Open in new window / Try shogun cloud
--- Log opened Thu Aug 23 00:00:17 2012
shogun-buildbot_build #63 of nightly_none is complete: Success [build successful]  Build details are at
shogun-buildbot_build #74 of nightly_default is complete: Failure [failed test]  Build details are at
-!- emrecelikten1 [] has joined #shogun04:26
-!- emrecelikten [] has quit [Ping timeout: 248 seconds]04:27
-!- av3ngr [av3ngr@nat/redhat/x-kzxmpmjwtrcqadom] has joined #shogun05:56
-!- av3ngr [av3ngr@nat/redhat/x-kzxmpmjwtrcqadom] has quit [Client Quit]05:56
-!- cheng [80fa500f@gateway/web/freenode/ip.] has joined #shogun08:16
-!- uricamic [~uricamic@2001:718:2:1634:6dc1:2e18:bbfe:a408] has joined #shogun08:35
-!- blackgun|shoburn [5bdfb203@gateway/web/freenode/ip.] has joined #shogun08:37
blackgun|shoburnhey cheng08:37
CIA-52shogun: Michal Uricar master * r3713c0c / examples/undocumented/libshogun/so_multiclass_BMRM.cpp : BMRM libshogun example fix (runnable without cmd arguments) -
CIA-52shogun: Michal Uricar master * rf1a5b95 / examples/undocumented/python_modular/graphical/ : fix -
CIA-52shogun: Sergey Lisitsyn master * rdea19d8 / (2 files in 2 dirs): Merge pull request #754 from uricamic/BM_SOL_EXAMPLE -
chengblackburn: Hi08:43
chengblackgun|shoburn: Hi again. :-)08:43
blackgun|shoburnokay let get not so crazy nickname08:44
-!- blackgun|shoburn is now known as blackburn08:44
chengblackburn: You want to start a new channel so that we don't flood shogun?08:45
blackburnwe could just send messages directly08:45
blackburncheck other tab08:45
shogun-buildbot_build #337 of bsd1 - libshogun is complete: Success [build successful]  Build details are at
shogun-buildbot_build #448 of deb3 - modular_interfaces is complete: Success [build successful]  Build details are at
shogun-buildbot_build #355 of deb2 - static_interfaces is complete: Success [build successful]  Build details are at
-!- n4nd0 [] has joined #shogun10:22
-!- emrecelikten1 [] has quit [Quit: Leaving.]11:15
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has quit [Quit: Page closed]11:16
-!- cheng [80fa500f@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]11:23
wikingn4nd0: ping12:06
n4nd0wiking: tell me12:14
wikingwhen u have time i want to discuss a patch with you12:14
n4nd0I have time now12:15
wikingit's a massively ugly patch12:15
wikingbut i really dont know what to do about this12:15
wikingnono mine12:15
wikingmmm i think i'll just pastebin it12:15
wikingn4nd0: header
wikingthis is the consequence of not being able to have class anything: public CStructuredModel, public CLatentModel12:17
n4nd0yeah, I see12:19
n4nd0but it could be bad to have multiple inheritance ...12:19
n4nd0why are there two classes, CLSOModel and CLatentSOModel?12:19
wikingwell the naming is now stupid12:20
n4nd0aham I see12:20
wikingbut basically CLSOModel goes for the structural machine12:20
n4nd0one extends LatenModel and the other one the StructuredModel12:20
wikingand in order to just12:21
wikingrequire one inheritance12:21
wikingi hardwired CLatentSOModel into CLSOModel12:21
wikingi.e. a user will only need to do an inharitance of CLatentSOModel12:21
wikingand define all the relevant application specific functions12:21
n4nd0I understand12:22
n4nd0I guess that, providing how it is one right now, a user should do inheritance of CLSOModel or?12:22
n4nd0I mean it since it is the class that has access to CLatentSOModel and not vice versa12:23
wikingit should do inheritance of CLatentSOModel12:23
wikingthe CLSOModel is completely fixed12:23
wikingand basically it's 'just' a proxy class12:23
wikingand we cannot do multiple inheritance as then some modular interfaces will fail12:25
wikinglike java12:25
