Open in new window / Try shogun cloud
--- Log opened Tue Mar 27 00:00:19 2012
@sonney2kblackburn, yeah I know (you switch topics every 5 minutes so it must be you :)00:00
blackburnsonney2k: well, sorry :)00:01
@sonney2kgsomix, submit what you have then I will try to have a look at the remaining examples tomorrow00:02
@sonney2kgsomix, always note that you are not alone - when you get stuck ask!00:02
blackburnsonney2k: is d/d(w_m) of \sum_m ||w_m||^2 actually equal to 2 \sum_m ||w_m||?00:03
blackburna bit tricky00:03
@sonney2kblackburn, no b in that formula?00:04
blackburnsonney2k: no, C-S makes no use of B00:04
blackburnsonney2k: btw did you know that there are really strong relations between crammer-singer and SO svm?00:05
@sonney2kblackburn, yes00:05
blackburnI was happy to know it00:05
@sonney2ksame idea00:05
@sonney2kor better SO extends this idea...00:05
gsomixsonney2k, thx. of course, I'm not alone. blackburn  lives just on one floor above. :)00:05
n4nd0blackburn: is it what we discussed yesterday a bit?00:05
blackburnn4nd0: yes you made me think about it00:06
@sonney2kgsomix, he doesn't need much sleep so feel free to wake him up at any time ;-)00:06
shogun-buildbotbuild #390 of ruby_modular is complete: Success [build successful]  Build details are at
@sonney2kblackburn, 2 || w_m|| you mean00:06
blackburnactually it is rather difficult to catch me there00:06
blackburnsonney2k: where is sum?00:07
blackburnI am confused with derivative of sum00:07
@sonney2kshouldn't it be zero for all w_i  i!=m00:07
@sonney2kmaybe you choose some other sub index00:08
blackburnI got it00:08
blackburnsonney2k: remember I told you about ECOC SVM?00:08
blackburnsonney2k: I think it could be a great idea00:09
blackburnI asked vojtech and he agreed00:09
blackburnidea is to extend any MC formulation00:09
blackburnin means of ECOC00:09
@sonney2know I don't know what you are talking about00:09
blackburnyes, yes, I'll describe00:10
blackburnsonney2k: recall s.t. in structural svm00:10
@sonney2kwe just received the first student application!00:10
@sonney2k(officially submitted via gsoc melange)00:11
blackburnF is the same as in C-S, but \delta(y_i,y) is code distance (say hamming)00:11
blackburn<w_yi,xi> - <w_m,x_i> >= hamming distance - \xi_i00:11
blackburnsonney2k: crazy, is it someone you know already?00:12
n4nd0harshit said he was planning to start quite early with it ...00:12
@sonney2kblackburn, no00:13
@sonney2kyou all should - this way you can get feedback00:13
n4nd0but not directly to submit via melange right?00:14
n4nd0sonney2k: maybe mail it to you or something?00:14
blackburnn4nd0: why not?00:14
@sonney2kn4nd0, you can alwasy update until the deadline00:14
n4nd0ah ok, forgot about that!00:15
n4nd0then I think it is the most comfortable thing for all00:15
blackburnsonney2k: last thing before I leave - is my idea clear for you?00:15
@sonney2kand we (mentors) are being spammed like hell too :)00:15
@sonney2kblackburn, yes - I have no feeling though how it will change the learned result and if it will be useful00:17
blackburnsonney2k: no serious change00:17
blackburnbut something similar to GMNPSVM00:17
blackburnscaling mostly00:17
blackburnwhile ECOC is more disbalanced than OvR it may impact00:18
blackburnsonney2k: I gave this idea to my gf to derive dual task and stucked myself :D00:18
blackburnthat's why I asked00:19
@sonney2kpoor gf :)00:19
@sonney2kweird way of showing appreciation00:19
blackburnwell she has to come up with some great bachelor work, right? ;)00:20
blackburnsonney2k: my paper on sign recognition is almost ready btw
blackburntake a look on that funny characters lol00:22
@sonney2kblackburn, ok so what you would hope for is that e.g. hamming code is more robust than ovr code above?00:22
blackburnsonney2k: yes00:22
blackburnsonney2k: first shogun mentioning paper in russian ever00:22
@sonney2kblackburn, it is so funny that this is pure russian :)00:23
@sonney2kahh english abstract :)00:23
@sonney2klet me try to decipher how much I understand from the russian one :)00:23
blackburnsonney2k: I bet not much ;)00:24
@sonney2khaha maschina opornich vectorow :)00:25
blackburnyes that is svm00:25
@sonney2knever heard of anyone translating 'SVM' to german00:25
@sonney2kit would be Stuetzvektormaschine00:25
@sonney2kwhich sounds incredibly funny00:25
@sonney2kblackburn, why do you publish in russian?00:26
blackburnsonney2k: cause it is a useless crap00:26
n4nd0when I came last year to Spain after studying some machine learning at KTH I herad a professor talking about MVS00:26
@sonney2kn4nd0, MVS?00:27
n4nd0I think he said it like three times until I realized it was the short for SVM in Spanish :)00:27
n4nd0M?quinas de Vectores Soporte00:27
