Open in new window / Try shogun cloud
--- Log opened Tue Apr 03 00:00:19 2012
-!- wiking [] has joined #shogun00:02
-!- wiking [] has quit [Changing host]00:02
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun00:02
-!- tibi_popa [tibi_popa@] has left #shogun []00:03
n4nd0shogun-buildbot: oh, C# successful, good job!00:04
shogun-buildbotWhat you say!00:04
gsomixshogun-buildbot, test.00:05
-!- harshit_ [~harshit@] has quit [Ping timeout: 265 seconds]00:06
-!- harshit_ [~harshit@] has joined #shogun00:06
n4nd0gsomix: how is it going with your proposal man00:07
gsomixn4nd0, i'm writing now.00:08
n4nd0gsomix: cool! do you find useful the one I sent you?00:08
-!- flxb [] has joined #shogun00:11
gsomixOh. Jimi Hendrix and Duane Allman are ones of a greatest guitarists in history... I can listen to them a very long time.00:21
gsomixBut it's time to sleep.00:22
gsomixGoog night.00:22
-!- karlnapf [] has joined #shogun00:24
harshit_n4nd0: please send me sample proposals also ..00:29
harshit_having sometrouble making mine :(00:29
n4nd0harshit_: trouble? why so man?00:29
harshit_For creating the detailed timeline00:30
harshit_what material were you talking about with gsomix ?00:30
n4nd0I can send you that part then00:30
n4nd0in any case, what projects are you applying for?00:30
n4nd0it would be kind of weird to share proposal for the same project :P00:30
harshit_C5.0 + regression solvers of liblinear00:30
harshit_I think yours is SO learning . right ?00:31
n4nd0ok, so give me an address and I can send you the timeline for the multiclass project I have done today ok?00:31
n4nd0or SO, the one you prefer, they are pretty similar in structure actually00:31
harshit_here : harshit.syal@gmail.com00:32
n4nd0but I warn you, I think that mine's not pretty detailed at all00:32
n4nd0it's quite basic actually :S00:32
harshit_but still i'll get an idea00:32
n4nd0yeah, sure00:33
n4nd0please use it just as an idea ... it would be weird if we all send a very similar proposal :P00:33
harshit_yeah sure , i wont copy it :)00:34
harshit_but i thought they ask for weekly timeline00:34
harshit_i mean details of each weak00:35
n4nd0I don't really know00:35
n4nd0I suggest you to ask S?ren to get a better idea what do they want00:35
n4nd0try to ask when you meet him around here so I can get also an answer for that :)00:36
harshit_sure ;)00:36
harshit_g2g its about 5 here , need to sleep00:37
n4nd0yeah sure00:37
n4nd0I was going to ask you about that man00:37
n4nd0becuase I see that you are with us in the chat lot of times00:37
n4nd0but you have another time !00:37
harshit_yeah have to manage somehow.otherwise my work wont progress00:38
harshit_hasta luego00:38
-!- harshit_ [~harshit@] has quit [Quit: Leaving]00:38
-!- harshit_ [~harshit@] has joined #shogun00:39
harshit_forgot to mention, Thanks for sending me your timeline00:40
-!- harshit_ [~harshit@] has quit [Client Quit]00:40
n4nd0you are welcome, good night00:41
-!- av3ngr [av3ngr@nat/redhat/x-tfqcnnyvsaidfgdq] has joined #shogun00:50
-!- karlnapf [] has left #shogun []01:03
shogun-buildbotbuild #425 of python_modular is complete: Success [build successful]  Build details are at
n4nd0shogun-buildbot: great!!01:20
shogun-buildbotWhat you say!01:20
-!- mohataher [52c9fc37@gateway/web/freenode/ip.] has joined #shogun02:12
mohataherHi, I'm mohamed, applying for GSoC and I have a question regarding the application form for it. Is there anyone to answer my question now ? :)02:13
-!- nickon [] has quit [Quit: ( :: NoNameScript 4.22 :: )]02:19
n4nd0mohataher: hey mohamed! maybe I can help you02:21
n4nd0tell me02:21
mohataherThank you. Regarding sending a pull request, does it have to include any modification to the code/documentation or just a to try it out ?02:57
n4nd0it should include a patch03:00
n4nd0I think there was a conversation about that in the mailing list03:01
n4nd0but definitely, an empty patch is not enough03:02
-!- LeStealth [] has joined #shogun03:14
mohataherokay, thanks for your answer. I will search for it.03:15
n4nd0let me know if you need help to find it03:16
mohataheri found the mailing list and I'm skimming it now. I will let you know if I need help. Thank you.03:21
mohataherI found the answer. Thanks.03:24
-!- mohataher [52c9fc37@gateway/web/freenode/ip.] has quit [Quit: Page closed]03:24
-!- LeStealth [] has left #shogun ["Leaving"]03:27
-!- hpsoar_ [7c108901@gateway/web/freenode/ip.] has joined #shogun03:37
-!- hpsoar_ [7c108901@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]03:44
-!- n4nd0 [] has quit [Quit: leaving]03:45
-!- flxb [] has quit [Quit: flxb]03:58
-!- pythonroar [d2198538@gateway/web/freenode/ip.] has joined #shogun04:32
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]05:07
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun05:26
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]05:52
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Remote host closed the connection]06:19
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun06:19
-!- pythonroar [d2198538@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]06:50
-!- n4nd0 [] has joined #shogun07:55
@sonney2kn4nd0, so is our buildbot happy again?08:07
n4nd0sonney2k: it looked like that tonight :)08:08
n4nd0let's see if it continues well08:08
n4nd0sonney2k: yay! everythin all right08:08
CIA-64shogun: Heiko Strathmann master * rbad4740 / src/shogun/features/Subset.cpp : small cast fix -
CIA-64shogun: Soeren Sonnenburg master * r0e4f543 / src/shogun/features/Subset.cpp :08:10
CIA-64shogun: Merge pull request #413 from karlnapf/master08:10
CIA-64shogun: small SG_ADD type cast fix -
@sonney2kso shogun-buildbot work!08:11
@sonney2kn4nd0, heh even warnings down to nothing :)08:11
n4nd0sonney2k: cool!08:11
sonne|workn4nd0: btw how does the mkl picture look like?08:44
n4nd0sonne|work: this is for very few samples08:48
sonne|workn4nd0: could you run this with shading interp ?08:49
n4nd0sonne|work: I think it is already set like that08:49
n4nd0sonne|work:, line 11608:50
sonne|worksomething is wrong then08:52
sonne|workcontours look pretty ok IMHO08:52
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun08:53
sonne|workI guess the first test to fix this would be to use just 1 kernel08:53
sonne|workand if it still looks weird with 1 then figure out what the difference is...08:54
sonne|workn4nd0: I just tried - looks still totally broken ... but no time to go deeper for now08:58
n4nd0sonne|work: what do you suggest to use a CombinedKernel just appending one more kernel or one PolyKernel alone?08:58
n4nd0sonne|work: yeah, I tried changing a couple of things and it broke too :S08:59
sonne|workI even used a single GaussianKernel08:59
sonne|workstill all weird08:59
n4nd0I can try to take it a look and show you the results in some minutes09:00
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]09:11
-!- pythonroar [d2198538@gateway/web/freenode/ip.] has joined #shogun09:19
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun09:37
-!- blackburn [~qdrgsm@] has joined #shogun09:42
n4nd0blackburn: trying to scare people? :P09:44
blackburnno people scared me09:45
blackburnmoaaaar pull requests09:45
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]09:50
-!- Marty28 [~Marty@] has joined #shogun09:51
n4nd0Marty28: hey!09:52
-!- harshit_ [~harshit@] has joined #shogun09:57
sonne|workhey harshit_!10:10
sonne|worknice work10:10
sonne|workbtw I didn't understand your benchmark numbers10:10
sonne|workhow long did it take on covertype?10:11
sonne|workfor shogun and how long with just matlab?10:11
-!- harshit_ [~harshit@] has quit [Ping timeout: 246 seconds]10:14
-!- harshit_ [~harshit@] has joined #shogun10:19
-!- blackburn [~qdrgsm@] has quit [Ping timeout: 265 seconds]10:25
-!- n4nd0 [] has quit [Ping timeout: 276 seconds]10:27
-!- harshit_ [~harshit@] has quit [Ping timeout: 244 seconds]10:32
* gsomix went to classes about computer networks.10:34
-!- harshit_ [~harshit@] has joined #shogun10:49
-!- harshit_ [~harshit@] has quit [Quit: Leaving]11:03
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Remote host closed the connection]11:21
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun11:22
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Remote host closed the connection]11:22
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun11:23
-!- karlnapf [] has joined #shogun11:25
-!- flxb [] has joined #shogun11:31
-!- av3ngr [av3ngr@nat/redhat/x-tfqcnnyvsaidfgdq] has quit [Quit: That's all folks!]11:51
-!- nickon [] has joined #shogun11:56
-!- wiking [] has joined #shogun12:09
-!- wiking [] has quit [Changing host]12:09
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun12:09
-!- flxb [] has quit [Quit: flxb]12:26
-!- n4nd0 [] has joined #shogun12:26
-!- flxb [] has joined #shogun12:28
n4nd0I am taking a look to the seg fault Nicholas has pointed out12:48
-!- flxb [] has quit [Quit: flxb]13:03
n4nd0sonne|work: around?13:18
