Open in new window / Try shogun cloud
--- Log opened Fri Apr 13 00:00:19 2012
-!- av3ngr [av3ngr@nat/redhat/x-cvkzvkolelblsktc] has joined #shogun01:47
-!- blackburn [~qdrgsm@] has quit [Ping timeout: 252 seconds]02:01
-!- n4nd0 [] has quit [Quit: leaving]02:37
-!- vikram360 [~vikram360@] has quit [Ping timeout: 260 seconds]03:01
-!- vikram360 [~vikram360@] has joined #shogun04:45
-!- vikram360 [~vikram360@] has quit [Ping timeout: 252 seconds]04:53
-!- vikram360 [~vikram360@] has joined #shogun04:53
-!- siddharth_ [~siddharth@] has joined #shogun05:18
-!- siddharth_ [~siddharth@] has left #shogun []05:18
-!- PhilTillet [] has quit [Ping timeout: 272 seconds]07:35
-!- n4nd0 [] has joined #shogun07:47
-!- av3ngr [av3ngr@nat/redhat/x-cvkzvkolelblsktc] has quit [Ping timeout: 240 seconds]08:25
-!- av3ngr [av3ngr@nat/redhat/x-xtwvnksjinqwtsqn] has joined #shogun08:36
sonne|workhi n4nd0 - a few warnings are on your account:
n4nd0sonne|work: ok, I will take of them, thank you!08:42
-!- av3ngr [av3ngr@nat/redhat/x-xtwvnksjinqwtsqn] has quit [Quit: That's all folks!]08:43
sonne|workn4nd0: btw are you still happily covertree-ing?08:44
-!- xiangwang [~chatzilla@] has joined #shogun08:49
n4nd0sonne|work: yeah08:56
n4nd0sonne|work: right now I am stuck with some university work, but I during my shogun time I am covertree-ing :)08:57
n4nd0sonne|work: why?08:57
sonne|workn4nd0: I got used to hearing break through's from you side :)08:57
n4nd0sonne|work: :D I am trying to control myself more now08:58
sonne|worktoo bad08:58
n4nd0I discover yesterday though an interesting thing about JL covertree construction in my integration within shogun08:59
n4nd0sonne|work: I can show you now08:59
n4nd0sonne|work: the context is in CKNN:apply09:02
n4nd0sonne|work: here distance is set so distance(train_features, query)09:02
n4nd0sonne|work: but I need to build two cover trees before being able to do the queries09:03
n4nd0sonne|work: one with the train_features, and another with the queries09:03
n4nd0sonne|work: to build these trees I need distances, between those features09:03
n4nd0sonne|work: i.e to build the cover tree with train_features I need that distance is distance(train_features, train_features)09:03
n4nd0sonne|work: and to build the one for the queties: distance(queries, queries)09:04
n4nd0sonne|work: I have managed to do that using this trick
n4nd0sonne|work: I think it is a bit ugly, but it does work!09:04
sonne|workyeah for that one would need what you have said09:05
n4nd0sonne|work: the problem is that it turns out that the calls to replace_[rhs|lhs] take more time than the creation of the cover trees09:05
sonne|worka way to compute distance for rhs,rhs stuff09:05
sonne|workwhich distance?09:06
n4nd0sonne|work: what do you mean?09:06
sonne|workfor which distance does it take more tiem?09:06
n4nd0I don't think it is really the distance09:07
n4nd0but the calls to replace_lhs and replace_rhs09:07
n4nd0in JL's code, both calls (for set_of_points and set_of_queries in the code) to batch_create09:07
n4nd0take less time than in my shogun code, using the replace methods in between09:08
n4nd0within my shogun code the portion of code I have pasted for you takes 0.2482(s)09:09
n4nd0while for JL's, the two calls to batch create take 0.053942(s)09:10
n4nd0oh wait09:10
n4nd0this may be caused as well for the upper_bound distance thing I told you about!09:11
n4nd0sonne|work: haha I think I did my first breakthrough of the day :D09:11
sonne|workI don't understand a thing09:12
sonne|workyou mean upper_bound stuff is the reason why JL's thing is so fast09:12
sonne|workbut you don't use it?09:13
sonne|workin shogun09:13
n4nd0I have not implemented it yet09:13
sonne|workand that is why it is so slow?09:13
n4nd0I think that the reason why I see that the construction of the trees is slow in shogun09:13
n4nd0maybe is because of the thing of the upper_bound distance09:13
n4nd0do you understand?09:14
n4nd0I am not that good at explaining sometimes :S09:14
sonne|workyes OK09:15
sonne|workbut it is already faster right?09:15
sonne|work4 times I mean?09:15
sonne|workso you can speed it up quite a bit with this upper_bound stuff09:16
n4nd0right now it is already faster than the thing we had before09:16
n4nd0not as fast as standalone JL though09:16
n4nd0it might approach more standalanoe JL doing the upper_bound thing09:17
-!- wiking [] has joined #shogun09:21
-!- wiking [] has quit [Changing host]09:21
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun09:21
sonne|workn4nd0: now I wish you had time to try this out09:22
n4nd0sonne|work: the cover tree?09:22
sonne|workthe upper bound stuff w/i shogun09:26
n4nd0ah I will get some time to work on it today, I'm almost done with the test I wanted to do for my school project09:28
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]09:39
sonne|worklet me know :)09:42
-!- blackburn [~qdrgsm@] has joined #shogun09:43
blackburnn4nd0: hey there I heard some success with covertree?09:47
n4nd0blackburn: yeah, I managed to make it work from shogun09:47