wikingso we either go with something like this12:25
wikingor i just throw out latentmodel as is12:25
wikingand try to use structuredmodel12:25
wikingthe problem is that for simple LatentSVM12:25
wikingmost of the stuff in structuredmodel is not necessary12:26
wikingand some other stuff are missing...12:26
wikingso if possible i would like to keep it12:26
wikingand maybe just rename CLSOModel to CLatentSOProxy12:26
n4nd0that last renaming is good I think12:27
n4nd0but I don't think it is such an ugly patch!12:28
n4nd0I mean it is not a common thing to do12:28
n4nd0but it is not a common situation either12:28
n4nd0I think it is ok12:28
wikingbtw: this is lol:
wiking500 Terabytes of Data /day12:28
wikingand it's all running on hadoop12:29
wikingpretty impressive that it can handle it12:29
wikinghahaha the hdfs is 100+ PB12:30
n4nd0smart people12:30
wikingyeah maybe shogun would fail on trying to use this :)12:30
wikingStreamingFeatures ;)12:30
wikingok well if u don't find this patch soooo ugly12:32
wikingthen i can actually commit12:32
wikingjust need to finish calculating the subgrad in the risk function12:32
wikingdo u actually use in PrimalMosek the y value in the ResultSet?12:35
wikingor just psi_truth and psi_pred12:36
n4nd0wiking: you mean the argmax right?12:36
wikingyeah sorry12:37
wikingyeah ResultSet->argmax12:37
n4nd0I think not12:37
n4nd0let me check12:37
wikingyeah i've greped now12:37
wikingdidn't show up12:37
wikingjust wanted to make sure12:37
n4nd0just delta and psis12:37
n4nd0that one is used in the moment of prediction12:38
n4nd0in the PrimalMosek we are training so ignored12:38
wikingthnx a lot12:38
n4nd0you are welcome12:39
n4nd0this is just awesome12:42
n4nd0I don't know if you understand12:42
n4nd0but just look at the picture...12:42
n4nd0restoration, they said12:44
wikingyeah i saw that... the old lady went a little bit creative ;)13:07
wikinguricamic: here?13:16
-!- n4nd0 [] has quit [Ping timeout: 276 seconds]13:16
uricamicwiking: yep13:16
wikingso i'm just checking the multiclassmodel again, the risk function13:16
wikingand the subgrad is  R'({\bf w}) = \sum_{i=1}^{m} \Psi(x_i, \hat{y}_i) - \Psi(x_i, y_i),13:16
wikingi wanted to say that couldn't we just use the CResultSet->psi_truth and CResultSet->psi_pred13:17
wikingbut then i've realised that argmax is not being called in the ::risk function13:18
uricamicnot yet, but eventually it should be called there13:18
wikingok ok13:18
wikingthen i guess we could use those two values13:18
wikingas they are generated already by argmax13:18
uricamicbut the psi_pred has to be computed including the loss13:18
uricamicI mean argmax, computed on features infected by loss13:19
wikingok i think this one is ready let's test it on the example :)13:32
-!- cheng [73406f11@gateway/web/freenode/ip.] has joined #shogun13:57
-!- n4nd0 [] has joined #shogun14:03
-!- cheng [73406f11@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]14:22
n4nd0wiking: ?14:33
wikingscreaming at my code :)14:33
wikingwhy no work :)14:33
n4nd0does it work?14:33
n4nd0I mean when you scream :D14:33
wikingheheh yeah14:35
@sonney2kwiking, please do your final evaluations today!15:04
wikinghahah i looove when somebody is creating a release with -g flags15:04
wikingyou always get some nice info:15:04
@sonney2kwhy would one not?15:04
wikingsonney2k: in a release?15:04
@sonney2kahh closed source ok...15:04
wikingsonney2k: even if it's open source15:05
wikingi wouldn't create a release with -g15:05
@sonney2kthen you have no info when things crash15:05
@sonney2kwiking, seen the 'please do the final eval today'?15:05
wikingsonney2k: affirmitive15:05
wikingsonney2k: *nod* and on it!15:05
wikingsonney2k: i mean honestly i would just do a striped library for real release15:06
