Open in new window / Try shogun cloud
--- Log opened Thu Aug 02 00:00:01 2012
--- Day changed Thu Aug 02 2012
n4nd0and MOSEK libraries00:00
blackburnokay sure but why not to have it automatically?00:00
n4nd0no idea00:00
blackburnso should I do that? :)00:00
n4nd0maybe Soeren just didn't want that I'd spend time with that00:00
blackburnyeah I can00:00
n4nd0sure, why not00:00
blackburnn4nd0: can you help me with that a little?00:00
n4nd0blackburn: yes00:01
blackburni need a minimal example00:01
blackburnthat fails w/o mosek00:01
blackburnand flags you use00:01
blackburn-lmosek or so?00:01
n4nd0I can give you that00:01
n4nd0I will do that tomorrow morning00:01
n4nd0I am going early today to sleep00:02
blackburnwanted to make it now :)00:02
blackburnjust say what are flags00:02
n4nd0wait a moment, I sent Nico a mail with the exact steps00:03
n4nd0-- In the line that start with DEFINES add -DUSE_MOSEK.00:03
n4nd0-- In the line INCLUDES add -I/usr/local/etc/mosek/6/tools/platform/linux32x86/h00:03
n4nd0where /usr/local/etc/ is the path to your mosek installation directory that may00:03
n4nd0be different for you.00:03
n4nd0-- In the line LINKFLAGS add -L/usr/local/etc/mosek/6/tools/platform/linux32x86/bin00:03
n4nd0where it applies the same as before about the mosek isntallation directory.00:04
n4nd0- In POSTLINKFLAGS add -lmosek.00:04
n4nd0I love the function of search in the mails :D00:04
blackburnthanks got it00:04
blackburnn4nd0: did you go already?00:12
n4nd0blackburn: no00:14
n4nd0tell me00:14
blackburnn4nd0: these paths wouldn't work as something default I am afraid00:15
blackburnso one would need to set up these paths on configure00:15
n4nd0whereis mosek outputs /usr/local/etc/mosek in my machine00:16
n4nd0maybe that could work?00:16
n4nd0although I do remember I told Soeren the same thing and he said that it was not platform independent00:16
blackburndoes it have pkg-config?00:16
n4nd0no idea what that is00:16
blackburncool thing allowing to do00:16
blackburnsomething like00:16
blackburnpkg-config --cflags mosek00:17
n4nd0let me try to exec that00:17
blackburnbut probably it would not work00:17
n4nd0Package mosek was not found in the pkg-config search path.00:17
n4nd0Perhaps you should add the directory containing `mosek.pc'00:17
n4nd0to the PKG_CONFIG_PATH environment variable00:17
n4nd0No package 'mosek' foun00:17
n4nd0you were right00:17
n4nd0doesn't work00:17
blackburnokay I've added something00:19
CIA-18shogun: Sergey Lisitsyn master * ref83b92 / src/configure : Added mosek detection -
shogun-buildbotbuild #240 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Sergey Lisitsyn <>00:19
blackburnehmm okay00:21
n4nd0ok good night now then!00:40
-!- n4nd0 [] has quit [Quit: leaving]00:40
-!- needsch [] has quit [Quit: Verlassend]00:48
shogun-buildbotbuild #241 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Sergey Lisitsyn <>00:50
CIA-18shogun: Sergey Lisitsyn master * rbe2d76b / (2 files): Fixed apply method call of multitask logistic regression -
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]01:25
shogun-buildbotbuild #242 of deb3 - modular_interfaces is complete: Success [build successful]  Build details are at
-!- zxtx [] has quit [Ping timeout: 260 seconds]02:09
-!- av3ngr [av3ngr@nat/redhat/x-cwemmkxkyzgxvooo] has joined #shogun03:24
-!- av3ngr is now known as abn03:24
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has quit [Quit: Page closed]03:44
shogun-buildbotbuild #47 of nightly_default is complete: Success [build successful]  Build details are at
-!- n4nd0 [] has joined #shogun06:53
-!- abn [av3ngr@nat/redhat/x-cwemmkxkyzgxvooo] has quit [Quit: That's all folks!]07:41
-!- uricamic [~uricamic@2001:718:2:1634:a48f:d0e6:2584:7cff] has joined #shogun08:56
-!- cronor [] has joined #shogun09:34
-!- cronor [] has left #shogun []09:34
-!- flxb [] has joined #shogun09:34
-!- pluskid [~pluskid@] has joined #shogun09:35