n4nd0blackburn: is still work in progress though09:48
-!- wiking [] has joined #shogun09:49
-!- wiking [] has quit [Changing host]09:49
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun09:49
-!- blackburn [~qdrgsm@] has quit [Read error: Connection reset by peer]09:55
-!- pluskid [~chatzilla@] has joined #shogun09:56
-!- blackburn [~qdrgsm@] has joined #shogun09:56
sonne|workwiking: yo10:01
wikingfriday the 13th had it's course already on me :)10:01
sonne|workwiking - yeah you missed out 2 lines with warning fixes
sonne|workor was it sth else ?\10:02
wikinghahaha sure something else... internet got broken, warm water in the house suddenly stopped working10:02
wiking2? i just see one preprocessor/PNorm.h:37:10:04
wikingpull request for fixing that error sent out10:07
CIA-64shogun: Viktor Gal master * r92348cf / src/shogun/preprocessor/PNorm.h : Fix for PNorm constructor doxygen comment -
CIA-64shogun: Soeren Sonnenburg master * r6c0a969 / src/shogun/preprocessor/PNorm.h :10:09
CIA-64shogun: Merge pull request #447 from vigsterkr/master10:09
CIA-64shogun: yet another fix for PNorm doxygen comment -
sonne|workwiking: this one preprocessor/PNorm.cpp:120: warning: 'norm' may be used uninitialized in this function10:09
-!- pluskid_ [] has joined #shogun10:09
sonne|workactually more critical10:09
wikinghahahahah THE FIST line and i don't see it :)10:10
-!- pluskid [~chatzilla@] has quit [Ping timeout: 260 seconds]10:11
-!- pluskid_ is now known as pluskid10:11
wikingactually that's a good question what it should be... afaik the best is to return 1.0 rather than 0.010:11
wikingalthough that's not 'correct' but it won't hurt as much as if i would return 0.0 and then the caller right away uses the return value without checking to divide a number => divide by zero10:12
sonne|workwiking: huh?10:19
sonne|workyou should init it with 010:19
* sonne|work thinks wiking didn't have enough sleep10:19
wikingand if that gets returned...?10:19
sonne|workit doesn't10:19
wikingit might...10:19
sonne|workin all if () { } sections norm is computed10:20
sonne|workonly in the first one it is incremented10:20
sonne|workthat is why you need the init to 010:20
wikingwell it still can happen that 0 gets back10:21
wikingbut that's the caller's problem then...10:21
sonne|workand an independent problem10:21
sonne|workno matter what you set norm to10:21
blackburnTHE FIST10:22
wikingpull req sent10:23
blackburnhey guys have you heard how nice is DPRK's rocket?10:23
pluskidrocket ?10:24
blackburn totally broken lol10:24
blackburnwiking: should I merge your stuff? :)10:24
wikingblackburn: first one is already merged :)10:25
wikingbut yeah would be great10:25
wikingnow it's time to go back and try to speed up gmm10:26
-!- vikram360 [~vikram360@] has quit [Ping timeout: 246 seconds]10:27
CIA-64shogun: Viktor Gal master * r7f6d2d2 / src/shogun/preprocessor/PNorm.cpp : PNorm: initialise the norm to 0.0 in the beginning -
CIA-64shogun: Sergey Lisitsyn master * r1f77383 / src/shogun/preprocessor/PNorm.cpp : Merge branch 'master' of git:// -
blackburnwiking: I glanced over your twitter and I admit I do not understand anything10:28
wikingsome of it is on hungarian :)10:28
blackburnyes I meant that ;)10:28
blackburnwhat is csinalok?10:29
wikingactually i'm doing10:29
wikingcsinalni = to do10:29
wikingcsinalok = i'm doing10:29
wikingor i do10:29
n4nd0use a lot twitter guys, I am making a project to rank users and stuff!10:29
wikingdepends on the context10:29
blackburnn4nd0: same with your twitter btw10:29
blackburnwiking: there is no difference between I'm doing/to do in russian :D10:30
n4nd0blackburn: like in Swedish, no continuous forms!10:30
blackburnhowever you probably know that10:30
blackburnthat is one of the reasons why we fail with speaking english properly sometimes10:30
n4nd0blackburn: do you have articles in Russian?10:32
n4nd0blackburn: like "the house"10:32
n4nd0the thing I have noticed the most in Russian speakers is that10:32
blackburnthat's why I (and other russians) mix it up as well hah10:32
blackburnhowever I get used with it a little10:33
blackburnearlier it was worse10:33
n4nd0I don't think you do that one, at least I didn't notice10:33
blackburnno, I still do a lot of mistakes and crazy sentences10:34
blackburnbut year ago it was really worse hah, right sonne|work?10:34
blackburnn4nd0: we also have cases (dative,genitive,etc) there10:36
n4nd0blackburn: that must be difficult10:36
blackburnnot while you are native with it :D10:36
blackburnfor example10:36
blackburnthere is a fish10:36
blackburnthere is no fish10:37
blackburnwould be10:37
blackburnah bad example10:37
blackburnword with same cases :D10:37
blackburnI mean word is changing in different cases10:37
n4nd0the fish is red10:37
n4nd0I like fish10:37
blackburn???? ???????10:37
blackburn? ????? ????10:37
n4nd0when it is subject or DC10:38
n4nd0I actually know to say fish in Russian10:38
blackburnryyba? ;)10:38
blackburnn4nd0: aren't cases there in spanish?10:39