wikingthat way small memory footprint15:06
wikingbut it's good now as my code crashed big time15:06
wikingah null pointer referencing :)15:07
wikingn4nd0: still here?15:07
n4nd0wiking: yeah15:07
n4nd0uricamic: are you around?15:07
wikinghave u checked the static analyzer report for structure/*15:07
uricamicn4nd0: yep15:07
n4nd0wiking: not really to be honest15:08
wikingdo u use DynProg atm somewhere?15:08
wikingi mean it's jsut that it'd be great to clear out some of those problems15:08
n4nd0uricamic: have you analyzed the Dual Lib QP BM SOSVM for memory leaks? I get some in my example when I am using it :S15:09
@sonney2kwiking, how do you create the static analysis btw?15:09
wikingi'm happy to create a new report with the latest commit15:09
@sonney2kwiking, I mean we could run this on the bb's regularly...15:09
n4nd0wiking: I am using PLiFs but those are defined in Plif.h and so not in DynProg15:09
wikingsonney2k: yeah but we need clang/llvm for it15:09
wikingsonney2k: at least for this particular one15:09
n4nd0wiking: so no, I am not using anything from DynProg directly15:09
uricamicn4nd0: to be honest not15:09
uricamicI mean not using e.g. valgrind15:09
wikingsonney2k: clang-analyzer.llvm.org15:09
@sonney2kwiking, that shouldn't be the problem I think15:09
uricamicbut according to some report system it should be ok15:10
n4nd0uricamic: aham ok, I think there may be some leaks there - at least it is what valgrind tells me15:10
wikingsonney2k: i'm happy to do it if u give me access where i can have clang15:10
uricamicn4nd0: ok, can u send me the log then?15:10
n4nd0uricamic: if you are busy with another thing I can try to fix it if you don't mind15:10
uricamicn4nd0: ok, feel free to do that :)15:11
n4nd0uricamic: I will let you know about any change15:11
wikingsonney2k: i'm not 100% sure that u really need to build the whole thing from scratch15:11
uricamicn4nd0: ok15:11
wikingsonney2k: but yeah it'd be cool to have that report within the buildbot system15:12
wikingsonney2k: easy peasy ;)15:12
wikingsonney2k: apt-get install clang15:13
wikingsonney2k: after that you just need scan-build ./configure && scan-build make15:14
wikingand your report is ready15:14
@sonney2kwiking, will that do the usual compile too?15:14
@sonney2kI mean is the analysis in addition?15:15
wikingsonney2k: good question15:19
wikingsonney2k: it did compile me the libshogun15:20
n4nd0wiking: have you used resize_vector in SGVector?15:22
n4nd0wiking: valgrind detects leaks with it but I don't see anything weird in the code15:22
wikingn4nd0: noup i haven't15:24
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has joined #shogun15:33
blackburnworst day on job ever :D15:37
n4nd0blackburn: why so?15:40
blackburnsupport is on fire with 3-4 blocker issue15:40
blackburnmay be even 515:40
n4nd0what is that?15:42
blackburnissue - well just like github issue15:42
blackburnblocker means we have to respond immediately or close to that15:42
n4nd0it was the blocker what confused me hehe15:44
n4nd0but that means some job to do, right?15:44
@sonney2kblackburn, so down to 0 blockers now?15:45
@sonney2kblackburn, btw any ideas where gsomix is hiding?15:45
blackburnn4nd0: blocker is like  OMG DAMN WE ARE GOING TO DIE issue15:46
blackburnsonney2k: no I am still wroking on it15:46
n4nd0blackburn: hehe15:46
n4nd0blackburn: I have a blocker issue here now15:46
@sonney2kblackburn, I at least know what you are talking about15:46
blackburnsonney2k: hehe15:46
n4nd0blackburn: valgrind complains quite a bit of SGVector::resize_vector, any clue about it?15:46
n4nd0have you used this method before?15:46
blackburnn4nd0: no, not really15:47