n4nd0the way to do CV for kernel machines is to use the constructor that takes a machine and no features right?13:20
n4nd0the kernel has to be set to that machine13:20
sonne|workI think so - karlnapf is the expert for that13:24
karlnapfhum? just read my name :)13:24
karlnapfwell you can specify a kernel before13:25
karlnapfdepends on what you want to do13:25
karlnapfjust simple CV for evaluation or parameter search?13:25
n4nd0karlnapf: I am trying to fix the example of the guy that sent the mail to the list13:26
n4nd0his problem is that when he does CV with LibSVM it seg faults13:26
karlnapfah, ok13:26
karlnapfCV means what?13:26
n4nd0the kernel has never been set, we should lock that13:26
n4nd0cross validation, sorry13:26
karlnapfjust for simple evaluation or for a parameter search?13:27
n4nd0so now I am trying to find out how to do it for LibSVM, I have managed to get something running13:27
n4nd0I am afraid I cannot answer that :S13:27
n4nd0I can so you the code of the example if so and you can tell me if it is all right, ok?13:28
karlnapfmmh, well13:28
karlnapfyes sure13:28
n4nd0so in the line 4513:28
n4nd0before it was simply LibSVM(), and it turned out to seg fault for this issue that we get to execute a method for the kernel member of LibSVM13:29
karlnapfyes you have to set the kernel before training in that case13:29
n4nd0and the kernel had never been initialized13:29
karlnapfthe parameter search does not involve the kernel at all13:29
n4nd0so this is cross validation then?13:29
karlnapfno its a parameter search13:29
karlnapfbut this is based on cross-validation13:30
karlnapfsince you have to test your current parameters somehow13:30
karlnapfand thats done with CV13:30
n4nd0ok, I think I get the main idea13:30
karlnapfthere are two seperate classes for that, you can see them in the example13:30
n4nd0my only concern is that now to create the LibSVM13:30
karlnapfdoesnt that example work?13:30
n4nd0I have to use a value for C, and the labels13:31
n4nd0yes yes, it does13:31
n4nd0but idk if it does what it should do ;)13:31
n4nd0so the point is13:31
n4nd0when the instance of LibSVM is created now13:31
n4nd0this values, are used somewhen?13:31
karlnapfwhich values?13:32
n4nd0the values for C and the labels13:32
karlnapfah ok13:32
karlnapfwell the labels are used in training, in the CV, they are needed13:32
n4nd0I mean it because C is supposed to be tried out within an interval13:32
karlnapfand the C value is just overwritten13:32
n4nd0but the labels are given again in the CV constructor, line 5513:33
karlnapfoh, yes, true13:33
karlnapfwell, then you could create the SVM with an empty constructor13:33
karlnapfand set just the kernel by hand instead13:33
n4nd0I think that is fairly reasonable13:33
karlnapf(or so)13:33
karlnapfwe ineed do sometimes not check for good initialization13:34
karlnapfthats currently all a bit complicated i guess :)13:34
n4nd0it works as well with the way you've suggested13:34
karlnapfnice then13:35
n4nd0and I like it more than using the labels in the constructor13:35
n4nd0I found that confusing13:35
karlnapfyes its nicer that way13:35
karlnapfsay n4nd0, are you fernando?13:36
n4nd0so I think that I should just add a fix in the method that was seg faulting to fire SG_ERROR in case the kernel is null13:36
n4nd0karlnapf: yeah! you are Heiko right?13:36
karlnapfI am a bit confused by all these nicknames13:36
karlnapfyes :)13:36
n4nd0but yours is even more difficult to guess :P13:37
karlnapfyes, an error message if there is no kernelk would be nice13:37
karlnapfyes, true ;)13:37
karlnapfwhen blackburn shows up, we can talk about that subset stuff13:37
n4nd0karlnapf: sure13:37
n4nd0in the meantime I will add a pull request with the fix and answer Nicholas13:38
karlnapfnice one13:38
karlnapfI will continue on my work now, but will have one eye on the irc, see you!13:38
n4nd0thank you, bye!13:39
n4nd0sonne|work: jasmine's pull request looks nice!13:58
-!- n4nd0 [] has quit [Ping timeout: 240 seconds]14:06
-!- sidsniper [~piggy@] has joined #shogun14:27
CIA-64shogun: iglesias master * rce07314 / (2 files): * fix seg fault when trying to do CV with no kernel initialized, * fix indent -
CIA-64shogun: Soeren Sonnenburg master * r2b4bdc2 / (2 files):14:30
CIA-64shogun: Merge pull request #415 from iglesias/fix-modelsel14:30
CIA-64shogun: Small fix to avoid seg fault -
-!- sidsnipe [~piggy@] has joined #shogun14:32
-!- flxb [] has joined #shogun14:32
-!- sidsniper [~piggy@] has quit [Ping timeout: 260 seconds]14:33
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has joined #shogun14:33
-!- sidsnipe [~piggy@] has quit [Client Quit]14:33
blackburnkarlnapf: hey there14:34
karlnapfhey blackbutrn14:41
karlnapfblackburn, hi14:41
blackburnkarlnapf: I heard you wanted to kill me because of machine subsetting"? ;)14:42
karlnapfkrkr ;)14:42
karlnapfno, just asking about it14:42
karlnapfexplain it so me14:42
blackburnlets take a look14:42
blackburnonce you train machines OvO14:43
blackburnyou need to subset your trainset14:43
karlnapfOvO is one versus one?14:43
karlnapfok, why the need for subsets?14:44
blackburnbecause it would be painful14:44
blackburnto create new features14:44
blackburneach time you want to construct ovo machine14:44
karlnapfno I meant why needs OvO that?14:44
karlnapfsorry doorbell14:44
blackburnhmm because all the classes but some (m,n) pair should be ignored14:45
blackburnany better way would be welcome but I can't see any14:46
sonne|workblackburn: so this is 'only' to avoid copying of data around right?14:47
blackburnsonne|work: yes probably14:47
karlnapfHow does LibSVMMutlicalss do this?14:47
sonne|workkarlnapf: they have some kind of subset index internally14:47
karlnapfand why not subsetting features?14:48
sonne|workkarlnapf: possible...14:48
sonne|workwhat about model selection then?14:48
karlnapfblows up14:48
sonne|workcan we have subsets of subsets14:48
karlnapfYes, one solution would be to have multiple subsets14:49
sonne|workif we had a function to say train(features, subset) it would work14:49
karlnapflike we talked about one year ago or so14:49
karlnapfthats possible on the locked data14:49
sonne|workor a function that returns a subset given a subset as input14:49
karlnapfbut thats not good for OvO14:49
sonne|workthat might be sufficient ...14:49
karlnapfI mean14:49
sonne|workso a function that computes a new subset based on the current subset14:49
karlnapfmodelselection on kernel machines without locking is a pain anyway14:49
sonne|workonly if you can store the kernel in memory...14:50
sonne|work(that was never the case in the applications I had)14:50
karlnapfyes true14:50
karlnapfok so subsetting the machine14:51
karlnapfthat seems to be a good solution, however, we got many subset mechanisms then14:51
karlnapfand they basically all take subsets of features14:51
karlnapfif there was a way to register multiple subsets on features14:51
karlnapfwe wouldnt need all that stuff14:51
sonne|workok true14:52
karlnapflike give each subset a unique identifier14:52
karlnapfand pass it before accessing them14:52
sonne|worksome dynarray of subsets...14:52
sonne|workbut actually that would kill performance14:52
karlnapfand how to access?14:52
sonne|workmaybe one can add subsets as one wants14:53
karlnapffeatures subsets are currently used implicitly14:53
sonne|workand then a 'current' subset is computed based on the subsets in the list14:53
sonne|workso accessing features uses only this 'current' subset14:53
sonne|workmaybe current is the wrong term 'merged subset' is better14:53
karlnapfoh yes, they would have a  hierachy14:54
karlnapfstill the problem:14:54
karlnapfA sets subset A14:54
karlnapfB sets another subset on that14:54
karlnapfB is not ready yet14:54
karlnapfA wants to access14:54
karlnapfand because A is just setting the subset once and then accesses implicitly this is a problem14:55
sonne|workwell I would say that B has to free its subset before passing on to A14:55
sonne|workotherwise this gets too difficult14:55
blackburnI do not really understand why you need any hierarchy or other stuff14:56
karlnapfthen it would be quite straightforward to implement right?14:56
blackburnwhat are you trying to solve right now?14:56
karlnapfblackburn, the problem is that modelselection subsets features14:56
karlnapfand if the multiclass svm also does that, there is a kind of conflict14:56
karlnapfa "subset on a subset"14:56
karlnapfwould be needed, but currently thats not possible14:57
blackburnyes got it14:57
karlnapfand having multiple subset mechanisms (features, multiclass, locking) is a bit ugly14:57
sonne|workkarlnapf: yes I think this is doable14:57
blackburnmulticlass do not do subsetting actually14:57
blackburnit subsets features14:57
blackburnor kernel14:57
blackburnnothing moer14:57
sonne|workjust some 'add_subset' / 'remove_subset' functions14:57
sonne|workand whenever add/remove is called the 'merged' subset is updated (which is what we currently have)14:58