wikingblackburn:  ;) just saying ... usually indo-european language based people cannot grasp the idea of having only 3 tenses: past, present and future ;p10:39
n4nd0blackburn: no, no cases10:39
-!- uricamic [9320543b@gateway/web/freenode/ip.] has joined #shogun10:41
blackburnwiking: isn't english indo-european? :)10:41
blackburnah only10:42
wikingyes it is and it has about 8 tenses :)10:42
blackburnyes got you10:42
blackburnuricamic: can it be that you do not precompute kernel matrix in your gs krr?10:44
blackburnin case it is too large10:44
blackburnokay have to go guys10:44
blackburnwiking: how much tenses there are in hungarian then?10:45
wikingas said 310:45
uricamicblackburn: yes, I can change that10:45
n4nd0blackburn, wiking the tenses in Spanish are not normal ... too many10:46
wikingn4nd0: it's latin based language it's normal10:46
blackburnok see you10:47
-!- wiking [~wiking@huwico/staff/wiking] has quit [Remote host closed the connection]10:47
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]10:47
-!- wiking [] has joined #shogun10:47
-!- wiking [] has quit [Changing host]10:47
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun10:47
n4nd0too many anyway :)10:47
pluskidonly 1 tense in Chinese ! :D10:47
pluskidI guess10:47
n4nd0you don't have difference between past / present / future at least?10:47
pluskidwe do not change form of a word10:48
pluskidbut add more words to indicate the tense10:48
pluskidso there's no difference between "eat" and "eating"10:48
pluskidbut the meaning depends on the context10:49
n4nd0not even eat and ate?10:49
pluskidthere will be words in the context to indicate whether "eat" or "eating"10:49
pluskidbut there are multiple tenses in Japanese10:50
pluskidthough Chinese and Japanese looks very similar, they are quite different actually10:50
pluskidbtw, shogun (the japanese word) is "jiang jun" in Chinese :p10:51
pluskidthe same characters, different pronounciations10:51
pluskidand the same meaning10:52
sonne|workuricamic: can you do any kind of benchmark for KRR?10:52
sonne|workuricamic: and one more thing - could you create a *separate* pull request for the GS-KRR?10:53
uricamicsonne|work: I plan to implement also the version which will not precompute the kernel matrix for large data. What kind of benchmark do you mean?10:53
sonne|workuricamic: just measure if GS is faster than the pinv based one10:54
sonne|workbtw I would do it slightly differently10:54
uricamicsonne|work: I have tried to create another pull request, but it didn't work10:54
sonne|workgsomix, n4nd0, wiking, pluskid - please help uricamic with this :)10:55
pluskiduricamic: are you in the mailing list?10:55
uricamicis it possible that I can have only one opened pull request?10:55
uricamicpluskid: yes10:55
pluskidthere was an email describing the work flow10:55
sonne|workuricamic: I would just add a new train function to KRR10:55
shogun-buildbotbuild #651 of octave_static is complete: Failure [failed test_1]  Build details are at  blamelist: sonne@debian.org10:56
sonne|workand the epsilon parameter10:56
wiking!!!! sonne|work :>10:56
sonne|worknot create a new class for that - I mean the rest is the same :)10:56
pluskidI think you have to create multiple branches to make multiple pull request10:56
sonne|workwiking: I promise I didn't do anything!10:56
wikingwhy did it fail now? :)10:56
sonne|worksomething wrong with perceptron10:57
uricamicsonne|work: ok, that sounds reasonable, I will change it10:57
sonne|worka rhyme :)10:57
sonne|workuricamic: yeah then one can have a 'master' train function that selects some selected solver that one can specify as arg to constructor10:57
sonne|workwiking: I don't even see any error?!10:58
uricamicpluskid: what was the subject of that email ?10:58
n4nd0uricamic: yeah I agree with pluskid, you need multiple branches for multiple pull requests10:59
pluskiduricamic: git workflow10:59
pluskiduricamic: this link:
uricamicpluskid: thanks11:00
pluskidu r wel~ ^_^11:00
* sonne|work is happy to see some interaction here :)11:02
-!- vikram360 [~vikram360@] has joined #shogun11:22
n4nd0sonne|work: I got news11:28
* sonne|work listens up11:31
sonne|workn4nd0: what news?11:31
-!- gsomix [~gsomix@] has quit [Ping timeout: 264 seconds]11:31
n4nd0sonne|work: so doing this upper bounded computation11:32
n4nd0the time improves, but just a little bit11:32
n4nd0before the construction time in shogun was like 5 times the one in standalone JL11:33
* pluskid is trying some IRC command11:33
n4nd0now it is sth between 4 and 5 times11:33
* n4nd0 likes this IRC command :)11:33
* pluskid thinks it cool, too11:34
n4nd0the thing is that the distance function used in standalone JL is pretty optimized11:35
n4nd0take a look11:35
sonne|workahh ok11:35
n4nd0the loop is unrolled I think11:36
n4nd0the big issue is that it assumes that the dimension of the points is multiple of 811:36
sonne|workand it is based on floats11:37
-!- pluskid [] has quit [Quit: ChatZilla [Firefox 11.0/20120312181643]]11:37
n4nd0does that make a big difference too?11:37
sonne|workfactor 2 at least11:37
n4nd0so what do you think?11:37