n4nd0I really didn't know what they were talking about when the said it00:27
@sonney2kI guess wikipedia has it all00:27
blackburnsonney2k: that's the difficult question why I publish in russian00:28
@sonney2kand japanese :
-!- av3ngr [av3ngr@nat/redhat/x-dhyqmsdozibizipk] has joined #shogun00:29
blackburnsonney2k: chervonenkis used to call SVM as it is in wiki00:29
blackburnV-C one you know00:30
@sonney2kwell I guess russians have a good reason00:30
n4nd0hahaha it is even in Swedish00:30
n4nd0I bet the people in the department here have not heard it either00:30
@sonney2kanyway seriously bed time for me now00:30
@sonney2kcu all00:30
blackburnsee you00:30
n4nd0good night00:31
blackburnfor me too probably00:31
blackburnsonney2k: I guess main answer is this paper just supports my bachelor's work and contains nothing really new00:31
blackburnso it would fit only to some paid journals or local russian journal like this one00:32
-!- PhilTillet [~Philippe@] has joined #shogun00:32
blackburnn4nd0: see you tomorrow00:32
n4nd0blackburn: bye!00:33
n4nd0good night00:33
blackburngsomix: hey you00:33
blackburngood night00:33
gsomixblackburn, "Hey you" is the Pink Floyd's song.00:33
gsomixblackburn, good, ok00:33
blackburnreally? I am not a big fan of pink floyd00:33
blackburnI used to listen to stoopid music :)00:34
PhilTilletHey :)00:34
-!- blackburn [~qdrgsm@] has left #shogun []00:35
n4nd0hi PhilTillet00:38
PhilTillethi n4nd0  :)00:39
gsomixsonney2k, i sent pull request.00:41
gsomixsorry for commit "Merge branch 'master'..."00:42
gsomixgood night, guys00:48
n4nd0good night!00:49
gsomixtomorrow I will have a day full of theoretical mechanics. =____=00:50
n4nd0good luck with that ;)00:50
n4nd0gsomix: what do you study btw?00:50
gsomixn4nd0, physics, math, cs. bit of everything.00:53
n4nd0gsomix: cool stuff00:53
-!- wiking [~wiking@huwico/staff/wiking] has quit [Ping timeout: 244 seconds]01:04
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun01:20
-!- makokal [~makokal@] has quit [Quit: makokal]01:23
-!- n4nd0 [] has quit [Quit: leaving]01:39
-!- makokal [] has joined #shogun01:39
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]02:11
-!- makokal [] has quit [Ping timeout: 246 seconds]02:39
-!- makokal [] has joined #shogun02:40
-!- av3ngr [av3ngr@nat/redhat/x-dhyqmsdozibizipk] has quit [Quit: That's all folks!]03:36
-!- pluskid [~chatzilla@] has joined #shogun03:40
pluskidwhere is LinearByteKernel? is it removed? or renamed to something else?03:57
-!- derekhh [8f59be22@gateway/web/freenode/ip.] has joined #shogun04:06
-!- pluskid_ [~chatzilla@] has joined #shogun04:16
-!- pluskid [~chatzilla@] has quit [Ping timeout: 260 seconds]04:18
-!- pluskid_ is now known as pluskid04:18
-!- pluskid [~chatzilla@] has quit [Quit: ChatZilla [Firefox 11.0/20120312181643]]04:24
-!- vikram360 [~vikram360@] has joined #shogun05:14
-!- Varun_Agrawal [] has joined #shogun05:41
Varun_AgrawalHey guys!05:41
Varun_AgrawalSorry for the long hiatus05:41
Varun_AgrawalHad my GRE yesterday.05:41
-!- Varun_Agrawal [] has quit [Client Quit]05:44
derekhhwish you gl :)05:52
-!- derekhh [8f59be22@gateway/web/freenode/ip.] has quit [Quit: Page closed]05:52
-!- n4nd0 [] has joined #shogun07:30
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun08:12
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]08:27
-!- PhilTillet [~Philippe@] has quit [Ping timeout: 248 seconds]09:02
-!- Miggy [~Piggy@] has joined #shogun09:03
-!- Miggy [~Piggy@] has quit [Quit: Leaving]09:16
n4nd0gsomix: hey!09:59
n4nd0gsomix: do you think you can help me a moment with git?09:59
gsomixn4nd0, let's try. however, i sometimes experience problems with it.10:07
n4nd0gsomix: so I was on the point of doing a pull request10:07
n4nd0but it didn't allow me to push before pulling first10:08
n4nd0so I fetched shogun's repo and rebased it10:08
n4nd0and later I pushed10:08
n4nd0and it has turned out to be like this10:08
-!- harshit_ [~harshit@] has joined #shogun10:08
n4nd0you can see that lot of the commits are done twice :(10:08
n4nd0one authored by me and the other the original10:09
n4nd0and now if I try to pull request10:09
n4nd0it understands that I am pushing all that stuff as new10:09
n4nd0not just the changes I have done but also the changes I got when I rebased10:09
n4nd0do you see what I mean?10:10
-!- in3xes [~in3xes@] has joined #shogun10:16
n4nd0harshit_: hey! yesterday we spoke a bit with sonney2k about deep learning and stuff, I suggest you to check the irclogs10:20
n4nd0if you have not already done so of course :)10:21
harshit_okay i'll have a look at it10:21
harshit_actually i am in my class right now, so i'll look at it in an hour or so..10:22