sonne|workso no need to change the code we have now - just some additions14:58
blackburnsubset union subset?14:58
sonne|workactually one needs to change all code that sets a subset14:58
sonne|workit needs to use add/remove_subset14:58
blackburneach time you add subset it subsets the subset?14:58
sonne|workblackburn: it computes the subset of the subset of the subset - yes14:59
blackburnsonne|work: it may break things then14:59
-!- harshit_ [~harshit@] has joined #shogun14:59
karlnapfwe should set a maximum subset level to avoid abuse of this system ;)14:59
blackburnearlier it was assumed that there is only one subset14:59
sonne|workblackburn: what does it break?15:00
blackburnsonne|work: no idea but karlnapf could know probably15:00
karlnapfyou lost me, what?15:00
sonne|workkarlnapf: blackburn is asking when setting subsets of subsets would break in the current system15:00
karlnapfno I think not15:01
blackburnok then15:01
karlnapfone just HAS to remove subsets after doing work on them15:01
blackburnone issue is that add and remove are not interchangeable or so15:01
sonne|workyeah so it should be safe15:01
blackburnI mean add applies it on top of that15:01
blackburnbut remove removes all of them15:01
blackburnor I got wrong?15:02
sonne|workremove removes 115:02
blackburnyou want to do stack there?15:02
harshit_sonne|work : I have updated the pull request, I think now its ready to be merged15:02
blackburngood idea probably15:02
sonne|workjust an array of subsets15:02
karlnapfPerhaps this removing could be possibly  done automatically in some destructor15:02
blackburnI thought you wanted to merge subsets actually15:02
sonne|workharshit_: Thanks - I will have a look later15:02
sonne|workblackburn: yes for performance15:02
karlnapfthen we would not have seperate subset systems at three places15:03
sonne|workwe have one 'merged' subset15:03
blackburnwell two subsets is ok probably15:03
sonne|workwhich is computed from the array of subsets15:03
sonne|worksuch that we don't have to go through the whole array of subsets all the time15:03
blackburnkarlnapf: which three places?15:03
karlnapfyes, all the subset instances are saved but everytime the stack is touched, atemporary "current" one  is created for performance thight?15:03
karlnapfmodsel, locking, multiclass15:03
blackburnwell I see no other way for multiclass ovo15:04
karlnapfBut when we have this new thing, you can just subset the features before training the class pairs15:05
karlnapfand then the machines just train on "all" data (from their perspective)15:05
sonne|workfight karlnapf fight15:05
karlnapflol ;)15:05
karlnapfI went kick-boxing yesterday, still have Muskelkater, so, no thanks ;)15:06
blackburnkarlnapf: actually I think subset stack is a good idea not only wrt to ovo15:06
sonne|workI also like it15:06
sonne|workso who wants to implement it?15:06
blackburni.e. what if you want to do cross-validation on subset of your features?15:06
karlnapfthats possible then15:07
sonne|workblackburn: features -> you mean subset of dimensions15:07
sonne|workor subset of examples?15:07
blackburnoh no sorry15:07
blackburnfeature vectors15:07
karlnapfthat would work then15:07
blackburnsubsets of features would be interesting however it is a different system15:07
sonne|workthat's fine15:07
karlnapfyou just would have to check for class indices everytime, not just once in the beginning15:07
blackburnis it required btw?15:08
karlnapfyes, subsets of dimensions would be cool for feature selection15:08
karlnapfbut thats probably another story15:08
karlnapfI cou?d implement the stuff15:08
blackburnsonne|work: btw what do you think about sparse linear machine?15:08
blackburnkarlnapf: please do if you have time ;)15:09
blackburnI can but you would probably be able to test it better15:09
karlnapfIts not ultra urgent, is it?15:09
sonne|workblackburn: what for? what does it do?15:09
blackburnkarlnapf: no, not urgent at all15:09
blackburnsonne|work: lasso15:09
sonne|workblackburn: please elaborate15:10
karlnapfbecause my parents are visiting me tomorrow and I only have a few hours per day, and also want to come up with the hypothesis test interfaces this week15:10
blackburnsonne|work: lasso has only a few non zeroes in [w]15:10
sonne|workkarlnapf: but it should be rather easy15:10
sonne|workmaybe n4ndo would want to do it?15:10
blackburnor gsomix?15:10
sonne|workI mean it would just mean to add these to add/remove functions15:10
sonne|workblackburn: no gsomix is busy with python315:11
karlnapfits ok, Ill do it15:11
sonne|workand the merge function15:11
sonne|workthat's all or?15:11
karlnapfthere might be trouble with x-val15:11
blackburnsonne|work: do you have any other tasks for gsomix?15:11
sonne|work...can be fixed when it occurs :)15:11
-!- harshit_ [~harshit@] has quit [Quit: Leaving]15:11
sonne|workblackburn: e.g. octave_modular typemap for sparse features15:11
blackburnsonne|work: because I want somebody to do LARS and he would do that15:11
blackburnI see - then let him do this octave15:12
sonne|workblackburn: you still didn't explain what sparselinearmachine is!15:12
sonne|workgsomix: are you busy?15:12
blackburnsonne|work: yes,yes - what if w is sparse?15:12
karlnapfsonne|work, I read the mail yesterday about the parameter selection for GPs, I think there might be some potential for creating new model selection classes15:13
sonne|workblackburn: ahh you mean not to store whole w vector15:13
blackburnrather not to compute outputs with whole one15:13
blackburnit is the case of lasso15:13
sonne|workblackburn: or one just adds a sparsify_w function ...15:13
karlnapfone could generalize the modsel class to use specific evaluation functions (like model evidence or LOO) and then also specify gradients of them and create new GradientDescentModelSelection class or so15:14
sonne|workfor training it is tough...15:14
blackburnsonne|work: sparse training you mean?15:14
blackburnyes no need to15:14
sonne|workkarlnapf: you would probably have to help this student then15:14
karlnapfyes, I like that kind of stuff15:15
sonne|workblackburn: one way would be to then call compress_w and then apply uses the compressed representation of w15:15
blackburnSNOW in april!15:15
karlnapfhehe :)15:15
sonne|workwe just had that on sunday15:15
blackburnit is all dirty here15:15
karlnapfI got foggy rainy Londonish sky here :)15:15
blackburnand snow again15:16
blackburnwhat the fog!15:16
karlnapfthere was snow in germany?15:16
blackburnsonne|work: ok then lars goes to n4ndo probably15:17
blackburnI hope I do not overload him :D15:17
sonne|workblackburn: if you know better what gsomix is interested in - tell me - or gsomix tell me :)15:18
blackburnsonne|work: he is at uni right now probably15:18
blackburnI saw him hour ago15:19
blackburnsonne|work: director stuff btw15:20
blackburnsonne|work: what is # of proposals?15:20
sonne|workblackburn: he is more fond of that?15:20
sonne|workdirectors I mean15:21
blackburnsonne|work: I do not know15:21
blackburnI'll ask15:21
blackburnor you15:21
sonne|workit is tough though and I guess he would need to write an isolated thing15:21
sonne|workblackburn: 21 proposals by now15:22
sonne|work3 days left...15:22
blackburnI think new candidates should finish some small successful tasks15:22
sonne|workyes - even after deadline but before we decide15:22
blackburnthat's why I am afraid of sparse matrix and other stuff15:23
blackburncan you estimate men-day's? :D15:23
sonne|worksparse matrix?15:24
blackburnsonne|work: software eng managers used to estimate man hours and man days :D15:25
blackburnsonne|work: sparse matrix support for octave15:25
sonne|workthat is very easy (for me)15:26
sonne|workso lets say 2 days max15:26
blackburnok nice15:27
blackburnsonne|work: will you have time to guide him?15:27
karlnapfhey guys, Do you know this possibility of approximating a kernel matric using incomplete cholesky factorisation?15:27
karlnapfI recently used that and found it quite cool, so perhaps there would be some room for it in shogun15:28
blackburnkarlnapf: what is *incomplete* cholesky?15:28
blackburnah let me google that15:28
karlnapfsent you an email15:29
karlnapfits an extract from the kernel methods book of john shawe taylor and nello crist....15:30
blackburnoh I'll check it later15:31
blackburnkarlnapf: did you send the proposal alread15:31
karlnapfhope you like it :)15:32
blackburnkarlnapf: this year we are going to receive 20% of last year ones :D15:32
karlnapfwhat do you mean?15:32
karlnapfah like re-applying? :)15:33
blackburnin total count15:34
blackburnkarlnapf: only you and me are re-applying afaik15:34
sonne|worklast year we had like 70 applications - this year it will be around 30 I guess15:35
karlnapfreally? how does that come?15:35
sonne|workthis is because we required patches even before applying15:35
karlnapfprojects too hard?15:35
sonne|workand applications/candidates are much stronger15:35