sonne|workwe don't want to switch to float11:44
sonne|workbut all other speedups are welcome :)11:44
n4nd0so since that implementation assumes that the dimension of the features is multiple of 811:46
n4nd0(if the points given are not, there is a function that pads them with zeroes)11:47
n4nd0how would you handle that in shogun?11:47
xiangwanghi, does shogun include the cutting-plane algorithms?11:47
n4nd0xiangwang: hey! we have OCAS :)11:47
xiangwangin which folder of classifier ? thanks11:48
n4nd0classifier/svm I guess11:48
n4nd0yeah I see it in classifier/svm/SVMOcas11:49
n4nd0sonne|work: do you think it pays off to reallocate CFeatures* lhs and rhs to make them multiple of 8?11:50
n4nd0could it be done in another way?11:50
sonne|workn4nd0: well one could use this algorithm for the multiple of 8 dims that fit and for the remainder the slow one11:58
n4nd0sonne|work: I like that solution11:59
n4nd0sonne|work: we document it, that is recommended to use covertree with multiples of 8 dimensions :)11:59
sonne|workI mean sth else12:00
sonne|workone can do fast processing as long as there are blocks of 8 dims to process12:00
sonne|workonly for the remainder we use the slow method12:00
sonne|workthat is fast too so no problem12:00
-!- elango_ [] has joined #shogun12:01
n4nd0also, this can only be done for SimpleFeatures, can't be?12:01
sonne|workso I think this should be a function in dot/simplefeatures then12:03
sonne|workbut hey covertree first then this speedup stuff :)12:03
n4nd0all right12:04
n4nd0avout the warnings before12:06
n4nd0ups aBout12:06
n4nd0they're related with the other covertree so when I do a pull request with the new one12:06
n4nd0that code disappears12:07
n4nd0I mean, no need to fix them before right?12:07
sonne|worklooking forward to it then12:07
shogun-buildbotbuild #652 of octave_static is complete: Success [build successful]  Build details are at
-!- vikram360 [~vikram360@] has quit [Ping timeout: 276 seconds]12:15
-!- wiking_ [~wiking@huwico/staff/wiking] has joined #shogun12:17
-!- wiking [~wiking@huwico/staff/wiking] has quit [Read error: Operation timed out]12:19
-!- wiking_ is now known as wiking12:19
-!- vikram360 [~vikram360@] has joined #shogun12:28
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has joined #shogun12:43
blackburnxiangwang: MulticlassOCAS and SVMOcas are cutting-plane based (libocas)12:44
-!- elango_ [] has left #shogun ["Leaving"]12:47
blackburnsonne|work: hey now I understand you a little more with 10 hrs of sleeping for last 3 days :D12:48
blackburnn4nd0: my suggestion is to add a flag/option for upper_bounded thing12:53
n4nd0blackburn: if SimpleFeatures are provided then do the special one?12:55
blackburnn4nd0: why features?12:55
n4nd0blackburn: I think this thing can only be done for SimpleFeatures12:57
blackburnn4nd0: I'd it depends on distance rather than features12:57
n4nd0but don't you think that the function I have posted can just be done if we have SimpleFeatures?12:58
n4nd0we can upper bound for all ok (as long as Euclidian is used this far)12:58
n4nd0but to do it with that unrolled loop and the multiples of 812:59
blackburnI mean some distances could provide that13:00
blackburnand could not13:00
blackburnsonne|work: can all these fancy string kernels/distances be upper bounded?13:00
blackburnn4nd0: however you said it is not as fast as we expected with this upper bound?13:02
n4nd0sorry I was afk13:08
n4nd0blackburn: it is expected because it is not just the upper bound what makes JL faster but also the optimizations13:08
blackburnn4nd0: loop unrollment?13:08
n4nd0blackburn: probably13:09
n4nd0sonne|work pointed out that the floats also, we use double13:09
blackburnn4nd0: loop unrollment is dangerous thing to implement explicitly I think13:09
blackburnwe can't be sure it would be faster everywhere13:09
blackburnand actually blas is unrolled so we should stick to blas I think13:10
n4nd0but there we cannot cut the computation if the upper bound is passed right?13:10
blackburnwe can compute batches with blas13:11
blackburncould you please try that?13:11
blackburnor SNRM213:11
blackburnn4nd0: it is pretty interesting how did they implement that
n4nd0blackburn: why don't you like just to implement that piece of code?13:13
n4nd0I mean the distance used in JL13:13
blackburnn4nd0: implement where?13:14
n4nd0to optimize the distance computation for cover tree13:14
blackburncheck if it is faster than dnrm2 ;)13:14
n4nd0how can dnrm2 be faster?13:15
n4nd0don't you think this one is already very specific?13:15
blackburnin which means?13:17
-!- wiking [~wiking@huwico/staff/wiking] has quit [Remote host closed the connection]13:18
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun13:18
n4nd0it looks like they have optimized the distance computation so it goes very fast here13:18
blackburnI am still think we should stick to blas ;)13:18
blackburnit is optimized as well13:19
blackburnn4nd0: for example dnrm2 is more stable13:19
n4nd0once I start to do this part of the optimization we can test it13:19