blackburnlet me get back to that a little bit later15:47
n4nd0I'll investigate a bit more then15:47
@sonney2kn4nd0, it is like hey we sold you an alarm clock but it will stop incrementing seconds after 7 days of use15:47
n4nd0sonney2k: :D I like your metaphors15:48
n4nd0but last day's one with the party to go by car or boat and the bundle methods for SO is remains my favourite :D15:49
n4nd0uricamic: let me ask you, why do you use resize_vector in libbmrm.cpp lines 528-53015:51
blackburnsonney2k: I hope you didn't have to use clearcase15:51
uricamicn4nd0: it is there just because of the history tracks of Fp, Fd and wdists15:52
n4nd0uricamic: it looks to me that at that point there is nothing else to do in the algorithm, then why to resize those vectors?15:52
blackburnand we, wiking and I, are the only victims of it15:52
uricamicand resize is there since I guess in the beginning there is initialization to the BufSize15:52
uricamicso to remove the rest after nIter15:53
n4nd0uricamic: aham, I understand15:53
uricamicbut probably this could be done in more clever way now15:53
uricamicI was writing it in a hurry and also changed the way how these values are represented, so it is possible that the recent version is cumbersome15:54
n4nd0uricamic: no problem :)15:55
uricamicso the lines valgrind was complaining about were just these with resize_vector ?15:56
n4nd0uricamic: yes, it seems something is not working as it should with resize_vector15:56
uricamicn4nd0: I see, well positive is at least that the rest is ok :)15:57
n4nd0uricamic: there was another thing related to using get_label in MulticlassModel::risk15:57
n4nd0uricamic: but that was easy to solve15:57
uricamicn4nd0: I see, ok15:58
n4nd0uricamic: when you call get_label you need to do an UNREF later since this method increases the reference count of the object that returns15:58
n4nd0uricamic: I will show you when I prepare the patch15:58
uricamicn4nd0: aha, ok, thx15:58
@sonney2kn4nd0, btw is your friend bern4rd still alive? I hope I can install the new website rather soon that is why I am asking (in case we run into issues...)16:00
n4nd0sonney2k: yes he is alive, he's in an internship in Paris16:01
n4nd0sonney2k: I can tell him about it16:01
@sonney2kCool thanks16:02
n4nd0what day will that be?16:04
@sonney2kn4nd0, well we put it live September eleven errr 1 :)16:05
@sonney2kif blackburn doesn't have blockers to fix :D16:05
@sonney2kbut I want to put it live under some 'hidden' url fist16:05
blackburnnext 3 days should be free16:05
n4nd0sonney2k: ok16:05
blackburnthat's why I am finishing things with my burning head16:05
@sonney2kbunsenblackburner even!16:06
uricamicn4nd0: THANKS16:21
n4nd0that has solved the leaks I saw using so_multiclass.cpp16:21
n4nd0it seems that there are some others triggered when using so_multiclass_BMRM.cpp16:21
uricamicok, I will have to update valgrind and look on that16:22
n4nd0uricamic: I posted in my blog a good reference for valgrind if you need to check something16:23
-!- yoo [575b08cb@gateway/web/freenode/ip.] has joined #shogun16:23
yoohi all16:23
n4nd0hi yoo16:24
uricamicn4nd0: the only problem I encountered was that the last version I had haven't supported some instructions and therefore it actually didn't provide any useful information, the way how that could be overcome was to compile with some flag to disable those features16:25
n4nd0uricamic: aham ok16:25
uricamicbut somebody told me, that there should be now new version for ubuntu, that should fix that issue16:25
uricamicso I hope it will be ok, when I will update it16:25
n4nd0uricamic: but that thing was for python support or for C++ programs too?16:26
uricamicit was something with 64bit and the intel i7, I think16:26