n4nd0sure, no problem, they will remain there ;)10:22
harshit_and thanks for talking dude10:23
-!- in3xes [~in3xes@] has quit [Ping timeout: 264 seconds]10:25
gsomixn4nd0, unfortunately, I have no ideas about it.10:36
gsomixi will try to appeal to blackburn.10:36
n4nd0gsomix: no worries man! I just fixed it :)10:36
n4nd0 so nice again :)10:37
n4nd0gsomix: thank you very much anyway for taking a look to it ;)10:37
gsomixn4nd0, oh, ok.10:38
-!- blackburn [~qdrgsm@] has joined #shogun10:38
gsomixn4nd0, tell me what you did, please.10:39
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun10:40
blackburnwhat's up10:41
gsomixblackburn, hi10:41
wikingblackburn: yo10:41
blackburnwiking: yo10:41
blackburngsomix: please rebase your pull request up to latest master10:43
blackburnand push it again10:43
blackburngit push --force origin if needed10:44
n4nd0blackburn: that was the trick I had to use git push --force10:44
n4nd0blackburn: why does that happen?10:44
blackburnn4nd0: what exactly?10:45
n4nd0blackburn: that we need to use --force after rebasing10:45
blackburnn4nd0: origin at github becomes unsynchronized with your local git10:47
n4nd0n4nd0: aham, I see10:48
blackburnactually it shouldn't require force anytime10:48
n4nd0blackburn: fuck, talking to me again10:48
n4nd0I had to do it before, at the beginning I avoided it but I almost screwed my branch10:48
blackburnI don't get why it can't fast-forward it though10:48
wikingblackburn: don't force it10:49
wikingn4nd0: don't force it10:49
wikingyou'll fuck up your commit history...10:49
wikingand then people following you will have problems10:49
wikingso rather try a rebase10:50
wikingand if that does not work10:50
n4nd0wiking: how should we do it then?10:50
wikingdo a merge10:50
n4nd0wiking: ok I see10:50
n4nd0wiking: I did the rebase but later it didn't allow me to push to my branch because of this non-fast-forward not allowed10:50
n4nd0wiking: so when that happens, better to merge?10:51
wikingfirst try a git pull10:51
wikingso if you did rebase and it did not allow you to have a git push after that10:51
wikingtry a git pull10:51
wikingit'll try to do a simple merge10:51
wikingand then after that try a git push10:51
blackburnhah, that makes more sense, agree10:52
n4nd0ok, thank you guys :)10:52
wikingno worries10:52
blackburnactually I think force is required only in case of wrong rebasing10:52
blackburnit should work as fast-forward thing..10:52
n4nd0blackburn: I have new news about one-vs-one10:53
n4nd0blackburn: I had to change a couple of things more because the seg faults persisted
n4nd0the features in m_machines were NULL at training10:54
n4nd0I fixed that with the changes in MulticlassMachine:8910:55
n4nd0and the other thing I had to do, I don't like that much how it's though is MulticlassMachine:6810:57
blackburnn4nd0: I am afraid line 90 is wrong10:59
blackburnconsider case data=NULL and m_features was set before11:00
blackburnand train() is called11:00
blackburnit will init train machine with NULL11:00
n4nd0but init train machine doesn't set anything if data is NULL11:02
n4nd0it just refreshes the las set features for m_machines11:02
-!- harshit_ [~harshit@] has quit [Ping timeout: 248 seconds]11:03
blackburnhmm sorry I have to go11:03
blackburnI'll reach you in a hour or little more11:03
n4nd0no problem11:03
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]11:03
-!- harshit_ [~harshit@] has joined #shogun11:08
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds]11:18
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun11:19
-!- makokal [] has quit [Quit: makokal]11:36
-!- gsomix [~gsomix@] has quit [Ping timeout: 244 seconds]11:39
-!- gsomix [~gsomix@] has joined #shogun11:49
-!- harshit_ [~harshit@] has quit [Ping timeout: 276 seconds]12:06
-!- romovpa [bc2c2ad0@gateway/web/freenode/ip.] has joined #shogun12:22
-!- vikram360 [~vikram360@] has quit [Ping timeout: 264 seconds]12:30
romovpahi there, I want to participate in gsoc. I should demonstrate my familiarity with git and ability to build the sources12:31
romovpais there somebody who tell me how to do that?12:31
n4nd0hey romovpa12:40
n4nd0first you should fork the repo in github, checkout the source, get a bit used to it ...12:40
n4nd0so you can work on something and submit a small patch to the project! :)12:40
n4nd0if you and idea you'd like to do for this patch, it is ok then12:40
romovpaI have already forked and touched it12:40
n4nd0romovpa: good!12:41
n4nd0so would you like to contribute with something in special?12:41
romovpahere is my small patch:
n4nd0that one is already merged12:42
n4nd0good then12:42
romovpahmm... should I do anything more?12:42
n4nd0I mean it's up to you :)12:43
romovpaok, =) thanks12:43