blackburnyes just test13:19
-!- xiangwang_ [~xiangwang@2001:cc0:2020:2013:f2de:f1ff:fe50:9012] has joined #shogun13:23
sonne|workn4nd0: I think blas would be even faster13:26
n4nd0we should try then13:26
-!- PhilTillet [~Philippe@] has joined #shogun13:29
-!- xiangwang [~chatzilla@] has quit [Ping timeout: 276 seconds]13:34
n4nd0see you later guys13:37
-!- n4nd0 [] has quit [Quit: leaving]13:37
blackburnsonne|work: don't you think we should integrate cblas there?13:38
blackburnit is a common way I'd say13:39
-!- xiangwang [~chatzilla@] has joined #shogun13:40
sonne|workblackburn: yes that is what I said13:40
blackburnI mean to codebase13:40
blackburnsonne|work: many libs are integrating blas into their codebase13:40
blackburnhowever I see one great pitfall there - it would make packing .deb harder, right?13:41
blackburnbad way?13:43
sonne|workI don't want atlas/blas/gotoblas or anything in our codebase13:43
blackburnsonne|work: I am actually worried with arpack packaging - that's why I was thinking about it13:44
blackburnhowever better way would be to help with arpack package maintanence13:45
-!- xiangwang [~chatzilla@] has quit [Ping timeout: 265 seconds]13:46
blackburnsonne|work: how does it work? do maintaners have some public repo?13:49
sonne|workthere is a pts page13:50
sonne|workfor debian for each package13:50
sonne|workbut you can also do apt-get source <nameofpkg>13:50
sonne|workthen modify debian/ dir13:51
blackburnsonne|work: what if I want to fix something?13:52
-!- PhilTillet [~Philippe@] has quit [Ping timeout: 245 seconds]13:52
blackburnone should send patches to maintainers?13:52
sonne|workcreate a patch ans submit a bugreport via reportbug <packagename>13:52
blackburngot it13:52
-!- xiangwang [~chatzilla@] has joined #shogun13:53
blackburnsonne|work: btw did you get used to gnome3?13:54
blackburnI am becoming angrier with unity13:54
-!- harshit_ [~harshit@] has joined #shogun13:56
harshit_blackburn: hi !13:57
harshit_TSVM is almost complete, but there is some problem, should I create a pull request for TSVM ?13:58
-!- xiangwang [~chatzilla@] has quit [Ping timeout: 246 seconds]13:58
harshit_also do you happen to have old svmlin library's example with you, their link is  broken in web13:59
harshit_I sent a mail to vikas(author ) but no reply till now14:00
blackburnno, I haven't14:00
harshit_I think soeren might have it, coz he is author of svmlin in shogun ? Please ask him when he comes online, I may not be able to catch him.14:02
blackburnsonne|work: ^14:02
harshit_oh he is online !!14:02
blackburnbut I am afraid he just merged it14:03
harshit_what did he merge ??14:04
blackburnsvmlin to shogun's codebase14:05
harshit_no no , I am not asking for its code, that is available. All I am asking for is its 'standard' examples,which are available in a separate archive which in not downloadable14:06
harshit_Want to test shogun's TSVM and svmlin TSM on that dataset before proceeding14:07
blackburnah no idea14:09
harshit_no problem, btw how to issue *(some message) on IRC ?14:10
* blackburn this one?14:10
blackburn / me14:10
blackburnwithout spaces14:10
harshit_I cant see any thing14:11
-!- xiangwang [~chatzilla@] has joined #shogun14:11
harshit_ / checking some new command14:12
blackburn"/me something"14:12
* harshit_ hello world14:12
harshit_got it14:13
-!- wiking_ [~wiking@huwico/staff/wiking] has joined #shogun14:15
-!- wiking [~wiking@huwico/staff/wiking] has quit [Ping timeout: 246 seconds]14:18
-!- wiking_ is now known as wiking14:18
-!- xiangwang [~chatzilla@] has quit [Ping timeout: 260 seconds]14:24
-!- xiangwang [~chatzilla@] has joined #shogun14:32
-!- pluskid [~pluskid@] has joined #shogun14:34
-!- xiangwang [~chatzilla@] has quit [Ping timeout: 265 seconds]14:38
-!- n4nd0 [] has joined #shogun14:39
pluskidblackburn, have you looked at the LARS code?14:39
blackburnpluskid: yes14:39
pluskidany feedback?14:40
blackburnok not really I think it should be merged14:40
blackburnI'll do that once I get to home14:41
pluskidcool, thanks14:41
pluskidany suggestions on what to work on next?14:42
pluskidsmall ones14:42
blackburnhowever not so small..14:42
pluskidhmm, let's see what's it14:42
blackburnpluskid: have you seen how did we organized multiclass machine?14:44
pluskidno, haven't investigated that yet14:44
pluskidany problems there?14:44
blackburnyes as new multiclass machines (CS liblinear and multiclass ocas) are in multiclass14:45
-!- xiangwang [~chatzilla@] has joined #shogun14:45
sonne|workblackburn, pluskid: actually one could do the new subset stuff from heiko for oVo ?14:45
sonne|workOvO :D14:45
blackburnand are inherited from multiclass machine14:45
blackburnsonne|work: oVo looks nice ;)14:45
blackburnsonne|work: yes it is the task too14:45
blackburnsonne|work: my intention is to move larank and gmnp to multiclass14:46
sonne|workyeah but first finalize what n4nd0 started no?14:46
blackburnsonne|work: it is pretty unrelated14:46
blackburnwhile that stuff was generic and solvers do not make use of it14:46
n4nd0what is left from the part I did?14:46
