Open in new window / Try shogun cloud
--- Log opened Thu Dec 13 00:00:38 2012
shogun-buildbotbuild #173 of nightly_all is complete: Failure [failed compile]  Build details are at
shogun-buildbotbuild #216 of nightly_default is complete: Failure [failed compile]  Build details are at
-!- Hausas [~admin@unaffiliated/hausas] has joined #shogun03:12
-!- agas [~admin@] has joined #shogun03:35
-!- Hausas [~admin@unaffiliated/hausas] has quit [Disconnected by services]03:35
-!- agas is now known as Hausas03:35
-!- Hausas [~admin@] has quit [Changing host]03:35
-!- Hausas [~admin@unaffiliated/hausas] has joined #shogun03:35
-!- Hausas [~admin@unaffiliated/hausas] has quit [Ping timeout: 246 seconds]04:47
-!- Hausas [~admin@] has joined #shogun04:52
-!- Hausas [~admin@] has quit [Changing host]04:52
-!- Hausas [~admin@unaffiliated/hausas] has joined #shogun04:52
-!- Hausas [~admin@unaffiliated/hausas] has quit [Ping timeout: 248 seconds]06:34
-!- Hausas [~admin@] has joined #shogun06:36
-!- Hausas [~admin@] has quit [Changing host]06:36
-!- Hausas [~admin@unaffiliated/hausas] has joined #shogun06:36
-!- Hausas [~admin@unaffiliated/hausas] has quit [Ping timeout: 250 seconds]08:01
-!- n4nd0 [] has joined #shogun08:03
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]08:39
-!- Hausas [~admin@unaffiliated/hausas] has joined #shogun09:38
-!- n4nd0 [] has quit [Quit: leaving]09:53
-!- zxtx [~zv@] has quit [Ping timeout: 252 seconds]10:08
-!- zxtx [~zv@] has joined #shogun10:59
-!- zxtx [~zv@] has quit [Ping timeout: 245 seconds]11:04
-!- zxtx [~zv@] has joined #shogun11:10
-!- Hausas [~admin@unaffiliated/hausas] has quit [Ping timeout: 260 seconds]12:07
-!- zxtx [~zv@] has quit [Ping timeout: 245 seconds]12:10
-!- Hausas [~admin@unaffiliated/hausas] has joined #shogun12:12
-!- n4nd0 [] has joined #shogun12:51
wikingsonne|work: here?13:06
-!- Hausas [~admin@unaffiliated/hausas] has quit [Ping timeout: 245 seconds]13:06
-!- Hausas [~admin@unaffiliated/hausas] has joined #shogun13:06
sonne|workwiking: ye13:22
wikingsonne|work: i wonder what i was about to ask13:26
wikingah yeah one thing13:26
wikinghow's our release schedule13:26
wikingand 2) how do i apply for write permission into shoung13:26
wikingshogun repo13:26
sonne|workwiking: we have no schedule... if sb wants a release he should say so and we should then try to get things more stable / fix issues that appear (not that I know any...)13:27
sonne|workwiking: you get write permissions when you continue to stay active like you are now :)13:27
wikingsonne|work: but the question is when13:28
wikingi'm like almost a year now on the row13:28
wikingit'll be in february13:28
wikingthat i've submitted my 1st patch :)13:28
sonne|workohh indeed13:29
sonne|workthen *very* sooon13:29
wikingi think we should do a release still this year13:29
wikingbut there's certain things needs to be fix i think13:30
sonne|worknothing against it13:30
sonne|workplease start to fix things you come across then13:30
sonne|workI am not so sure what needs fixing though13:30
sonne|workI would have liked to attack sparse/stringfeatures to work with SG* datatypes13:30
sonne|workbut hey that can also wait for another release13:31
wikingi need to fix cc13:31
wikingwith dynamicarray13:31
wikingi know that13:31
sonne|workcc ?13:31
sonne|workwe don't have an example for ccsosvm right?13:31
wikingno not at all atm13:32
wikingi mean i've added it for13:32
wikingsonne|work: hehe yeah13:40
wikingyou've closed this PR13:40
wikingthere was the addition to the so_multiclass example13:41
wikingi'll make the changes though13:41
sonne|workthen fix it and reopen13:41
-!- yoo [2eda6d52@gateway/web/freenode/ip.] has joined #shogun13:45
yoohi all13:45
sonne|workso wiking if you want a relase start to test & fix things and some plan what should be stable by then and in NEWS would be nice too13:53