blackburnfails on i5 as well16:27
uricamicthat valgrind was not able to get some instructuons and then it just refuse to examine the code16:27
blackburnwhen I debug things I disable optimizations and it works16:27
CIA-52shogun: thereisnoknife master * r800c845 / src/shogun/CMakeLists.txt : update src/shogun/CMakeLists.txt -
CIA-52shogun: thereisnoknife master * r764509e / src/shogun/CMakeLists.txt : update src/shogun/CMakeLists.txt -
CIA-52shogun: Sergey Lisitsyn master * r00d3056 / src/shogun/CMakeLists.txt : Merge pull request #755 from thereisnoknife/master -
uricamicblackburn: yep, something like that16:27
uricamicbut the latest release of valgrind should hopefully fix that16:28
n4nd0blackburn: CMakeLists, does shogun support cmake? didn't know about that!16:36
blackburnyeah there was a effort of some guy before16:37
yoon4nd0: you cannot compile shogun using cmake because the lib dependencies are not up-to-date16:46
yoon4nd0: I have fill the CMakeLists.txt with new files just for visualization16:46
yoon4nd0: my IDE read cmake files, that's why ..16:47
n4nd0yoo: aham, but we will be able to do so, compile with cmake, in the future?16:47
yoon4nd0: then we have to had all the dependencies in the root cmakelists.txt16:48
yooI can try to do this in the future but I am not very used to all shogun dependencies since I am still a shogun-noob16:50
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has quit [Quit: Page closed]17:03
-!- yoo [575b08cb@gateway/web/freenode/ip.] has quit [Quit: Page closed]17:06
n4nd0sonney2k: around?17:15
n4nd0see you later guys17:20
-!- n4nd0 [] has left #shogun []17:20
-!- uricamic [~uricamic@2001:718:2:1634:6dc1:2e18:bbfe:a408] has quit [Quit: Leaving.]17:25
-!- n4nd0 [] has joined #shogun19:01
n4nd0nice, we have users from Ernst & Young now :)19:05
@sonney2kn4nd0, ?19:11
n4nd0sonney2k: the guy in the mailing list19:21
@sonney2kn4nd0, ahh I meant to reply to your 'around?'19:26
n4nd0sonney2k: ah ok, that was in case you could merge the PR I issued before since I wanted to put some other changes19:26
n4nd0but at the end I just put everything in the same PR19:26
n4nd0it is a small patch in any case though19:26
CIA-52shogun: Soeren Sonnenburg master * r10341ca / (2 files in 2 dirs): Minor warning fix. -
-!- n4nd0 [] has quit [Ping timeout: 272 seconds]19:49
-!- n4nd0 [] has joined #shogun20:47
-!- blackburn [~blackburn@] has joined #shogun20:59
blackburnoh so e&y uses shogun right?/21:00
blackburninteresting news21:00
n4nd0blackburn: yeah :)21:01
n4nd0blackburn: did you see my patch at github?21:01
blackburnnot yet21:01
blackburnwhat is it about?21:01
n4nd0the generic risk function, I fixed it a bit21:01
n4nd0I get same results with it as with the one michal did in MulticlassModel21:02
n4nd0what is good :)21:02
blackburnwhat was wrong?21:02
n4nd0initialization of subgrad and the computation of the risk21:02
n4nd0the term -<w, psi(xi,yi)>21:03
blackburndo we need to initialize subgrad?21:03
blackburndid you change score in multiclass21:03
blackburnI mean now you are adding loss21:03
blackburndid you subtract it in multiclass?21:03
blackburnsee what I mean?21:04
blackburnbtw I think argmax in training mode should do that21:09
blackburnit could be highly inefficient this way21:09
n4nd0highly inefficient because I substract the delta?21:13
n4nd0blackburn: ?21:20
-!- K0stIa [] has joined #shogun22:10
-!- K0stIa [] has left #shogun []22:10
@sonney2kblackburn, any news on gsomix?22:11
@sonney2kblackburn, btw please kick chris that he does your evaluation...22:12
@sonney2kand wiking - please do yours too22:12
blackburnsonney2k: oops22:17
blackburnwhat is deadline?22:17
blackburnn4nd0: inefficient because of <w,Phi(x,y_truth)>22:17