blackburnI'll finish that on this weekend14:47
blackburnn4nd0: ovo subsetting14:47
blackburnpluskid: so the task could be14:47
blackburnto move larank to multiclass14:47
* pluskid is listening14:47
blackburnand adjust it accordingly14:47
n4nd0blackburn: didn't Heiko did that? I understood the thing that was left was to test if the results were now correct14:47
blackburnn4nd0: remember there is no add_subset for labels and kernels?14:48
sonne|workblackburn: but wouldn't he need it if he does other ECOC schemes?14:48
sonne|workahh labels14:48
* pluskid waiting for you guys to arrive at a conclusion14:48
sonne|workyes true that needs to be done14:48
blackburnpluskid: currently larank is derived from multiclasssvm14:49
pluskidso we should make it derive from multiclass machine?14:49
blackburnthis can be done with rebasing multiclasssvm actually14:49
pluskidok, I'll try that14:49
pluskidis multiclasssvm derived from multiclass machine?14:50
blackburnthat could be painful so just let me know if you don't like do such refactoring14:50
pluskidyeah, I haven't read those code yet14:50
blackburnno, multiclasssvm should be inherited from kernelmulticlassmachine I think14:50
blackburnsonne|work: agree?14:50
pluskidI'll try to look at it and ask here if I have problems :)14:51
-!- xiangwang [~chatzilla@] has quit [Ping timeout: 260 seconds]14:51
pluskidBTW: do we have publically available online doxygen doc14:52
pluskidfor our dev code?14:52
pluskidlatest code, automatically updated doc?14:52
n4nd0blackburn: I don't quite remember it that much but I think I already did the thing for kernels14:53
sonne|workblackburn: yes makes sense14:53
sonne|workpluskid: no14:53
sonne|workbut you can just run make doc14:53
pluskidsonne|work, yes, but that makes my local repo dirty14:54
pluskidI mean git status shows some changed files in doc directory14:54
pluskidwhen I compile the doc14:54
n4nd0blackburn: I think that what is left to do is if it is correctly working after Heiko's job14:55
pluskidn4nd0, what you are talking about is that: what I intend to do is already done before?14:56
n4nd0pluskid: no no, no worries14:56
n4nd0pluskid: another thing14:56
pluskidok :)14:56
n4nd0pluskid: another think that appeared in the conversation ;914:56
pluskidgot it14:57
* n4nd0 has weak fingers for parentheses today14:57
pluskidhave to connect a VPN to see files in dropbox14:57
blackburnmad skillz painting loool14:57
n4nd0blackburn: haha that explanation ^ is just awesome14:58
blackburndamn did they block dropbox?14:58
pluskidhave you guys heard about the Great Fire Wall in China?14:58
n4nd0yeah ...14:58
pluskidthey block many valuable sites14:58
pluskidwill be back in a while after a reconnection14:59
-!- pluskid [~pluskid@] has quit [Quit: Leaving]14:59
n4nd0blackburn: have you checked the commit ^ ?14:59
-!- pluskid [~pluskid@] has joined #shogun14:59
blackburnn4nd0: oh how did I miss that/15:00
blackburnis it merged?15:00
n4nd0blackburn: no15:00
blackburnwhat is you waiting for ;)15:00
n4nd0blackburn: I think we decided to wait after Heiko'15:01
pluskidwhat is the refactor figure mean?15:01
pluskidis this already done or waiting to be done?15:01
blackburnthat should be done15:01
blackburnleft (as is) => right (should be)15:01
pluskidso they all become disconnected?15:01
blackburnno my links were overlapped by package :D15:02
blackburnstructure is obvious15:02
blackburnthe only thing - multiclasssvm should inherit from multiclass15:02
pluskidcool figure btw :)15:03
blackburnI think we should make use of uml for such things - much easier to grasp15:04
-!- xiangwang_ [~xiangwang@2001:cc0:2020:2013:f2de:f1ff:fe50:9012] has quit [Remote host closed the connection]15:05
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has quit [Quit: Page closed]15:10
-!- wiking [~wiking@huwico/staff/wiking] has quit [Remote host closed the connection]15:19
-!- wiking [] has joined #shogun15:20
-!- wiking [] has quit [Changing host]15:20
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun15:20
-!- xiangwang [~chatzilla@] has joined #shogun15:22
-!- vikram360 [~vikram360@] has quit [Ping timeout: 265 seconds]15:28
-!- harshit_ [~harshit@] has quit [Quit: Leaving]15:28
-!- xiangwang [~chatzilla@] has quit [Ping timeout: 260 seconds]15:29
-!- vikram360 [~vikram360@] has joined #shogun15:37
wikingheheh it's great to do cuda calculations on your desktop machine... suddenly the whole gui became unresponsice :P15:52
-!- karlnapf [] has joined #shogun15:56
-!- uricamic [9320543b@gateway/web/freenode/ip.] has quit [Quit: Page closed]16:03
-!- xiangwang [~xiangwang@2001:cc0:2020:2013:f2de:f1ff:fe50:9012] has joined #shogun16:13
sonne|workkarlnapf: hey - the man is back :D16:29
karlnapfsonne|work hi there :) yes busy week here16:30
karlnapflast coursework deadline today16:30
karlnapffamily visiting ...16:30
karlnapfexam preparation16:30
sonne|worktough tough16:31
karlnapfhow are things?16:31
sonne|workdid you see that we got 9 slots?16:31
sonne|workof which we will give back 1 due to not enough mentors?16:32