romovpaLet it be my success report (I'm filling up my application now)12:44
n4nd0romovpa: I hope you didn't take that wrong12:44
n4nd0good luck with it12:45
n4nd0may I ask what project are you applying for?12:45
gsomixromovpa, hello from Russia.12:45
romovpaI am going to implement a bundle method solver for SO-SVM12:46
romovpagsomix: hello, could we know each other?12:47
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has joined #shogun12:47
blackburnn4nd0: ok lets check12:48
gsomixromovpa, i do not think so.12:48
n4nd0blackburn: so what did you say before that could go wrong? something that was getting null12:49
blackburnok let me describe logic12:49
blackburnm_features/m_kernel is always a state12:49
blackburnso then you set features/kernel and call train() it should train machine and prepare for applying to exactly these ones12:50
blackburnwhen apply(CFeatures) is called it sets features for machines and applies12:50
n4nd0ok, I understand12:50
n4nd0the problem here is that somehow if we just do init_machines_for_applly before starting to train; it seg faults12:52
blackburninit_machines_for_apply should be called AFTER train12:52
n4nd0aham I see12:52
n4nd0just before the return?12:52
blackburnyes probably12:52
blackburndid you get this logic?12:53
n4nd0yeah, I think so12:53
n4nd0I think that solution is better12:53
blackburn3 cases12:53
n4nd0but still I don't know what may be wrond doing init_machines_for_train(NULL) in apply()12:53
blackburnin apply?12:54
blackburndamn this naming is pretty hard to understand12:54
n4nd0at it's right now in MulticlassMachine.cpp:6812:54
blackburnn4nd0: could you please rename it?12:54
blackburninit_training_machine or so12:54
blackburnmay be tihs way?12:54
blackburnmachine and machines look confusing12:55
blackburnlemme check the source12:55
blackburnn4nd0: ok let me describe a wrong case12:56
n4nd0blackburn: that was one issue, there is another to look at12:56
n4nd0blackburn: sorry, continue12:57
blackburnwhen you call it12:57
blackburnit sets features12:57
blackburnand calls apply()12:57
blackburnthis time apply() sets features with NULL12:57
blackburnand BADDABOOM12:57
n4nd0I don't think so12:58
n4nd0because init_machines_for_apply12:58
n4nd0doesn't set the features if they're NULL12:59
n4nd0if will skip the set features and reload the ones for m_machines12:59
n4nd0the latter is what we are interested in12:59
blackburnit was beyond my intention12:59
blackburnoh really12:59
blackburnI got wrong with my own logic12:59
n4nd0idk, maybe init_machines_for_apply doesn't do the right thing in LinearMulticlassClassifier13:00
n4nd0but it does so this far I think13:00
blackburndamn I've got some doubts13:02
blackburnuse cases are13:03
blackburnset features / set kernel13:03
n4nd0in any case I think that to do init_machines_for_apply at the end of train is better13:03
blackburnset features / set kernel13:03
blackburnno more cases I guess13:03
n4nd0I will check if they're ok13:04
n4nd0apart from that13:04
n4nd0I also changed MulticlassMachine::train_machine13:05
n4nd0the very beginning13:05
n4nd0lines 87 - 9013:06
n4nd0it's a similar idea, but the thing is that if it is not done this way, it seg faults in training13:06
blackburnit should fail if it !is_ready()13:07
n4nd0I am thinking now that maybe it is appropriate to do init_train_machines at the end of the constructor that takes features as argument13:07
blackburnand there is no data13:07
blackburnif there is no data*13:07
n4nd0yeah I think that it's true, if features was not initialized in the constructor13:09
n4nd0then the constructor with features must call init_machines_for_train13:10
n4nd0do you see my point?13:10
blackburnthat was forgotten probably13:10
blackburnnot machines :D13:11
blackburnyeah embarrasing naming13:11
n4nd0haha yeah I noted that13:11
blackburnfeel free to change it13:11
blackburnit is in guts of these classes so ok to change13:12
blackburnno API13:12
n4nd0can we impose somehow that class that inherit from MulticlassMachine13:12
n4nd0are compulsed to do these init_training_machine13:12
n4nd0in the constructor that accepts Features?13:13
blackburnit calls constructor of base class13:13
n4nd0I think I need to implement that constructor then13:13
blackburnso while init_training_machine is virtual13:13
blackburnoh wait13:14
n4nd0what happened?13:15
blackburnhmm no way to do that probably13:16
blackburnthen it shouldn't be in the constructor13:17
n4nd0wait I don't see it13:17
n4nd0why no way?13:17
blackburnwell multiclass machine constructor can't get any features13:17
n4nd0why not?13:17
blackburncurrently I mean13:18
n4nd0even if features is not a member13:18
blackburnwhy not to init_training_machine in train_machine?13:18