karlnapfmmh, was that a good idea then? the number of slots is determined by the number of applications thight?15:35
blackburnnot only I think15:36
sonne|workyeah from that perspective not - but who can wade through 70+ applications?15:36
sonne|workeven reading them is *a lot* of work15:36
blackburnsonne|work: especially if they are done by russians15:36
karlnapfhehe :)15:37
sonne|workand last year we had like 3 reviewers per student app15:37
sonne|workand scores / final ranking15:37
blackburnsonne|work: who did the job?15:37
sonne|workwe split work somehow15:37
blackburnsonne|work: I mean who have participated in that?15:38
sonne|workcandidates that we were more sure of we wanted were reviewed by more people then15:38
sonne|workof course15:38
sonne|workthat is how it works15:38
blackburnI know, I am just curious :)15:38
sonne|workI wish blackburn / karlnapf you both would be mentors instead of students15:38
karlnapfHopefully next year :)15:39
blackburnsonne|work: if it was same-rewarded15:39
karlnapfI desperately need money, unfortunately15:39
karlnapfso expensive here15:39
sonne|workno reward unfortuantely15:39
blackburnyes that's why we want to be students :)15:40
* sonne|work wants to be student to15:40
-!- pythonroar [d2198538@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]15:40
karlnapfHow is work at tomtom btw?15:40
sonne|workthey have great data...15:41
sonne|workand lots of15:41
sonne|workand even more data analysis problems15:41
karlnapfwhat are you actually doing there? Like what kind of problems?15:41
sonne|workso quite interesting15:41
blackburnsonne|work: I bet your year salary is larger than gsoc reward ;)15:41
sonne|workI am working in the HD traffic team - that is the one predicting jams / any kind of live traffic stuff15:42
blackburnsonne|work: btw can I intern there? ;)15:43
sonne|workblackburn: you should go to university and become professor15:43
blackburnwould be funny15:44
blackburnsonne|work: why should I?15:44
sonne|workotherwise we will never have students to work for you on shogun15:44
karlnapflol :)15:44
blackburnhah good reason to devote whole life for useless stuff there15:44
sonne|workkarlnapf: nothing to laugh here ;) you have the same destiny15:46
blackburnsonne|work: what about you then?15:46
sonne|workI left academia...15:46
karlnapfwe are all converging on growing shogun students ;)15:47
blackburnyes I heard that before but no way back?15:47
blackburnactually we've got one kind of brilliant  haha15:47
blackburnsonne|work: what I think is there is no need to worry we are not the mentors15:49
sonne|workkarlnapf: I am attempting to grow+educate 2 kids to become shogun developers :D15:49
blackburnfirst shogun girl?15:49
blackburnwe would have to wait at least 15 years more though15:50
sonne|workblackburn: I do - every help is welcome (as always)15:50
blackburnsonne|work: you do worry you mean?15:50
sonne|worknowadays programmers start 615:50
karlnapfhehe, your kids15:50
karlnapfI got a BASIC interpreter for my AMIGA for my 10th birthday from my father ;)15:51
blackburnsonne|work: I think being mentors would mean we have no time for develop..15:51
blackburnI got my first computer when I was 715:51
karlnapfAnd my parents always wanted me to do computer science, but I refused and went studying jazz guitar first ;)   so dont put too much pressure up, sonne|work ;)15:53
sonne|workkarlnapf: my son is 11 months and want to play games on my android phone :)15:55
sonne|workbut I don't understand how I got interested in this15:55
blackburnsonne|work: he is also used to multiclass stuff already15:55
sonne|workmaybe because my parents always wanted to limit computer-time15:56
karlnapfhehe, dont let him become a super-nerd ;)15:56
karlnapfyes same here15:56
sonne|workso I guess that is the key15:56
sonne|worktough limits (he wants to break)15:56
blackburnsonne|work: you should deny him to code perfectly later15:57
* sonne|work thinks that karlnapf would have finished this sub-subsetting business by now if we wouldn't fool around...15:57
karlnapfI am studying ;)15:58
blackburnand yes deny him to construct theory better than vapnik did15:58
karlnapfbut youre right, I should stop chatting15:58
karlnapftake care guys :) (wave)15:58
blackburnmy son, never develop general pattern recogntion theory better than VC and PAC15:58
blackburnsonne|work: that's the thing you should say everyday ;)15:59
sonne|workkarlnapf: hope to see you in chat more often (idling)15:59
karlnapfyes, will leave it on more, uni just stopped so I can be online all day : bye16:00
blackburnsnow stopped16:03
sonne|workkarlnapf: bye16:04
blackburnsonne|work: have you used clearcase ever?16:05
blackburnif someone asks you to use it16:05
blackburnbetter run16:05
blackburnor hide somewhere16:05
blackburnI heard it is legal to kill person that suggests to use clearcase16:06
-!- pluskid [chatzilla@] has joined #shogun16:08
-!- flxb_ [~cronor@] has joined #shogun16:09
-!- flxb [] has quit [Ping timeout: 264 seconds]16:11
-!- flxb_ is now known as flxb16:11
blackburnsonne|work: that paper I showed to you is to appear in shity 'computer optics' 36(2) :D16:15
sonne|workcomputer optics???16:16
blackburnyes that is the name of journal..16:17
blackburnsonne|work: do you think I am crazy? I am not welcome to Russia :D16:17
blackburnam not,*16:18
sonne|workthe world is crazy...16:18
blackburnsonne|work: it is a kind of local journal with rather easy submission16:20
blackburnsonne|work: and I needed it to 'power up' my bachelor's work16:20
blackburnthe most funny thing there is abstracts16:21
blackburn'In this paper is considered a method of forming a sequence of reliefs based on the specified reference terrain images by using a criterion based on computation a conjugation indexes between a current reliefs and a reference reliefs. An example of the implementation is given.'16:21
sonne|worksome lower machine learning conference would also have worked16:21
sonne|workanyways got work to do cu16:21
blackburnsonne|work: can you believe it is *real* abstract?16:21
blackburnoh sure16:21
blackburnsee you16:21
blackburnconference would work but no time for that16:21
wiking blackburn i've decoded the email!!! :DDD16:23
blackburnwiking: hahah16:23
blackburntook a while right?16:24
wikingwell yeah16:24
-!- n4nd0 [] has joined #shogun16:24
blackburnn4nd0: can you remind me what is your 'contribution roadmap'? ;)16:25
blackburnwiking: so what does he want?16:25
wikinga little bit of change16:25
wikingnot so much16:25
wikingbasically we would remove the whole CLatentFeatures16:25
n4nd0blackburn: what do you mean, contributions in shogun?16:25
blackburnn4nd0: yes16:25
wikingand just have CFeatures16:25
wikingbut they would contain all the features: x,y,h16:26
blackburnwiking: and include things to latentlabels?16:26
blackburnhow to separate it when?16:26
wikingwell you don't need to16:26
n4nd0blackburn: from the very beginning, CPLEX compilation, Mahalanobis distance, QDA, OvO, SPE16:26
wikingoptimizer doesn't need to know it16:26
n4nd0blackburn: not that much :S16:26
wikingwhat part of the feature is what16:26
wikingthe only thing that after the results16:26
wikingthe user of the tool will have to know what is what16:27
blackburnn4nd0: sorry I meant what is next?16:27
n4nd0blackburn: aham!16:27
blackburnwiking: ah I see16:28
n4nd0blackburn: so first we should manage to get merged SPE, did you see I updated it?16:28
wikingand then still one _could_ define an own function to speed up the optimization algo16:28
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds]16:28
blackburnwiking: that is hardly implied from the msg :D16:28
n4nd0blackburn: later, what about the OvO issue? I see you have been discussing long about it16:28
wikingbut if it's not provided it should just do a brute force search16:28
wikingi'm just wondering now16:28
blackburnn4nd0: we shall wait for new subset stack concept ready16:29
wikingwhether the whole class should be inherited.... i.e. have a protected virtual function for finding the argmax of the latent variable16:29
n4nd0blackburn: ok16:29
blackburnn4nd0: I've got a new task for you thought, LARS16:29
blackburnwiking: whole class of?16:29
-!- flxb [~cronor@] has quit [Remote host closed the connection]16:29
wikingor have a class that only has a function like this.. and that class should be inherited and redefine that function for finding the argmax16:29
wikingblackburn: LatentLinearMachine16:29
-!- flxb [] has joined #shogun16:30
n4nd0blackburn: tell me about it16:30
wikingany input on it?16:30
wikingi'd rather go with the latter option16:30
wikingit feels better for me16:30
blackburnn4nd0: generalized? simple would work as well :)16:30
blackburnwiking: then go with latter ;)16:31
n4nd0blackburn: give me a pointer if so16:31
blackburnn4nd0: scikits learn has one16:31
blackburnn4nd0: I'll take a look on spe later16:31
n4nd0blackburn: cool16:31
pluskidn4nd0: maybe this one?
