, I am fixing the csharp one22:26
@sonney2kbut the other?22:26
blackburnother like?22:26
@sonney2kother error22:27
blackburnwhich one/22:27
blackburnit makes no sense to me22:27
blackburnat all22:27
@sonney2kat least it has to do with your commit :)22:28
blackburn12 people in room22:29
blackburnheiko: !22:29
@sonney2kugh csharp compiles here22:29
heikoblackburn, hi :)22:29
blackburnheiko: long time no see as sonney2k likes to say22:30
heikoblackburn, yes he would be right22:30
heikobusy weeks here22:30
heikohow are you?22:30
@sonney2kheiko, trying to rescue the world from evil?22:30
heikosonney2k, I did, but usually nobody notices22:31
heikoeverything going alright here?22:31
@sonney2kthats the destiny of super heros22:31
heikohehe ;)22:31
heikosonney2k, blackburn, no just first few weeks of new job, and its all a bit different than expected so I had lots of things to do22:32
heikocurrently in Cambridge22:32
heikonext week in Warsaw22:32
@sonney2kheiko, we are tying hard to break all code - but it is not as easy ;-)22:32
heikoproject ends with this year thank god22:32
heikosonney2k, sounds good ;)22:32
blackburnyeah I use axe22:32
blackburnto break everything22:32
blackburnbut it still works sometimes22:32
@sonney2kheiko, but back in germany next year then?22:32
@sonney2kblackburn, csharp just compiles here!?22:33
heikosonney2k, no, I will stay in London22:33
@sonney2koh different job?22:33
@sonney2knot the GP classification one?22:33
heikosonney2k, well currently I have to finish a EU project with a company where GPs are used, but since so little time left, I just need to collect stuff - so no time for shogun implementation22:34
heikobut next year I will continue working in the stats department, but doing research22:34
blackburnI start on new job next week starting from monday22:34
-shogungit:#shogun- [shogun] sonney2k pushed 1 new commit to master:
-shogungit:#shogun- shogun/master 1d57638 Soeren Sonnenburg: include lib common to make data types know22:34
heikoand then, the cool MCMC stuff with GPs will happen - hopefully22:34
@sonney2kheiko, sounds more like you won't have time either :/22:35
heikono I think it will be better since no grant project deadlines22:35
blackburnsonney2k: have you ever thought it will be better some day? ;)22:35
heikothey basically did nothing during the summer and now I am supposed to do all the stuff22:35
@sonney2kheiko, poor slave22:36
@sonney2kso rescuing the world indeed!22:36
heikosonney2k, yeah a bit annoying, but at least I can do traveling and get a salary - and its interesting stuff in any way, so learning...22:37
blackburnsounds like my thought when I got into netcracker :D22:37
@sonney2kheiko, I hope you will find the time to add GPC to shogun - now that you've done this stuff once you should probably know best how to do it right :)22:38
wikingwhere should be the structure machines be defined?22:38
wikingsonney2k: ^22:38
@sonney2kwiking, what?22:38
heikosonney2k, yes I will do that for sure, implemented a lot of stuff in matlab recently22:38
wikingsonney2k: currently structured machines doesn't have a type22:38
heikomulti-class GPs, approximate inference, approximate exact inference, cool stuff22:38
wikingsonney2k: EMachineType22:38
heikobut currently only hacks,22:38
@sonney2kwiking, shogun/machines/StructuredOutputMachine.h ?22:39
@sonney2kalso there22:39
wikingsonney2k: they just have problem type22:39
wikingsonney2k: but i cannot differentiate the different SO machine implementations22:39
wikingonly if i try try catch dynamic csat22:39
wikingsonney2k: or of course get_name + strcmp ;)22:40
heikocurrently seeing christians pydata talk22:41
@sonney2kwiking, blackburn -
@sonney2kthere are still a few warnings22:45
wikingsonney2k: so?22:46
wikingsonney2k: (this is not a so about the warning but the machines..)22:46
@sonney2kwiking, in your code.22:47
wikingsonney2k:  i know that i see that :)22:47
@sonney2kwiking, feel free to just add to EMachineType in shogun/machine/Machine.h22:48
shogun-buildbotbuild #446 of cyg1 - libshogun is complete: Failure [failed compile]  Build details are at  blamelist: Soeren Sonnenburg <>22:50
wikingsonney2k: there u go22:51
-shogungit:#shogun- [shogun] vigsterkr opened pull request #840: Fix warnings in libncbm and StructuredModel (master...master)
-shogungit:#shogun- [shogun] sonney2k pushed 2 new commits to master:
-shogungit:#shogun- shogun/master 46a1f28 Viktor Gal: Fix warnings in libncbm and StructuredModel22:52
-shogungit:#shogun- shogun/master 11acf90 Soeren Sonnenburg: Merge pull request #840 from vigsterkr/master...22:52
-shogungit:#shogun- [shogun] vigsterkr opened pull request #841: Define EMachineType for the SO learners (master...master)
* sonney2k ZZzzz...23:03
shogun-buildbotbuild #447 of cyg1 - libshogun is complete: Failure [failed compile]  Build details are at  blamelist: Viktor Gal <>23:09
shogun-buildbotbuild #683 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Soeren Sonnenburg <>23:16
shogun-buildbotbuild #448 of cyg1 - libshogun is complete: Failure [failed compile]  Build details are at  blamelist: Soeren Sonnenburg <>23:20
shogun-buildbotbuild #684 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Soeren Sonnenburg <>, Viktor Gal <>23:46
--- Log closed Wed Nov 28 00:00:15 2012