blackburnin case of NULL data and is_ready()?13:18
blackburnisn't it exactly what we need?13:18
n4nd0yes ... I think so at least13:19
n4nd0then we will call init_training_machine with NULL13:19
blackburnI hope this all makes sense13:20
n4nd0we'll see13:21
n4nd0one thing about the doc, where should I put it? this might what S?ren told me yesterday but I might have not understood13:26
n4nd0I mean for example, for LDA13:26
n4nd0I would like to do sth similar for QDA13:27
n4nd0as it appears in doxygen doc13:27
n4nd0but I have not found that documentation in the code13:27
n4nd0wait, I said nothing13:27
n4nd0it is in the code ...13:28
blackburnn4nd0: yeah @brief part in .h13:29
n4nd0blackburn: I am working in the code, in the train_machine part13:32
n4nd0I really think that if ( !is_read() && !data ) then SG_ERROR13:32
n4nd0in any other case init_training_machine13:33
n4nd0init_training_machine(data) simply13:33
n4nd0i.e. as it is in my branch right now13:33
n4nd0is it normal that LibLinear doesn't always give the same results?13:35
blackburnn4nd0: multiclass?13:36
n4nd0the example I am using for MulticlassMachine uses LinearMulticlassMachine with LibLinear13:36
n4nd0blackburn: yes13:36
blackburnyes it is a little random13:36
blackburnresults shouldn't be *very* different though13:36
n4nd0every once in a while13:37
blackburnwhat kind of random you've got?13:37
n4nd0I get a bunch of class 3 detected for what it is class 013:37
blackburnrandom labels?13:37
n4nd0normally it classifies good13:37
n4nd0but every 1 out 4 or 5 let's say, that happens13:37
blackburnthat's strange13:37
n4nd0possible bug you think?13:38
blackburnyes it should be almost equal..13:39
n4nd0it happens just with ONE_VS_ONE actually :S13:39
n4nd0what's going on?13:41
blackburnI thought you mean MulticlassLibLinear13:41
blackburnand was very afraid of it13:41
n4nd0no no13:41
n4nd0so what do you think then?13:42
n4nd0bad result?13:42
blackburnn4nd0: well depends on data actually but strange anyway13:42
-!- romovpa [bc2c2ad0@gateway/web/freenode/ip.] has quit [Quit: Page closed]13:46
blackburnn4nd0: valgrind is the tool ;)13:46
blackburnn4nd0: what is the email addr you reached nico with?13:50
-!- n4nd0 [] has quit [Quit: leaving]14:10
-!- Peter_Gr [~Peter_Gr@] has joined #shogun14:43
-!- vikram360 [~vikram360@] has joined #shogun14:45
wikingblackburn: here15:16
wikingok so now it's starting to be ready15:20
wikingjust one more commit15:20
blackburnwiking: what exactly? ;)15:21
wikinglatent svm15:21
blackburnlets see15:22
wikingcheck this15:25
wikinglast two commits15:25
wikingor rather15:25
blackburnI'd suggest to move these things to shogun'/latent15:27
blackburni think it is ok to move all latent domain specific things to shogun/latent15:28
blackburnnew dir15:28
wikingbut why?15:28
wikinganyhow here you go a new stupid pull request for shogun :))15:29
blackburnwiking: I think domain specific separation is ok15:30
blackburnI actually plan to move ready multitask things to shogun/multitask15:30
blackburnand multiclass things to shogun/multiclass15:30
wikingbtw is this your email adds: Sergey Lisitsyn <>15:31
blackburnwiking: yes both works15:31
blackburnblackburn91 at gmail15:31
blackburnand lisitsyn.s.o15:32
wikingok i'll send there now the email to you as well15:32
wikingthat goes for alex as well15:32
blackburnah ok sure15:32
wikingok email sent... apply my pull request :D15:41
blackburnwiking: can't right now15:47
blackburnhmm didn't received your mail15:47
blackburnwould be funny if it was considered as spam15:47
wikingmaybe it is15:49
blackburnah it is ok15:51
-!- harshit_ [~harshit@] has joined #shogun15:52
-!- harshit_ [~harshit@] has quit [Ping timeout: 252 seconds]16:08
-!- harshit_ [~harshit@] has joined #shogun16:22
-!- harshit_ [~harshit@] has quit [Ping timeout: 248 seconds]16:47
-!- harshit_ [~harshit@] has joined #shogun16:47
-!- jckrz [] has joined #shogun16:51
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has quit [Quit: Page closed]16:51
-!- harshit_ [~harshit@] has quit [Ping timeout: 248 seconds]16:52
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]17:08
-!- blackburn [~qdrgsm@] has joined #shogun18:01
-!- harshit_ [~harshit@] has joined #shogun18:02
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds]18:14
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun18:20
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]18:30
-!- flxb [] has joined #shogun18:34
flxbhey all18:37
flxbI saved a classifier with pickle.dump() and then loaded it. When I do apply it throws "terminate called after throwing an instance of 'shogun::ShogunException'". Is there a way to get a more meaningful error?18:39
-!- flxb [] has left #shogun []18:42