wikingi can start doing that next week13:56
wikingthis week is very busy for me estill13:56
sonne|workyoo please ask in the channel14:10
yooI can read in the code that OvRStrategy is used. Is CGMNPSVM OneVsRest or OneInAll ?14:13
n4nd0yoo: it should be one vs. rest as it says in the header and because they are using one vs rest in the implementation file14:20
n4nd0yoo: why do you wonder if you saw OvR was used?14:21
sonne|workyoo, n4n0 - it is true multiclass14:21
yoobecause it says "true multiclass svm fashion"14:21
sonne|workso all hyperplanes are learned simultaneously14:22
yooand true multiclass mean AllInOne ,14:22
n4nd0yoo: yeah ok, that is because in shogun you can do multiclass based on binary classifiers14:22
sonne|workbut when applying you have nr classes many classifiers14:22
n4nd0yoo: but there are also some classifiers that are natively multiclass, like this one14:23
yooallright because I used MKLMulticlass that use CGMNPSVM then for N classes I get global optimiaztion and one set of kernels weights14:23
yoowhat if I want differents set of kernels weights for each class ?14:24
yoois there a framework for that in shogun ?14:24
n4nd0yoo: I've never used it but it looks like MKLOneClass could make sense14:25
yooAnd what Multiclass machine resolve the OvR multiclass problem ?14:27
n4nd0yoo: CGMNPSVM? :)14:28
yooI mean c(x) = argmax(f_i) and not c(x) = argmax(w_i x + b_i) as CGMNPSVM14:28
yooor maybe I misunderstood smthg14:28
n4nd0I didn't really get the question to be honest14:29
n4nd0so an SVM intuitively puts a line (hyperplane) somewhere in the space14:30
n4nd0that can only separate two things14:30
n4nd0to do multiclass, you have to put several lines or use several SVMs14:30
n4nd0depending on how this lines are put you get one vs. rest or one vs. one14:31
n4nd0yoo: when you were saying one in all I guess you meant one vs. all?14:31
n4nd0in that case, one vs. all and one vs. rest is the same concept. It just makes more sense to say one vs. rest, it is more semantically correct I think14:32
yooOvR : N binary classfiers are trained to separate one class from the rest and winning class is the one that corresponds to the SVM with highest output value c(x) = argmax(f_i)14:34
yooAll in One : N two class rules the ith function separates training vectors of the class i from the other vectors14:35
yooN decisions but all are obtained by solving one problem14:35
n4nd0aham I see14:35
yooAll in One is more a generalization of the binary classification algorithm14:37
yoothen it is called "true multiclass"14:37
n4nd0yoo: I am not sure about this, but I think that CGMNPSVM solve OneVsRest, since they are using that MulticlassStrategy14:37
yoothat's I was saying ^^14:37
n4nd0hehe ok :)14:38
yoojust wanteed to be sure what it is exaclty14:38
n4nd0let me check a moment just in case14:38
yoobecause you can read in the header "true multiclass"14:38
yooThe question is: the N binary machine are learned independently or not  ?14:39
n4nd0yoo: I think that they are not learnt independently14:41
yoothen it is not OvR14:41
yoowhat make you say that ?14:41
n4nd0I ignored this distinction between OvR and all in one14:42
n4nd0I didn't know it had that name14:42
sonne|workn4nd0: no there is a true multiclass svm14:43
sonne|workn4nd0: that simultaneously learns all hyperplanes - so in the end you get a OvR classifier but hyperplanes are not based on learning serveral OvR binary SVMs but learnt in one go14:43
yooallright, then if you wanted to resolve a multiclass pb with OvR strategy what shogun framework will you use ?14:43
n4nd0sonne|work: I see, but even if learnt silmultaneously one could do that with OvR or OvOne I gues, right?14:44