karlnapfwhat? wow :)16:32
karlnapfno not yet16:32
karlnapfawsome this is16:33
sonne|workif 2/5th of them stay with us after GSoC we are in shogun heaven16:33
karlnapfvery cool16:34
karlnapfthis will probably happen16:34
karlnapfhow many proposals were there in the end?16:34
sonne|workfrom 38 students16:35
sonne|workstill a lot16:35
sonne|workprobably a dozen w/o patch though16:35
karlnapfindeed, very nice congratulations for the 9 slots16:35
karlnapfwhats next?16:36
karlnapfranking probably16:37
karlnapfsonne|work, talk to you later got a meeting now. I will be around in the evening and on the weekend16:39
wikingfinally we are on the same page with alex!16:41
-!- karlnapf [] has quit [Ping timeout: 252 seconds]16:43
wikingnow it's coding time \o/16:45
wikingand god damn cuda :)16:46
-!- karlnapf [] has joined #shogun16:49
-!- pluskid [~pluskid@] has quit [Quit: Leaving]16:51
wikingmmm was there a final decision whether shogun will depend on viennaCL ?16:53
-!- n4nd0 [] has quit [Quit: leaving]16:53
sonne|workno not yet16:54
sonne|workoptional in any case16:54
wikingi mean that the opencl solution will be actually based on viennacl?16:55
wikinganyhow i'll give my comments on the pull request as well16:56
-!- PhilTillet [] has joined #shogun17:19
-!- sonne|work [~sonnenbu@] has left #shogun []17:22
-!- gsomix [~gsomix@] has joined #shogun17:31
gsomixhome, sweet home17:31
-!- PhilTillet [] has quit [Ping timeout: 245 seconds]17:38
-!- PhilTillet [] has joined #shogun18:06
-!- xiangwang [~xiangwang@2001:cc0:2020:2013:f2de:f1ff:fe50:9012] has quit [Remote host closed the connection]18:11
-!- gsomix [~gsomix@] has quit [Ping timeout: 260 seconds]18:19
-!- gsomix [~gsomix@] has joined #shogun18:20
-!- nickon [] has joined #shogun18:44
-!- nickon [] has quit [Quit: ( :: NoNameScript 4.22 :: )]18:53
gsomixsonney2k, moin. how are you?19:19
-!- naywhayare [] has quit [Remote host closed the connection]19:27
-!- blackburn [~qdrgsm@] has joined #shogun19:42
wikingblackburn: have u ever worked with DomainAdaptationSVM?19:48
blackburnplanning to actually19:48
blackburnare you going to try whole shogun? :)19:48
wikingsome stuff19:48
wikingi'd like to test something19:48
wikingi have a pretty well trained svm19:49
wikingbut for some images i don't have some features19:49
blackburnat all?19:49
wikingsince some of the features are actually textual features19:49
wikingthat are not dependent on the image itself19:49
wikingbut where the image resided19:49
wikingand i have some random images about which i have no other info19:49
wikingso some features are basically missing19:49
wikingand i was wondering if i could still19:50
wikingclassify those images19:50
blackburndamn you have hard task :D19:50
wikingthis DomainAdaptationSVM19:50
wikingseems to be something19:50
wikingthat is worth a shot19:50
blackburnis it a multiclass problem you have?19:52
blackburnwould need to design new DA SVM then :D19:53
blackburnhow much time you have?19:53
wikingyeah just realized from the code19:53
wikingthat actually it seems to be binary classifier19:53
blackburnso no need to multiclass?19:53
wikingi would19:53
wikingbut src/shogun/classifier/svm/DomainAdaptationSVM.h is binary right?19:54
wikingmm yeah19:54
wikingsince it's derived from SVMLight19:54
wikingdamn i thought i could just use it out of box :)19:55
blackburnwiking: while this DA SVM was implemented by chris I can ask him19:58
blackburnand we can come up with linear multiclass da svm19:58
blackburnda svm lol19:59
wikingyeah would be great to extend it20:00
wikingmmm it has a kernelized version as well...20:00
blackburnwiking: yes while it is dual20:01
blackburnwiking: okay dinner time now and after I'll send chris a mail20:02
wikingu can cc me as well20:02
-!- harshit_ [~harshit@] has joined #shogun20:26
blackburnwiking: are you particularly interested in kernelized version?20:38
wikingwell that'd be great20:38
wikingbut i can do a hkm shit :)))20:38
-!- PhilTillet [] has quit [Ping timeout: 260 seconds]20:41
blackburnon last pages there are some DA techniques20:50
-!- PhilTillet [~Philippe@] has joined #shogun21:01
-!- gsomix [~gsomix@] has quit [Read error: Connection reset by peer]21:01
-!- puffin444 [230bf329@gateway/web/freenode/ip.] has joined #shogun21:01
-!- gsomix [~gsomix@] has joined #shogun21:16
blackburnwiking: sent21:17
-!- PhilTillet [~Philippe@] has quit [Ping timeout: 260 seconds]21:23
-!- naywhayare [] has joined #shogun21:24
-!- n4nd0 [] has joined #shogun21:46
blackburnwiking: time to take pencil! deriving CS with DA22:05
wikingcs =?22:05
blackburncrammer singer22:06
blackburnmy favourite multiclass formulation with simple dual ;)22:06
wikingtell me tell me22:07
wikingi make NOW :)22:08
blackburnwiking: make what? ;)22:10
wikingda with cs22:10
blackburnI'm going to implement it once I formulate it22:10
blackburnhowever that would be not so trivial22:10
-!- nickon [] has joined #shogun22:11
blackburnafaik liblinear solver do not support easy linear term setting22:11
wikingLog-likelihood: 860587008.00000022:12