blackburnyes this one16:32
pluskidyou wish to include LARS in shogun for, solving LASSO ?16:32
n4nd0regression, cool16:32
pluskidLARS is cool, but I heard that there are many more efficient algorithm for L1 regularized problems?16:33
n4nd0aham, e.g.?16:33
blackburnpluskid: I have not much expertise with linear regression, what would you suggest?16:33
pluskidI do not know the literature (not familiar with optimization), actually, I always use LARS in my own work :p16:34
pluskidblackburn: I heard some discussion about something like ??? gradient descent algorithms, but I do not understand the details16:35
n4nd0I know gradient descent16:35
pluskidBut I do think LARS has an advantage that it gives you the whole path16:35
pluskidI think many other algorithms only gives you the solution for a particular \lambda (regularization coefficient)16:35
pluskidMaybe any of you could check out this package: , if you are good at optimization16:36
blackburnnice reference16:37
blackburnhmm that's something I should take care of16:38
blackburngroup based lasso is of interest for my project16:38
pluskidhmm, I think those ideas interesting, too. But I never used them in practice yet, except simply L116:39
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun16:40
pluskidTrying to find the discussion thread about LASSO optimization I have read before16:41
pluskidGot it!16:43
pluskidI list some references here, in case any one would be interested in.16:44
pluskidThey said that this algorithm is very fast:
pluskidAnd for really huge data, online learning algorithms also exist:16:44
pluskidThat's it. :)16:46
blackburnpluskid: thanks! nice references16:46
pluskidTime for sleep here.16:46
pluskidgood night/afternoon/morning guys!16:46
-!- pluskid [chatzilla@] has quit [Quit: ChatZilla [Firefox 11.0/20120312181643]]16:47
blackburngood night16:47
n4nd0good sth ;)16:49
n4nd0so after the ideas that you suggested yesterday and the new one today, LARS ...16:51
n4nd0let's go for LARS because it is the newest :)16:51
blackburnn4nd0: what have I suggested yesterday?16:52
n4nd0the tree16:53
n4nd0the cover tree for knn too16:53
n4nd0parzen windows16:54
blackburnn4nd0: what about parzen window one with covertree internally? later we would merge good design from it to knn16:55
n4nd0blackburn: idk if you're serious now16:55
blackburnyes I am16:56
n4nd0so parzen windows can be implemented with cover trees??16:56
blackburngood question16:56
blackburnprobably not16:58
blackburnn4nd0: actually it could if there was some threshold based covertree17:00
blackburnbut it would require patching covertree that is present in shogun17:00
n4nd0blackburn: let's do that new threshold based covertree and write a paper :P17:02
blackburnI am afraid it is way too simple idea17:02
blackburnoriginal covertree supports it IIRC17:02
@sonney2kblackburn, is covertree only for euclidian distances?17:05
@sonney2kor can it work with any distance?17:05
blackburnsonney2k: any17:05
blackburnit requires nothing but pairwise distances17:05
@sonney2kok then n4nd0 this definitely is my favourite then covertree knn :)17:06
n4nd0i'll focus on that then sonney2k17:08
CIA-64shogun: Harshit Syal master * r85ffdcd / src/shogun/classifier/svm/NewtonSVM.cpp : removed unwanted SG_SPRINT -
CIA-64shogun: Soeren Sonnenburg master * ra934dcb / (7 files in 5 dirs):17:10
CIA-64shogun: Merge pull request #409 from harshitsyal/master17:10
CIA-64shogun: Newton SVM -
n4nd0I'll be back later, see you guys17:15
-!- n4nd0 [] has quit [Quit: leaving]17:15
@sonney2kshogun-buildbot, ?17:15
-!- sidsniper [~piggy@] has joined #shogun17:17
sidsniperhi blackburn17:19
sidsniperhow can i test dim red algos? is there any data?17:19
blackburnsidsniper: what kind of testing you want to do?17:19
sidsniperreducing dimension of feature vector17:20
sidsniperhow did you test the correctness of algoriths implemented?17:20
blackburnmanually ;) and compared to other implementations17:21
sidsniperso there must be some dataset which u had tested them upon?17:21
blackburnyes, I tested it on swissroll, ORL and MIT-CBCL17:22
blackburnswissroll is the simplest test I'd say17:22
sidsniperso these are availaible online right?17:23
blackburnsidsniper: yes17:23
sidsniperok thanks :)17:23
blackburnsidsniper: have you doubts about its correctness?17:23
sidsniperi have yet to test them..i m sure they will be correct :)17:24
blackburnI'm going to implement unit-tests as well..17:24
sidsniperif you can give me links from where you have implemented them it would be great17:24
sidsniperORL and CBCL are face database right?17:26
blackburnsidsniper: implemented what?17:26
blackburnsidsniper: yes17:26
sidsniperimplemented the algorithms17:26
sidsniperdim red17:27
blackburnhmm I mainly refer to dr toolbox17:27
sidsnipernice :)17:28
sidsniperi have to implement some of them for my project17:28
sidsnipercan u suggest how to proceed?17:28
sidsniperi mean which are the easier ones17:28
blackburnsidsniper: LLE is pretty simple17:30
blackburndiffusion maps are as well17:30
blackburnthey are all actually pretty simple17:30
blackburnbut pretty slow in case of straightforward implementation17:31
sidsniperso implementation part is hard?17:31
blackburndepends on speed you need17:31
blackburnI had to implement superlu and arpack wrappers17:31
blackburnand to parallelize it17:31
blackburnbut in matlab it could be 20-liner17:32
blackburnas for correctness17:33
sidsniperbut we have to do it in R language17:33
blackburnthey are all ill-posed17:33
sidsniperwill that be tough?17:33
blackburnno idea17:33
sidsniperi saw that shogub also suppports R17:33
blackburnhow much data you have?17:33
-!- harshit_ [~harshit@] has joined #shogun17:33
sidsnipercureently i dont have17:33
blackburnharshit_: congrats your code is merged ;)17:34
sidsniperi will first try on swissroll as suggested by u17:34
blackburnsidsniper: yeah try ;)17:34
harshit_hey blackburn: is there any way to get precision , recall, accuracy etc on tested dataset in shogun itself ?17:34
harshit_and thanks i didnt check it ,17:35
blackburnharshit_: yes, ContingencyTableEvaluation17:35
blackburnok heading home finally17:35
blackburnsee you17:36
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has quit [Quit: Page closed]17:36
harshit_github is strange doesn't notify you for getting your code merged !17:36
sidsniperharshit_: which project are u applying for?17:37
harshit_sidsniper :Its a collection of things actually17:39
gsomixsonney2k, sonne|work slap.17:39
harshit_2 projects for which i am applying are: integrate regression in liblinear + implement some cdotfeatures17:40
harshit_and one is trees + liblinear17:40
harshit_sidsniper : what about you ?17:40
gsomixsonney2k, sonne|work I want to finish my proposal today. That's all I want. :)17:50
gsomixhowever I can do smth else.17:50
shogun-buildbotbuild #455 of octave_modular is complete: Failure [failed compile]  Build details are at  blamelist: harshit.syal@nsitonline.in18:00
shogun-buildbotbuild #442 of csharp_modular is complete: Failure [failed compile]  Build details are at  blamelist: harshit.syal@nsitonline.in18:04
-!- n4nd0 [] has joined #shogun18:08
shogun-buildbotbuild #431 of lua_modular is complete: Failure [failed compile]  Build details are at  blamelist: harshit.syal@nsitonline.in18:09
-!- khayyam [~chatzilla@] has joined #shogun18:11
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]18:13
shogun-buildbotbuild #434 of java_modular is complete: Failure [failed compile]  Build details are at  blamelist: harshit.syal@nsitonline.in18:13
harshit_n4nd0: hey , do you know why all these build are are failing to compile ?18:15
harshit_As in does shogun-buildbot compiles every commit that i made earlier18:15
n4nd0harshit_: so shogun-buildbot compiles commit by commit18:15
n4nd0harshit_: exactly, can it  be that some of those commits where not complete?18:16
harshit_I think there were some commits that I made earlier,in which there were some conflicts18:17
n4nd0I can see this error ^18:17
n4nd0make[1]: *** No rule to make target `../../shogun/classifier/svm/NewtonSVM.h', needed by `modshogun_wrap.cxx'.  Stop.18:17
harshit_So what do I need to do now18:17
n4nd0for me it looks like NewtonSVM.h was missing at that moment from the interfaces18:17
n4nd0but I am not sure18:17
n4nd0harshit_: don't worry about it18:17