-!- Marina [82955874@gateway/web/freenode/ip.] has joined #shogun09:35
Marinahello, if i train a svm with dna sequence for splice site recognition, do the sequences with the splice site must have the +1 or the -1 label?09:43
n4nd0hi Marina09:49
n4nd0yes, if you use a binary classifier, for example one of the binary SVMs, the labels have to be either +1 or -109:50
Marinahi n4nd0 so, there is no difference, if i use the +1 one label for the sequences with the splice site, i.e. the positive examples, or if i use the -1 label for  the sequences with the splice site?10:00
n4nd0Marina: what are the two classes in your problem? I am not familiarized with splice site recognition10:01
n4nd0but I'd say that in any binary classification problem you should learn the same model if the labels are swapped10:03
n4nd0so it shouldn't make a difference10:03
Marinaafter training the svm i make a poim with the svm result - and the picture look different when training with -1 or +1 for the splice site sequences.10:07
n4nd0Marina: even if they look different, aren't they equivalent in some sense?10:10
n4nd0Marina: have you tried comparing the results of the weight vector after training with +1 and -1?10:11
Marinai will proof it! Thanks a lot10:16
n4nd0you are welcome10:16
n4nd0Marina: what svm classifier are you using by the way?10:16
Marinai use LibSVM10:24
n4nd0I have just a test with libsvm10:26
n4nd0I have train my svm two times, the second time swapping the labels of the training examples with respect to the first time10:27
-!- pluskid [~pluskid@] has quit [Quit: Leaving]10:28
n4nd0and I have printed the alphas, which define the model learnt, both times10:28
n4nd0the values are the same in both cases, with the exception that the signs are changed10:28
n4nd0e.g. the first time I get something like [0.68 -1 ...], the second time [-0.68 1 ...]10:29
n4nd0and that actually makes sense, graphically the hyperplane is the same10:29
n4nd0but the meaning of each region changes when the labels are swapped10:30
n4nd0Marina: could it be that the change in the signs made the poim different?10:31
koen-shogunby default, the costs (e.g. weights) of the classes are the same11:03
koen-shogunso you'd expect nearly the same solution11:03
Marinan4nd0 yes - that should be the reason!11:03
-!- Netsplit *.net <-> *.split quits: audy11:15
-!- Netsplit over, joins: audy11:17
-!- blackburn [~blackburn@] has joined #shogun12:01
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]12:34
-!- Marina [82955874@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]13:36
-!- koen-shogun [~ksande@] has left #shogun []14:50
-!- yoo [2eda6d52@gateway/web/freenode/ip.] has joined #shogun15:17
-!- n4nd0 [] has quit [Quit: leaving]15:33
-!- yoo [2eda6d52@gateway/web/freenode/ip.] has quit [Quit: Page closed]15:39
-!- blackburn [~blackburn@] has joined #shogun16:03
-!- yoo [2eda6d52@gateway/web/freenode/ip.] has joined #shogun16:30
yoohi all16:30
yooblackburn: reading your mail, thx for answering16:42
blackburnyoo: btw why do you use C++?16:44
yooblackburn: simply because I am more comfortable with c++16:48
yooblackburn: and since I have made some modifications in c++ shogun code ..16:49
yooblackburn:do you usually use python interface ?16:49
-!- audy is now known as octopine16:49
blackburnwell it is just the same actually and loading data in  C++ is usually a kind of pain16:50
yoothe thing is when I create my own c++ class, I have no experience in wrapping in python form16:51
yoocould be the time I get into this ..16:51
-!- yoo [2eda6d52@gateway/web/freenode/ip.] has quit [Quit: Page closed]17:05
blackburnhmm I am unsure what do you mean17:08
-!- uricamic [~uricamic@2001:718:2:1634:a48f:d0e6:2584:7cff] has quit [Quit: Leaving.]17:10
-!- yoo [2eda6d52@gateway/web/freenode/ip.] has joined #shogun17:33
-!- yoo [2eda6d52@gateway/web/freenode/ip.] has quit [Quit: Page closed]18:05