-!- flxb [] has joined #shogun18:44
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun18:45
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds]18:55
-!- flxb [] has quit [Quit: flxb]18:55
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun19:00
-!- blink_123 [~blink@] has joined #shogun19:08
-!- harshit_ [~harshit@] has quit [Ping timeout: 240 seconds]19:15
-!- blink_123 [~blink@] has quit [Ping timeout: 246 seconds]19:19
-!- makokal [~makokal@] has joined #shogun19:20
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds]19:26
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun19:29
-!- makokal [~makokal@] has quit [Quit: makokal]19:45
-!- Marty [9eb54c39@gateway/web/freenode/ip.] has joined #shogun20:04
MartyHi, anybody on?20:07
MartyRunning svm.train() Traceback (most recent call last):   File "", line 64, in <module>     svm.train()   File "/usr/local/lib/python2.7/dist-packages/", line 20344, in train     return _modshogun.Machine_train(self, data) SystemError: [ERROR] assertion num<get_num_vectors() failed in file features/StringFeatures.cpp line 26420:09
-!- Marty [9eb54c39@gateway/web/freenode/ip.] has quit [Quit: Page closed]20:13
-!- flxb [] has joined #shogun20:19
-!- MartysBrother [~chatzilla@] has joined #shogun20:22
-!- wiking [~wiking@huwico/staff/wiking] has quit [Read error: Connection reset by peer]20:23
-!- MartysBrother is now known as Marty2820:24
Marty28testing again20:25
Marty28Running svm.train()20:25
Marty28Traceback (most recent call last):20:25
Marty28  File "", line 64, in <module>20:25
Marty28    svm.train()20:25
Marty28  File "/usr/local/lib/python2.7/dist-packages/", line 20344, in train20:25
Marty28    return _modshogun.Machine_train(self, data)20:26
Marty28SystemError: [ERROR] assertion num<get_num_vectors() failed in file features/StringFeatures.cpp line 26420:26
Marty28This I get with WeightedDegreeStringKernel20:26
Marty28My small toy example works.  Are there size limitations for feature data? I have 2200 x 141 DNA base pairs as training data.20:28
-!- makokal [] has joined #shogun20:28
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun20:38
-!- wiking [~wiking@huwico/staff/wiking] has quit [Read error: Connection reset by peer]20:40
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun20:42
-!- wiking_ [~wiking@huwico/staff/wiking] has joined #shogun20:46
-!- wiking_ [~wiking@huwico/staff/wiking] has quit [Client Quit]20:47
-!- wiking [~wiking@huwico/staff/wiking] has quit [Read error: Connection reset by peer]20:47
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun20:47
-!- Marty28 [~chatzilla@] has quit [Quit: ChatZilla [Firefox 11.0/20120310010446]]20:47
CIA-64shogun: Viktor Gal master * r306bc99 / src/shogun/distance/ChiSquareDistance.h : A little typo fix of define in ChiSquareDistance -
CIA-64shogun: Soeren Sonnenburg master * r1f71363 / src/shogun/distance/ChiSquareDistance.h :20:49
CIA-64shogun: Merge pull request #396 from vigsterkr/master20:49
CIA-64shogun: There was a little typo in the ChiSquareDistance header -
-!- wiking [~wiking@huwico/staff/wiking] has quit [Client Quit]20:50
-!- flxb_ [] has joined #shogun20:58
-!- flxb [] has quit [Ping timeout: 246 seconds]21:01
-!- flxb_ is now known as flxb21:01
-!- blackburn [~qdrgsm@] has joined #shogun21:14
@sonney2kgsomix, why did you close
@sonney2kexcept for missing rebase it looks very valid to me and should be merged...!21:26
blackburnsonney2k: blame him in russian ;)21:39
-!- n4nd0 [] has joined #shogun21:45
flxbI saved a classifier with pickle.dump() and then loaded it. When I do apply it throws "terminate called after throwing an instance of 'shogun::ShogunException'". Is there a way to get a more meaningful error or what could this be?21:45
blackburnflxb: try to add init_shogun_with_defaults() once you import shogun21:46
@sonney2kblackburn, he is using python already (pickle...)21:50
@sonney2kflxb, well you could do
flxbsonney2k: ok i don't know what effect init_shogun_with_defaults has. i'll try bla io log level 021:51
flxbsonney2k: ha, but what is bla? it's not in modshogun21:54
blackburnflxb: any shogun object you use21:54
@sonney2kflxb, some shogun object :-)21:54
flxbamusing, i'll give it a try22:02
n4nd0sonney2k: hi! I have a doubt related to solving SVM optimization with the primal22:08
@sonney2kn4nd0, ?22:08
n4nd0sonney2k: so can it be solved directly using a QP solver (let's say CPLEX or libqp)?22:09
n4nd0sonney2k: is there no need in using an specific SVM solver for that?22:09
@sonney2kn4nd0, svm's can be solved (in primal or dual) with off-the-shelve QP solvers22:10
@sonney2kbut that's slow and memory demanding22:10
@sonney2kso tuned solutions are much faster / require less memory22:10