n4nd0harshit_: it should compile the following commits soon and recover the state18:18
harshit_ohkay got it18:18
shogun-buildbotbuild #430 of python_modular is complete: Failure [failed compile]  Build details are at  blamelist: harshit.syal@nsitonline.in18:18
n4nd0harshit_: you can monitor it a bit if you are interested18:18
n4nd0just if you are curious to know how it works18:19
n4nd0I find it quite cool18:19
n4nd0(I think I like everything that is colorful :P)18:19
harshit_^ dont blame me, :(18:20
harshit_So that was the reason why blackburn said to make big commits rather than small ones18:20
harshit_looks great18:21
-!- naywhaya1e is now known as naywhayare18:21
n4nd0there's another way it seems18:21
n4nd0that sonney2k proposes using git commit -amend18:22
n4nd0I don't know how to use it yet though18:22
harshit_what does -amend do ?18:22
shogun-buildbotbuild #430 of r_modular is complete: Failure [failed compile]  Build details are at  blamelist: harshit.syal@nsitonline.in18:22
n4nd0I have not really gone into it yet, but I think it is used to replace already published-commits18:24
n4nd0take a look to man git commit18:24
harshit_oh great, I think we need to use it more often then18:24
n4nd0and search for --amend18:24
harshit_yeah doing the same18:25
shogun-buildbotbuild #426 of ruby_modular is complete: Failure [failed compile]  Build details are at  blamelist: harshit.syal@nsitonline.in18:27
-!- harshit_ [~harshit@] has quit [Quit: Leaving]18:30
-!- blackburn [~qdrgsm@] has joined #shogun18:34
shogun-buildbotbuild #456 of octave_modular is complete: Success [build successful]  Build details are at
shogun-buildbotbuild #443 of csharp_modular is complete: Success [build successful]  Build details are at
n4nd0great, our friend shogun-buildbot is healthy again18:42
shogun-buildbotbuild #432 of lua_modular is complete: Success [build successful]  Build details are at
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun18:48
shogun-buildbotbuild #435 of java_modular is complete: Success [build successful]  Build details are at
-!- flxb [] has quit [Quit: flxb]18:59
-!- wiking [~wiking@huwico/staff/wiking] has quit [Ping timeout: 276 seconds]19:02
-!- PhilTillet [] has joined #shogun19:03
shogun-buildbotbuild #431 of python_modular is complete: Success [build successful]  Build details are at
n4nd0bah not much19:06
n4nd0what about you?19:06
n4nd0our friendly shogun-buildbot is almost recovered again :)19:07
PhilTilletlol :D19:07
PhilTilletwell not much19:07
PhilTilleti'm tired19:08
PhilTilletand hungry19:08
n4nd0dinner time?19:09
PhilTilletsort of19:09
PhilTilletbut can't have dinner19:09
PhilTilletthe canteen is not opened yet19:10
shogun-buildbotbuild #431 of r_modular is complete: Success [build successful]  Build details are at
-!- harshit_ [~harshit@] has joined #shogun19:15
shogun-buildbotbuild #427 of ruby_modular is complete: Success [build successful]  Build details are at
-!- Juggy [~Razor@] has joined #shogun19:20
JuggyHi blackburn19:21
JuggyCan you give me the link for tutorial on Dimensionality Reduction in shogun?19:21
Juggy@blackburn: are you there ?19:33
-!- harshit_ [~harshit@] has quit [Ping timeout: 252 seconds]19:40
-!- harshit_ [~harshit@] has joined #shogun19:44
-!- PhilTillet [] has quit [Ping timeout: 264 seconds]19:46
@sonney2kharshit_, I am doing some whitespace cleanups and will commit soon - please have a brief look then. I hope this helps to make clear what I had in mind coding style wise19:46
n4nd0sonney2k: I am sorry to insist but, did you have time to take a look at multiclass proposal?19:47
gsomixsonney2k, moin19:50
@sonney2kn4nd0, sorry not yet19:50
@sonney2kgsomix, so how is it going?19:50
n4nd0sonney2k: all right, I am sorry that I rushed19:51
CIA-64shogun: Soeren Sonnenburg master * rcfab94a / (2 files): whitespace and coding style fixed in newtonsvm -
gsomixsonney2k, i'm working on my proposal and python3 interface now.19:54
gsomixsonney2k, are there other tasks for me?19:55
@sonney2kharshit_, please have a look at
-!- puneetgoyal [~puneet@] has joined #shogun19:56
@sonney2kgsomix, most important now is to write your proposal19:57
@sonney2kgsomix, apart from that  it depends a bit what you are most interested in doing20:00
@sonney2kI guess python3 stuff is close to final, so other simple task is to write swig typemap for octave modular20:00
-!- PhilTillet [] has joined #shogun20:00
harshit_sonney2k : sorry having issues with Internet20:01
@sonney2kyeah I've seen...20:01
@sonney2kharshit_, btw can you write some more description about newton svm in the class description and mention the paper there?20:02
@sonney2kharshit_, ahh and before I forget - you didn't mention how shogun's newtonsvm performs on covertype data set vs. matlab version20:03
@sonney2kwould be interesting to know20:03
harshit_sure, i'll do it today20:03
gsomixsonney2k, ok. i will consider this issue later.20:04
PhilTilletsonney2k, what are the "most used" CKernel children? DotKernel I suppose, but any other? :p20:04
harshit_and for comparison there was a problem running the dataset on matlab version, so i compared it with liblinear results on that dataset20:04
harshit_liblinear took about 3-4 sec to train vs my algo which took about 6-8 sec20:05
harshit_and the final weight matrix was almost similar20:05
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun20:05
@sonney2kgsomix, please ask questions / update us - I am not really sure if the tasks you have are too boring for you.20:06
@sonney2kPhilTillet, the most used kernel is the gaussian  kernel20:06
harshit_sonney2k : so after inclusion of documentation in NewtonSVM.cpp, do i need to issue a new pull request ?20:06
-!- PhilTillet_ [] has joined #shogun20:06
@sonney2kharshit_, sure20:06
@sonney2kPhilTillet_, the most used kernel is the gaussian  kernel20:07
@sonney2kharshit_, what was the problem on matlab?20:07
PhilTillet_yes, Actually every DotKernel could be more or less implemented the same way on OpenCL20:07
PhilTillet_but might be different for some other kernel classes20:07
PhilTillet_brb eat :D20:07
harshit_sonney2k: matlab said there is some problem with minres() function when i used that dataset20:08
harshit_but otherwise with other dataset it ran perfectly20:08
-!- PhilTillet [] has quit [Ping timeout: 250 seconds]20:09
@sonney2kharshit_, and if you use a subset of this data set (say first 10000 examples only?)20:09
harshit_I think that dataset is pretty big20:09
harshit_to handle20:09
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Remote host closed the connection]20:09
harshit_oops forgot to do that !20:09
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun20:09
@sonney2kPhilTillet_, dotkernel is most important... / maybe a function to compute whole kernel/distance matrix based on opencl would be nice too20:10
harshit_I'll do it today and report you, for now preparing my gsoc proposal ( i think i am running late on that)20:10
@sonney2kharshit_, yes that is most important true20:10
harshit_sonney2k: after submitting my proposal i want to start working on lbp features implementation .20:11
harshit_do you think opencv's implementation is portable ?20:12
@sonney2kharshit_, we have one in shogun that needs fixing - matlab / c code is avilable20:13
harshit_sonney2k: LBP features is already in shogun ! I thought that needs to be implemented from starting20:15
@sonney2kharshit_, it is in unfinished state20:16
harshit_anyways that is great I can start working on it ~20:16
harshit_please just give me any link to the matlab/c code from which it was ported20:17
@sonney2kharshit_, when it is in you can do some nice demo using opencv and these LBP features - to detect gender of a person based on an image of the face20:17
harshit_sonney2k : actually I have done something like that using deep learning "sparse autoencoders"20:18
harshit_I think it would be easy for me20:18
harshit_great thought20:18
@sonney2kkarlnapf, around?20:20
gsomixsonney2k, boring problem - this is also the problem. :)20:20
karlnapfsonney2k, yes hi20:20
karlnapfwhats up?20:20
@sonney2kkarlnapf, would you mind if I give you github commit rights - I mean you are a core developer without this right for some time now...20:21
@sonney2kyou can still do the same workflow20:21
karlnapfwhy should I mind this? :)20:21
karlnapfShould I have a look at people's requests then?20:21
@sonney2kkarlnapf, I knew that you like to break things20:22