yooI mean I want to optimize mkl weights for each class and MKLMulticlass use CGMNPSVM ..14:44
sonne|workn4nd0: I don't understand what you are asking14:45
n4nd0sonne|work: I understand the difference between learning all hyperplanes at the same time or not, we refer to this as TRUE multiclass when all of them are learnt together14:46
n4nd0sonne|work: ok?14:46
n4nd0sonne|work: but in true multiclass I suppose that one can do it either using an OvR strategy or a OvO strategy14:47
yoono I dont think so14:47
sonne|workn4nd0: true multiclass is what I mean with learning all hyperplanes together14:49
n4nd0sonne|work: ok, then we are speaking the same language :)14:50
sonne|workn4nd0: OvR OvO learning strategies are just for binary classifiers14:50
n4nd0yoo: anyway, you can use any of the binary classifiers in shogun to build a multiclass classifier using one vs. rest14:51
yoosonne|work: then if I want to do MKL and  optimize mkl weights for each class ?14:51
n4nd0yoo: see the KernelMulticlassMachine14:52
yoon4nd0: ok14:52
n4nd0sonne|work: I understand now, thanks! although I think that doesn't why is MulticlassOneVsRestStrategy used in GMNPSVM14:53
n4nd0doesn't explain*14:53
n4nd0I am guessing that it appears to follow the MulticlassSVM interface14:54
sonne|workn4nd0: well they use the same code when applying the classifier... argmax f_c(x)14:54
n4nd0got it14:57
-!- n4nd0 [] has quit [Ping timeout: 244 seconds]15:26
-!- yoo [2eda6d52@gateway/web/freenode/ip.] has quit [Quit: Page closed]15:43
-!- Hausas [~admin@unaffiliated/hausas] has quit [Read error: Connection reset by peer]15:56
-!- Hausas [~admin@] has joined #shogun15:57
-!- Hausas [~admin@] has quit [Changing host]15:57
-!- Hausas [~admin@unaffiliated/hausas] has joined #shogun15:57
-!- zxtx [] has joined #shogun16:13
-!- yoo [2eda6d52@gateway/web/freenode/ip.] has joined #shogun16:24
yoohi again16:25
yoosonne|work: I tried the following CKernelMulticlassMachine* mkl = new CKernelMulticlassMachine(new CMulticlassOneVsRestStrategy(), cker, new CMKLOneClass(), train_labels); but it raises the error: Interleaved MKL optimization is currently only supported with SVMlight16:27
yoocker for CCombinedKernels16:27
sonne|workwell you should not use oneclass16:28
sonne|workbut twoclass16:28
yooCMKL instead of COneClass ?16:29
-!- zxtx [] has quit [Ping timeout: 264 seconds]16:51
-!- zxtx [~zv@] has joined #shogun17:15
-!- heiko [] has joined #shogun17:25
-!- blackburn [~blackburn@] has joined #shogun17:53
-!- yoo [2eda6d52@gateway/web/freenode/ip.] has quit [Quit: Page closed]18:32
-shogungit:#shogun- [shogun] karlnapf opened pull request #845: some clean ups towars kernel selection framework for MMD (master...master)
-shogungit:#shogun- [shogun] karlnapf pushed 5 new commits to master:
-shogungit:#shogun- shogun/master 7303295 Heiko Strathmann: removed old kernel selection examples due to soon to be introduced new framework18:54
-shogungit:#shogun- shogun/master 33f4eb1 Heiko Strathmann: documentation18:54
-shogungit:#shogun- shogun/master 836d10a Heiko Strathmann: initialise optimisation parameters18:54
-!- heiko [] has left #shogun []19:07
-!- n4nd0 [] has joined #shogun19:19
shogun-buildbotbuild #720 of deb3 - modular_interfaces is complete: Failure [failed compile octave_modular]  Build details are at  blamelist: Heiko Strathmann <>19:39
-!- n4nd0 [] has quit [Ping timeout: 265 seconds]20:57
-!- blackburn [~blackburn@] has left #shogun []21:33
-!- Netsplit *.net <-> *.split quits: naywhayare21:47
-!- Netsplit over, joins: naywhayare21:48
-!- Hausas [~admin@unaffiliated/hausas] has quit [Ping timeout: 245 seconds]23:07
-!- Hausas [~admin@] has joined #shogun23:25
-!- Hausas [~admin@] has quit [Changing host]23:25
-!- Hausas [~admin@unaffiliated/hausas] has joined #shogun23:25
--- Log closed Fri Dec 14 00:00:38 2012