wikinggmm-ing my dataset :P22:12
blackburnlog? oh that's nice22:12
n4nd0good night guys!22:14
wikingwell this would have had taken like about a week to calculate22:14
blackburnn4nd0: nite22:14
-!- n4nd0 [] has quit [Quit: leaving]22:14
wikingw/o having a nvidia tesla22:14
blackburnwiking: is the dataset so big?22:15
wiking5*10^6 vectors with each of 64 dimension22:15
wikingit was weird though22:16
wikinghad to do a *10^3 for each vector22:16
wikingas otherwise gmm couldn't find any cluster centers22:16
wikingso i don't know if on the end this is going to work out :)22:17
wikingi'm still amazed by this tesla... now it's 116W is being used now when it's running on 100%22:17
blackburnwiking: why not kmeans?22:17
wikingblackburn: since i'm trying to generate some fisher vectors22:17
wikingand i'd be interested as well in the soft clustering of gmm22:18
blackburnwiking: I see22:18
blackburnI just recalled there is an algorithm called ??????22:18
-!- puffin444 [230bf329@gateway/web/freenode/ip.] has quit [Quit: Page closed]22:20
blackburnwiking: idea of regularized DA svm is to22:21
blackburnoptimize 1/2 ||w-w0||^2 + C \sum_i \xi_i with same s.t.22:22
blackburnw0 is SVM trained on source domain22:22
blackburnand predictor is22:22
blackburnsounds reasonable but I need to think about it in MC mean22:23
-!- gsomix [~gsomix@] has quit [Ping timeout: 276 seconds]22:32
-!- karlnapf [] has left #shogun []22:46
@sonney2kblackburn, wiking well you can just use all the w_0 you obtained from the original training and learn new w's22:50
blackburnsonney2k: how?22:51
@sonney2kso it is basically the same thing like two-class - one just chooses the corresponding w_022:51
@sonney2kblackburn, I assume you do true MC22:51
@sonney2kso you have one-vs-rest classification in the end22:51
blackburnsonney2k: the problem is:22:52
blackburnmin 1/2 ||w-w0||^2 + C ..22:52
blackburns.t. <w,x>-...22:52
blackburnnot w-w022:52
blackburnso it can't be 'virtual' w-w022:52
-!- harshit_ [~harshit@] has quit [Read error: Connection reset by peer]22:52
blackburnsonney2k: got me?22:52
-!- harshit_ [~harshit@] has joined #shogun22:53
@sonney2knot yet22:54
blackburnsonney2k: I mean dual task will change in that case22:54
@sonney2kmin 0.5||w||^2 +0.5 ||w-w0||^2 + ...22:54
blackburnsonney2k: and liblinear does not provide an easy way22:54
blackburnwow what is it?22:54
@sonney2ksame thing like you use22:55
blackburnbut grad will change22:55
@sonney2kby some constant22:56
@sonney2kIIRC there is some gamma or whatever factor in front of the second term22:56
blackburnsonney2k: but still needs some patching there right?22:57
@sonney2kso you can scale gamma and get22:57
blackburnsonney2k: have you seen I've got a competitor?22:57
@sonney2k0.5 ||w||^2 + gamma w^T w_0 + ...23:00
@sonney2kyes you need to do some patching23:00
@sonney2kin the same way liblinear is patched now for binary svm23:00
@sonney2knot just one :D23:01
blackburnsonney2k: but one is going to contribute ;)23:01
blackburnsonney2k: do you think following is reasonable then?23:02
blackburn0.5 \sum_m ||w_m||^2 + gamma \sum_m wm^T w_0m + C \sum_i \xi_i23:02
blackburnhowever I can hardly see what is the trade-off we want to achieve there23:03
blackburnsonney2k: what is the role of w^T w_0?23:04
blackburnah stupid question it is w-w023:04
blackburnwiking: we should employ sonney2k more in future ;)23:11
-!- nickon [] has quit [Quit: ( :: NoNameScript 4.22 :: )]23:23
@sonney2kblackburn, and the ||w0||^2 part is const so can be left out from objective23:24
blackburnsonney2k: yes that's clear23:24
blackburnsonney2k: what stays unclear is how that addition changes the solution?23:25
@sonney2kdepends on value of gamma :)23:28
blackburnI mean does the solution change if we +1 to gradient?23:29
blackburnhmm it seems so but strange23:29
-!- wiking_ [] has joined #shogun23:33
-!- wiking_ [] has quit [Changing host]23:33
-!- wiking_ [~wiking@huwico/staff/wiking] has joined #shogun23:33
blackburnsonney2k: how many multitask proposals then?23:35
-!- wiking [~wiking@huwico/staff/wiking] has quit [Ping timeout: 248 seconds]23:36
-!- wiking_ is now known as wiking23:36
wikinghave i missed something? :S23:37
wikingthis was the last message i've seen: blackburn: hmm it seems so but strange23:37
blackburnwiking: no, the only thing is I asked how many multitask proposals we received23:37
wikingah there was a gsoc idea about this?23:41
blackburnwiking: what is do you think I am applying for then? ;)23:43
wikingmm maybe i should have tried this project :)))23:44
blackburnwiking: will you stop changing areas some day? :D23:44
-!- harshit_ [~harshit@] has quit [Read error: Connection reset by peer]23:44
wikingno i just need this one23:44
wikingi can test with 1 class23:45
wikingto see how it would work23:45
-!- harshit_ [~harshit@] has joined #shogun23:45
wikingthat'll do out of box...23:45
blackburnwiking: it doesn't depend on your problem already - I got interested to implement that :D23:47
wikingthere's code23:52
--- Log closed Sat Apr 14 00:00:17 2012