@sonney2kkarlnapf, if you have time - that's welcome20:22
karlnapfsure, of course then20:22
karlnapfI will still ask you guys to take a look on stuff which involves larger changes20:22
karlnapfsonney2k, I am just wondering what is our favorite way of doing intending in python, spaces or tabs?20:24
blackburnJuggy: still there?20:24
blackburnsonney2k: finally!20:25
blackburnJuggy: probably that one, it is unaccurate though20:25
blackburnPhilTillet_: btw I have something really big to propose20:28
blackburnI have some collection of dimensionality reduction algorithms there.. ;)20:29
@sonney2kkarlnapf, - welcome to the team but now officially :D20:29
karlnapfnice one :D20:29
blackburnsonney2k: what a long way home20:29
-!- gsomix [~gsomix@] has quit [Read error: Operation timed out]20:29
@sonney2kkarlnapf, well *I* do tabs as usual20:29
blackburnI prefer tabs as well btw20:30
@sonney2kblackburn, sing some russian love song for heiko please - to celebrate this event!20:30
karlnapfyeaaah :)20:30
blackburnsonney2k: love song? like 'it is ok to be gay'?20:30
* sonney2k turns the volume to the max!20:30
karlnapfblackburn, gogo20:30
@sonney2kblackburn, no sth. like ~~~oh~~~heiko~~~in~~~your~~~code~~~ohh~~~we~~~trust~~~20:31
blackburn??????????? ???? ??????? ???? ??????20:31
@sonney2kkarlnapf, regarding spaces - it might be that python3 wants spaces ... not sure20:31
blackburn? ?????????? ? ????????? ???? ?????20:31
karlnapfsing louder blackburn, cant hear you20:31
blackburn???? ??????? ?? ???? ????????20:31
blackburn?? ??????????? ????????????????????????!20:31
karlnapfsonney2k, ok, but we can convert, just asking20:32
blackburnkind of old love song hahaha20:32
blackburnjust like sonney2k asked20:32
karlnapfI love these letters20:32
karlnapfwant to learn Russian at some point20:32
blackburnfunny it is a war song probably hahah20:32
blackburnWW2 yes20:32
karlnapfIll double check with google translate20:32
blackburnkarlnapf: would be pretty senseless20:32
karlnapfGr?ne Ahorn Riegelahorn, aber ;)20:33
karlnapfbut I can heart it, the friendly google woman sings it20:33
blackburnkarlnapf: this one is better hah20:34
-!- sidsniper [~piggy@] has quit [Ping timeout: 250 seconds]20:34
blackburnkarlnapf: why do you want to learn russian?20:35
karlnapfWant to learn another language which is not romanian20:35
karlnapfand I want to travel to the place20:35
blackburnkarlnapf: you are very welcome btw ;)20:36
-!- Juggy [~Razor@] has quit [Ping timeout: 260 seconds]20:36
karlnapfthanks blackburn20:36
karlnapfits very nice to be in your team guys20:36
blackburnkarlnapf: seriously I would like someone of you to visit me here20:36
karlnapfI was also serious, really wanna go there, so perhaps next year or so, well meet20:37
blackburnsonney2k: let you organize shogun code sprint at berlin? ;)20:37
karlnapfalso a nice idea20:38
-!- gsomix [~gsomix@] has joined #shogun20:38
@sonney2kat my flat with my kids jumping around - why not20:40
blackburnhow big is your flat? ;)20:40
@sonney2kno but seriously we could do that from gsoc money20:40
karlnapfhehe, well I know many people to stay at20:40
karlnapfyes would be kind of cool to meet20:41
karlnapfblackburn, how is your german? :)20:41
blackburnkarlnapf: ~ to your russian ;)20:41
blackburnich bin mude20:42
@sonney2kich bin immer muede20:42
blackburnkarlnapf: wo bist du?20:42
karlnapfich bin in england20:42
blackburnsonney2k: ??? ?????20:42
karlnapfich bin sehr m?de20:42
blackburnyes I know umlaut is there20:43
blackburntoo lazy to type it20:43
karlnapfhow to type it on russian keyboard?20:44
blackburnkarlnapf: do I remember that correctly - u mit dem umlaut = ?20:44
blackburnusing compose key20:44
blackburnu+right alt+"20:44
blackburnkarlnapf: I had been studying german up to 7th grade at school20:45
@sonney2kanother very strong proposal got submitted... many are from phd students this time20:45
blackburnfor two years probably20:45
blackburnsonney2k: w/o contributions?20:45
karlnapfah really, thats more than my russian20:45
karlnapfwhich project?20:45
gsomix>> phd students *__*20:45
karlnapfyes, phd all over the place, gsoc is very academic20:46
blackburnmy belief is proposal w/o contributions would not work20:46
@sonney2kkarlnapf, last year that was not the case20:47
karlnapfsonney2k, probably means that stuff is more advanced, or they checked last years projects or so20:47
karlnapfI mean if you dont do machine learning the stuff is pretty tough20:48
karlnapfeven if you do its not trivial20:48
@sonney2kmany things this year are really tough20:48
@sonney2kfor example for SO we didn't have any proposal last time20:48
@sonney2kthis year quite a bit20:48
blackburnthis year we have mighty n4nd020:49
blackburnstructured ouput20:49
karlnapfah ok20:49
karlnapfyes, cool thing is that20:49
blackburnn4nd0: btw checking your spe right now20:50
@sonney2kanyway we have 9 mentors so max slots is 920:52
karlnapflets hope to get all that !20:52
blackburnpretty impossible probably20:52
@sonney2kkarlnapf, actually in retrospect I am happy that we got just 5 slots last year20:53
@sonney2kit really was a lot of work20:53
karlnapfsonney2k, ok I believe that20:53
@sonney2kthis year though we have n4nd0, blackburn, karlnapf and others helping others20:53
@sonney2kit is not soo much all on my shoulders any more20:53
karlnapfyes, you did almost everythign last year20:53
@sonney2kso maybe this process can scale20:53
@sonney2kand actually students this year all submitted github pull requests before20:54
@sonney2kwith pretty nice work that goes deep into shogun20:54
@sonney2klast year we were switching from svn -> git in the process20:55
karlnapfthis year we got the buildbots20:55
@sonney2kwe were doing conversions like double* -> sgvector20:55
@sonney2kwe created buildbots20:55
-!- flxb [] has joined #shogun20:55
@sonney2klots of infrastructure that is now there and helps to keep the project in a sane state20:55
@sonney2kto conclude - we should be realistically able to handle 9 students20:56
blackburnsonney2k: what was there before gsoc 2011?20:56
karlnapfsonney2k, wow that would be amazing.20:56
karlnapfdo you know how google allocates slots?20:56
karlnapfonly in terms of application #20:56
karlnapfor also other stuff?20:56
blackburnkarlnapf: no, there are some criteria20:58
blackburnI'll refer in a min20:58
-!- Juggy [~Razor@] has joined #shogun20:59
-!- sidsniper [~piggy@] has joined #shogun20:59
n4nd0blackburn: cool man, I was having dinner, tell me if I should do / say sth21:01
blackburnkarlnapf: to be fair we fit 1), 2), 4)21:01
karlnapfsonney2k, just trying out a request21:01
karlnapfdoes the workflow change?21:01
karlnapfdo I still make pull requests and then merge them myself?21:01
blackburnn4nd0: I like it pretty much21:01
blackburnkarlnapf: I think this type of commits would be better to be direct21:02
karlnapfok so only pull request for larger stuff21:02
karlnapfand direct means I clone the actual shogun repo?21:03
blackburnkarlnapf: yes21:03
blackburnkarlnapf: ah yes21:03
blackburnyou need to push to21:03
karlnapfok then21:03
blackburngoogle, shogun and github21:04
@sonney2kkarlnapf, btw you can from now on just do (C) Heiko S... no need to mention TU or so any longer21:04
blackburnsonney2k: btw21:04
blackburnI had idea before21:04
blackburnWritten and Copyright now are pretty same21:04
@sonney2kblackburn, karlnapf I disagree21:04
blackburnsonney2k: about?21:04
@sonney2kI am very much favoring pull requests21:05
@sonney2kand then have someone else looking at it21:05
blackburnsonney2k: for small fixes?21:05
@sonney2kbut yeah I know blackburn you disagree21:05
@sonney2kblackburn, for small fixes pull request + self merge21:05
blackburnthis takes more time..21:05
karlnapfok like examples and stuff, but its always good to get stuff reviewed when its more complicated21:06
@sonney2kand avoids quite a bit of stupid errors21:06
blackburnsonney2k: these stupid errors emerge pretty rarely..21:06
karlnapfblackburn, I have a talent of finding stuff like this ;)21:07
blackburnI still don't like this kind of self-merge21:07
CIA-64shogun: Heiko Strathmann master * rf5c30fd / examples/undocumented/python_modular/ : added example for modelselection of different kernels -
CIA-64shogun: Heiko Strathmann master * r67c3782 / examples/undocumented/python_modular/ : fixed copyright and excluded TUB -