n4nd0sonney2k: aha, I understand22:10
n4nd0sonney2k: is there something already in shogun to solve the SVM in primal?22:11
blackburnhey any crazy guy wanting to implement sparse matrix class?22:11
@sonney2kn4nd0, liblinear / libocas22:11
n4nd0blackburn: aren't CSparseFeatures for that?22:12
blackburnn4nd0: I need pretty specific one22:12
@sonney2kprobably more :)22:12
@sonney2kblackburn, what's wrong with CSparseFeatures?22:12
n4nd0sonney2k: oh all these can be used for primal? cool!22:12
@sonney2kn4nd0, primal (with no kernel!)22:12
blackburnsonney2k: nothing but I need dymamic CSC matrix22:13
blackburnsonney2k: I thought ocas is dual :(22:14
blackburnmy bad22:14
n4nd0sonney2k: hmm I have to check if they could be used for the problem I am thinking of, for SO-SVM optimization22:14
@sonney2kblackburn, it uses cutting planes...22:14
blackburnhmm true22:14
blackburnI do not understand it thoroughly but I know it stands for s.t. part in primal ;)22:15
@sonney2kn4nd0, I suspect you would need sth like libqp but ask nico22:18
n4nd0sonney2k: no way to do it with OCAS because of the kernel issue?22:19
@sonney2kn4nd0, actually vojtech has an extension to ocas idea for SO22:20
@sonney2kthat is the other SO task btw22:20
@sonney2kn4nd0, one is more the framework and some baseline algorithms, the other SO one would be one application of the framework but bleeding edge and potentially even better than svm struct22:21
n4nd0sonney2k: mmm but this idea other talks about libqp and not OCAS22:22
gsomixsonney2k, I had some problems with repo.22:22
@sonney2kgsomix, so can you open a new pull request?22:22
@sonney2kn4nd0, this idea
gsomixsonney2k, yep.22:23
@sonney2khmmhh we have 5 student submissions by now - but none is from any of the active students here...22:24
n4nd0sonney2k: oh sorry, I'm kind of retarded ... I was checking another one22:24
gsomixsonney2k, but only tomorrow. I have some changes and I will complete it tomorrow.22:24
n4nd0we have to get hands on to work with the application guys22:25
blackburnoh not now22:25
n4nd0I think I am going to start tonight with it :)22:25
blackburnsonney2k: would you mind I attach a list of my commits?22:25
n4nd0blackburn: :D22:26
@sonney2kblackburn, the post gsoc ones only please :)22:26
blackburnsonney2k: ok, I've commited about 200 after gsoc :D22:27
* gsomix was trying to do home work today. But he did little. (tag: whining.)22:29
n4nd0blackburn: so should we attach the commits as well? I mean, isn't enough to give a brief description about what have we done?22:30
blackburnthat was joke!22:30
n4nd0haha ok22:31
n4nd0at first I took it like that22:31
n4nd0but after the post gsoc comment22:31
n4nd0sonney2k: do you know if Nico has access to the applications? I'd like to ask him for feedback on the timeline for my application22:32
blackburnwell sonney2k is a great joker ;)22:32
n4nd0I wonder if I should send it to him directly22:32
@sonney2kn4nd0, yes - he is a mentor but it doesn't hurt to explicitly ask him22:34
n4nd0thank you :)22:35
n4nd0sonney2k: is it important to write some references (e.g. professors we have interactuated with)? or not really that much22:43
@sonney2kn4nd0, everything that separates you from other students helps!22:46
blackburne.g. if you've got three legs22:47
n4nd0blackburn: :D22:48
blackburnn4nd0: want to try dimensionality reduction?22:50
blackburngsomix: want to try dimensionality reduction?22:51
n4nd0blackburn: yeah! sure22:51
n4nd0blackburn: what do you have in mind?22:52
blackburnn4nd0: examples, I lack examples22:52
n4nd0blackburn: yes, I like the idea22:52
n4nd0blackburn: have you thought anything in particular?22:53
blackburnok simplest22:53
blackburndownload a few images somewhere22:53
blackburncompute hue histograms22:53
blackburnembed it with isomap/mds22:53
blackburnwill be a nice picture :)22:53
blackburncheck applications/edrt/ to see how to place images on plot with matplotlib22:54
blackburnoh actually I had a script somewhere22:54
blackburnbtw it is my gf ;)22:54
n4nd0check the previous to last page22:55
n4nd0sth like that would be nice?22:55
blackburnI'll give you some code in a min22:56
n4nd0blackburn: so these slides are from one of the courses I told you the other day22:57
blackburnaha I got it22:57
n4nd0blackburn: we could prepare an example like this one we are talking about22:57
n4nd0blackburn: and I can mail the professor and show her22:58
n4nd0blackburn: maybe that way we can start some shogun-university courses contact22:58
blackburnwhy not :)22:58
blackburnn4nd0: ^ that's how it worked22:58
n4nd0blackburn: looks nice22:59
n4nd0blackburn: you have the code for that right?22:59
blackburnyeah let me find it22:59