n4nd0blackburn: yeah sure, but it is the price of having it generic22:18
n4nd0blackburn: I think the best idea is to do as wiking suggested22:18
blackburnwe could assume argmax does that22:18
n4nd0blackburn: aham, we could do that too maybe22:18
blackburnit sounds more efficient22:19
n4nd0I think so too22:20
wikingsonney2k: hye22:21
wikingsonney2k: here?22:22
-!- blackburn1 [~blackburn@] has joined #shogun22:22
wikingsonney2k: managed to get the scan-build working?22:22
-!- blackburn [~blackburn@] has quit [Read error: Operation timed out]22:22
blackburn1sonney2k: chris says he has done it two days ago22:23
blackburn1sonney2k: are you sure?22:24
blackburn1sonney2k: about gsomix he floats in some space probably - I said to him you are missing him22:27
-!- gsomix [~gsomix@] has joined #shogun22:28
blackburn1sonney2k: hooray22:29
blackburn1n4nd0: lets discuss that issue tomorrow22:29
n4nd0blackburn1: ok22:29
wikingblackburn1: yo22:29
gsomixsonney2k, sorry for absence. was helping parents with home t last days22:29
n4nd0blackburn1: in any case I think the patch should be ok since the current StructuredModel::risk is wrong22:30
wikingblackburn1: seems EY people will really going to try using shogun for catching tax fraud :DD22:30
blackburn1yeah I've seen22:30
blackburn1sounds crazy but pretty nice to hear22:30
wikingit'd be great if he could share us what they are trying to use for22:31
wikingi'll email him and ask him... maybe he can share a little bit22:31
wikingthat reminds me: anybody knows how to prepare an .rpm?22:31
wikingmmm as i can see the .deb creator scripts are not within the repo either22:32
wikingsonney2k do you maybe have them somewhere?22:33
wikingi knew once how to create deb package but i think things have changed ever since 2004 :P22:34
-!- blackburn [~blackburn@] has joined #shogun22:42
-!- blackburn1 [~blackburn@] has quit [Ping timeout: 248 seconds]22:44
wikingdoh: incorrect checksum for freed object22:45
blackburnsonney2k: could you please send a mail to chris if there is a problem with my evaluation still22:49
@sonney2kblackburn, why don't you do it?22:50
@sonney2kgsomix, you have to do your final evaluations!22:50
@sonney2kgsomix, ASAP22:50
blackburnsonney2k: I said that because I didn't know you are around :)22:51
blackburnsonney2k: did you check it just?22:51
blackburnis it still not in system?22:51
@sonney2klet me check22:51
gsomixsonney2k, Final Various usability improvements Submitted 2012-08-23 17:30:03 2012-08-23 17:30:0322:52
gsomixalready done, hm22:52
@sonney2kalright - everyone submitted22:53
* sonney2k relaxes22:53
blackburnargh you made all nervous :)22:54
gsomixsonney2k, :) dont worry, be happy22:54
gsomixok, it's a working time22:54
@sonney2kblackburn, well I just (1 hr ago) received reminder emails - so things have just been done.22:56
@sonney2kgsomix, whats the status22:56
blackburnas I said chris was sure he did it two days ago22:56
@sonney2khe didn't :D22:57
blackburnokay anyway it is done and we are all happy22:58
@sonney2kblackburn, wait - he did, wiking, gsomix and arthur didn't22:58
@sonney2kbut now we are back to peace love and you know :D22:58
gsomixsonney2k, preparing PR now. plan to talk with heiko tomorrow (luck?)22:58
@sonney2kgsomix, you should email him - he is writing his masters thesis currently and probably won't be online22:59
@sonney2k(in chat I mean)22:59
@sonney2kreachable via email yes22:59
gsomixsonney2k, ok, I will do this23:00
blackburnsee you23:06
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]23:06
-!- gsomix [~gsomix@] has quit [Quit: Ex-Chat]23:21
n4nd0good night guys23:24
-!- n4nd0 [] has quit [Quit: leaving]23:24
--- Log closed Fri Aug 24 00:00:17 2012