CIA-64shogun: Heiko Strathmann master * r2801f8d / examples/undocumented/python_modular/ :21:07
CIA-64shogun: Merge pull request #417 from karlnapf/master21:07
CIA-64shogun: new example -
blackburnkarlnapf: sonney2k: one more thing I would like to suggest21:08
blackburnis to use #XXX stuff21:08
blackburnfor example if you are pretty ready to merge PR no need to wait21:08
blackburnyou may commit with 'blabla Closes #XXX'21:09
blackburnand actually we can use not only PRs but issues as well21:09
blackburnwe can commit with #XXX to incorporate commit to this issue21:10
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds]21:10
blackburnI believe it makes more sense to group commits with issues21:10
@sonney2k...if we used issues21:10
blackburnlet me show how it works21:10
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun21:11
shogun-buildbotbuild #198 of nightly_none is complete: Failure [failed compile]  Build details are at
blackburnhere we would have docu improvements21:11
n4nd0karlnapf: about the answer to the issue, can you elaborate a bit more on that? what is the idea then? not to serialize SGNArray stuff?21:13
n4nd0karlnapf: I think that might be a problem because for example QDA has some members of that type21:13
karlnapfn4nd0, no that was just a general thought21:13
karlnapfthis is possible (and very nice I think)21:13
karlnapfbut its more complicated than on the first impression21:13
blackburnand once I add new commit21:14
n4nd0karlnapf: I see, I have no idea how serialization stuff works in shogun21:14
karlnapfif you dive into the load_serializable method of CSGObject you will that there are many places where these types have to be handled21:14
karlnapfn4nd0, is it important that you need that?21:15
CIA-64shogun: Sergey Lisitsyn master * rb1b17cb / src/shogun/converter/DiffusionMaps.h : Added note about lapack performance with diffusion maps #418 -
blackburnit appears in #41821:15
n4nd0karlnapf: you mean if I need to use that type in QDA?21:15
blackburnsonney2k: karlnapf: what do you think?21:15
karlnapfn4nd0, yes21:16
n4nd0karlnapf: at the beginning I was simply using SGMatrix *21:16
n4nd0karlnapf: the idea is that I need a list of matrices21:16
karlnapfblackburn, I think thats nice, but is it easy to handle?21:16
n4nd0karlnapf: but that SGMatrix* turned out to be a bit problematic also with this issue, the serialization21:17
blackburnkarlnapf: yes all you need is to recall # of issue21:17
karlnapfn4nd0, i see21:17
n4nd0karlnapf: so sonney2k proposed to use SGNDArray instead21:17
@sonney2kblackburn, yeah I like that - we had this same feature with trac before but also there rarely used it - just because we never really managed to use issues all the time21:17
blackburnsonney2k: that would be faster for doc fixes and other small fixes21:17
blackburntoo much pull requests would slow down things I think21:18
karlnapfn4nd0, ok, probably good idea then21:18
@sonney2kblackburn, I am saying lets use it but I don't know how who will file issues all the time21:18
@sonney2knot sure if we are enough developers yet for that stuff21:18
karlnapfn4nd0, I just hate to touch the migration stuff ;), but should be doable, if you want, take a look21:18
n4nd0karlnapf: so the thing is that right now serialization in QDA wouldn't work right?21:18
karlnapfn4nd0, yes21:19
blackburnsonney2k: hmm just write somewhere (doc fixes goes to #418)21:19
karlnapfnot too bad now, but should be done on the long run21:19
n4nd0karlnapf: I will write that in the doc. at least21:19
blackburnwould be nice to make some other keyword referencing21:19
n4nd0karlnapf: yeah, I agree, sooner or later it must be done21:19
@sonney2kblackburn, can I close an issue from via a commit message?21:19
karlnapfn4nd0, yes, perhaps during the summer :)21:20
blackburnsonney2k: closes #41821:20
blackburnsonney2k: just like pull request21:20
blackburnbut no need to switch between fork and origin21:20
blackburnsonney2k: fixes, fixed, fix, closes, close, closed21:21
blackburnand of these words21:21
blackburnok have to clean up my damned traffic sign recognition paper be back in a hour or so21:24
gsomixsonney2k, at now. my last fixes and 2to3.21:24
n4nd0blackburn: see you later then, maybe we can finish with SPE then?21:24
blackburnn4nd0: hmm one more thing21:24
blackburnplease add reference21:25
blackburnand I'll check it once again this night and merge21:25
n4nd0blackburn: for the doc?21:25
n4nd0blackburn: ok21:25
-!- khayyam [~chatzilla@] has quit [Quit: ChatZilla [Firefox 11.0/20120312181643]]21:25
n4nd0blackburn: I am a bad guy ... I am always like saying TODO doc :P21:25
blackburnheh better have todo than forgot21:25
@sonney2kgsomix, you need to install liblzma-dev21:26
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]21:26
@sonney2kgsomix, and libjson0-dev21:26
gsomixsonney2k, done21:31
sidsniperblackburn can you give me the link to the tutorial on dim red?21:36
n4nd0sidsniper: I think blackburn is out now working on his stuff21:37
n4nd0sidsniper: what tutorial do you want?21:37
sidsniperthe one that is in shogun21:38
sidsnipertutorial showing how to use those algos21:38
n4nd0ok, have you taken a look to the examples?21:38
n4nd0maybe that is enough for your needs?21:38
gsomixsonney2k, but doesn't work. hmm21:40
gsomixon python3.21:40
gsomixand on python2.21:41
-!- gsomix [~gsomix@] has quit [Quit: ????? ? ?? ??? (xchat 2.4.5 ??? ??????)]21:41
-!- Afan [] has joined #shogun21:48
PhilTillet_Hello again21:48
-!- PhilTillet_ is now known as PhilTillet21:51
-!- gsomix [~gsomix@] has joined #shogun21:52
gsomixsonney2k, works.21:53
-!- uricamic [5546c513@gateway/web/freenode/ip.] has joined #shogun21:53
Juggyis there who knows about the DR tutorial in Shogun?22:01
JuggyI am not able to fine :(22:01
Juggyis there a list of implemented algorithms ?????22:02
n4nd0you can check them under the directory converter22:03
n4nd0there you can a find a list in doxygen doc22:04
n4nd0but doc might not be up to date22:05
@sonney2kn4nd0, strong MC proposal...22:07
@sonney2kyours I mean22:08
n4nd0sonney2k: oh thank you :)22:08
-!- PhilTillet_ [] has joined #shogun22:08
-!- PhilTillet [] has quit [Remote host closed the connection]22:08
-!- gsomix [~gsomix@] has quit [Quit: ????? ? ?? ??? (xchat 2.4.5 ??? ??????)]22:21
PhilTillet_sonney2k, do you have any idea of what "big idea" blackburn was talking about?22:38
@sonney2kbig idea?22:47
@sonney2kI guess I missed sth22:47
@sonney2kexcept he is dreaming of world domination again ;-)22:47
PhilTillet_<blackburn> PhilTillet_: btw I have something really big to propose22:49
PhilTillet_<blackburn> I have some collection of dimensionality reduction algorithms there.. ;)22:49
@sonney2kmany dim reduction methods are expensive so it sounds like he wants some opencl magic :)22:51
@sonney2kPhilTillet_, btw how are you progressing with your patch?22:51
PhilTillet_They are all about eigenvalues right?22:51
PhilTillet_Well, pretty hard22:51
PhilTillet_I cleaned my source22:52
PhilTillet_so that there is a clear OpenCL structure22:52
PhilTillet_to compile programs etc22:52
PhilTillet_and now I copy with direct memcopy, without using ublas22:52
PhilTillet_I mean, for now my patch is mostly the opencl code22:54
PhilTillet_and basically a shogun::ocl::svm::init()22:54
PhilTillet_which compiles all the opencl code22:54
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]23:03
PhilTillet_memory management is one of the first thing in my GSoC Timeline, but won't be able to do it for the patch23:04
-!- wiking [] has joined #shogun23:18
-!- wiking [] has quit [Changing host]23:18
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun23:18
-!- Marty28 [~Marty@] has quit [Quit: ChatZilla [Firefox 11.0/20120310010446]]23:18
-!- karlnapf [] has quit [Quit: Leaving.]23:22
PhilTillet_sonney2k, hope I'll be able to submit an opencl patch (not perfect from a design point of view XD) soon23:28
CIA-64shogun: Soeren Sonnenburg master * r21ae0ec / src/shogun/classifier/svm/NewtonSVM.h : add ifdef HAVE_LAPACK for NewtonSVM -
-!- puneetgoyal [~puneet@] has quit [Ping timeout: 246 seconds]23:32
PhilTillet_sonney2k, I'll have some design question though :p23:36
-!- uricamic [5546c513@gateway/web/freenode/ip.] has quit [Quit: Page closed]23:51
-!- PhilTillet_ [] has quit [Quit: Leaving]23:57
-!- harshit_ [~harshit@] has quit [Ping timeout: 244 seconds]23:57
--- Log closed Wed Apr 04 00:00:19 2012