gsomixblackburn, dimensional reduction? no, I do not want to live in two-dimensional world.22:59
gsomixif seriously, i will work on python3 and covertee.22:59
blackburnremember I always have tasks22:59
blackburnn4nd0: ok it is pretty crappy but here it goes23:00
blackburnthere is even a hog function I borrowed somewhere23:00
CIA-64shogun: Soeren Sonnenburg master * rc375b8f / (5 files in 3 dirs): make director based classes optional -
blackburnn4nd0: all you need is to get some pictures23:01
blackburnideally retrieve from google.. but it seems they do not provide such api23:01
n4nd0mmm not good23:02
blackburnn4nd0: any other options are welcome23:02
blackburnhaha I just realized all the pictures are upside down23:02
n4nd0I will make a search about how to do get images, if it doesn't work, I'll pick some manually23:03
n4nd0probably we don't need that many to make a nice-looking example23:03
n4nd0any suggestion/preference of the class of pictures?23:03
blackburnno, it is totally up to you23:03
blackburnah yes23:04
blackburninstall libsuperlu323:04
blackburnand libarpack23:04
blackburnto make this feasible23:04
blackburnin case of 3k-5k images it would be slow w/o superlu+arpack23:04
blackburnhowever I can hardly imagine that23:04
blackburnn4nd0: have you checked applications/edrt btw?23:05
blackburnthere are words embedding example as well23:05
n4nd0I have not executed it23:05
n4nd0which one is it?23:05
n4nd0... I feel stupid sometimes I ask :O23:06
n4nd0blackburn: libarpack2 is ok I guess23:06
n4nd0haha nice example23:08
blackburntry as well23:08
n4nd0fun one :)23:09
n4nd0isn't she sad a couple of times but very separated on the plane :P?23:09
blackburntoo complex for any dim reduction algorithm23:09
blackburnactually I should prepare data somehow23:10
n4nd0I see23:10
blackburnwith centering at leat23:10
blackburnthere are two more examples23:10
blackburnwith ORL and swissroll23:10
n4nd0I like swissroll23:10
blackburnyes it is easy but not real23:11
blackburnlocal tangent space alignment is ideal with swissroll23:11
blackburnbut not as good with real data23:11
@sonney2kn4nd0, I want pictures of giraffes, zebras, horses etc :)23:17
n4nd0sonney2k: I'll look for those then ;)23:21
n4nd0oh, didn't understand anything in that last mail :S23:24
blackburnheiko wrote in german oh lol23:27
blackburnI shall answer him in russian23:30
n4nd0can you read it?23:31
blackburnno, how can I23:31
n4nd0haha don't know ... maybe you know German man23:32
blackburnI wish I knew23:32
blackburnok sent23:33
blackburnn4nd0: one more mail you won't understand ;)23:34
n4nd0blackburn: hahaha you are fun man23:34
n4nd0blackburn: wanna do a guess what google translate tells me?23:34
blackburnsomething reasonable I think23:34
n4nd0quite reasonable actually23:35
n4nd0I generally do not understand what you wrote here! But to keep the conversation going write some thing about LibSVM too :)23:35
n4nd0so the reverser R in Russian is I23:35
n4nd0or flipped23:35
blackburnn4nd0: ya23:36
blackburnnot i23:36
blackburnlike in young23:36
blackburnah you mean "I"23:36
blackburnI thought of sound23:36
n4nd0ah ok23:36
blackburnyes it is "I"23:36
n4nd0I was like ... there is no i in young so I lost sth probably23:36
@sonney2kblackburn, heikos mail was a private mail to me ...23:40
blackburnsonney2k: why did he CC shogun list then?23:41
@sonney2kblackburn, I think he accidentally sent it to shogun list23:41
@sonney2kin the end it says 'migration business is complex'23:41
blackburnsonney2k: well why not to joke ;)23:41
@sonney2khe managed to fix things but it took quite long23:41
@sonney2kand he is going to submit a patch ...23:42
blackburnyes I did make use of google translate already23:42
gsomixgood night, guys23:47
n4nd0gsomix: good night23:49
@sonney2kyeah good night23:49
blackburnsonney2k: I don't think it was private ;)23:54
-!- vikram360 [~vikram360@] has quit [Ping timeout: 252 seconds]23:54
@sonney2kblackburn, you mean my reply? no it wasn't intended to be :)23:55
@sonney2kshould I have replied in german?23:55
blackburnsonney2k: yes or japanese according to way it was carried out ;)23:56
@sonney2kblackburn, you can still join the thread and reply in russian :D23:56
blackburnI already did once23:56
@sonney2kthen n4nd0 can continue in swedish or spanish :)23:57
blackburnn4nd0: yeah continue thread with something swenish23:57
blackburnor spadish23:57
blackburnI actually usually write in runglish23:58
n4nd0I am a shy guy to reply23:58
-!- flxb [] has quit [Read error: Connection reset by peer]23:58
-!- flxb_ [] has joined #shogun23:58
blackburnheiko from londongrad23:59
@sonney2kn4nd0, blackburn consumes a lot of vodka - you know he is *russian*23:59
blackburnyes I do23:59
--- Log closed Wed Mar 28 00:00:03 2012