-!- KMcQuisten [d8338942@gateway/web/freenode/ip.] has joined #shogun18:05
KMcQuistenHello all18:05
KMcQuistenI'm getting some really odd behavior form LibSVR, and I was wondering if you all could help.  I can set up a kernel, initialize it, load that kernel and some labels into the LibSVR object, and then train the machine.  However, when I try and make a prediction, either on the training set using apply() or on a set of unseen features using apply(Cfeatures), it predicts an output of 1.0 for everythign, regardless of what I input.18:10
KMcQuistenI have a script I can send you to reproduce this.18:11
blackburnhmm sure, show me18:12
KMcQuistenWhat do you think?18:22
blackburngive me 5 mins more :)18:24
KMcQuistenNo problem :)18:24
-!- heiko [] has joined #shogun18:27
KMcQuistenSo what's the verdict?18:37
blackburnKMcQuisten: for now use apply_regression18:37
blackburninstead of apply18:37
KMcQuistenAllrighty :)18:37
blackburnI will check why it calls apply_binary here18:38
KMcQuistenThanks for the help18:39
blackburna few things to discuss I have you to18:42
blackburnheiko: mind talk I yoda like?18:42
heikothou shall he18:42
blackburnI heard you like multiclass18:43
blackburnso I put multiclass into multiclass18:43
blackburnokay I stop joking18:43
blackburnheiko: first thing18:43
heikogo ahead!18:44
blackburnare you familiar how multiclass machines work currently?18:45
blackburnyou must be18:45
heikolets just say yes18:45
heikoI fixed a few bugs in these so almost18:46
blackburnhow do you think we could pass18:46
blackburnthese scores18:46
blackburnto some evaluation18:46
blackburnproblem is that18:46
blackburnmulticlass labels is num + confidence18:46
heikoyou mean the score from OvO etc?18:47
blackburnbut for multiclass ROC we would need more confidences18:47
blackburnOvO ROC is probably not really informativ18:47
heikoI see18:47
blackburnbut OvR one would be nice18:47
heikowe got confidences on the one side18:47
heikoand scores on the other18:47
heikoAs I see it we are currently abusing the term confidence by putting scores in there that are not calibrated probabilities right?18:48
heikoso why not have two fields: scores AND probabilities18:48
heikoprobabilities can be used as scores18:48
heikobut not the other way round18:48
heikoand also it would have the advantage that users are not confused (I tend to be confused with these things for example)18:49
blackburnbut wait18:51
heikoany problems with that?18:52
blackburnwhat do we need to compute OvR roc?18:52
heikoPossible with both18:53
heikooh wait18:53
heikoits not possible to compute real probabilities for OvR18:53
blackburnso for each vector18:53
heikoits not really a probabilistic framework18:53
blackburnwe need all outputs18:53
heikoI mean its possible to approximate this18:53
blackburnis OvO probabilistic?18:53
heikobut scores are fine for ROC are they18:54
heikoSVM have no probabilistic interpretation anyway18:54
heikoall probabilities are just approximations or fitted sigmoids18:54
heikocouldnt we make this all optional?18:55
heikoI mean currently shogun cannot compute any probabilities on the output18:55
heikoI wanted to add that anyway once the labels could do it18:55
heikobut I still think its a good idea to distringuish18:55
blackburnhmm why do you want probabilities?18:55
heikofor example, the GP has a probabilistic interpretation, there you have only probabilities and not really scoresy18:55
heikowell for multiclass its great to compute posterior probabilities isnt it?18:56
heikofirst class 33% second class 32%, which one to take18:56
heikousing the one with largest scores looses something18:56
heikoon the other side:18:56
heikowhenever you have probabilities, you dont need scores anymore18:56
heikoits just when you have no probabilities18:57
heikothen you need the scores18:57
blackburnproblem is that18:57
blackburnwe currently discard scores and probabilities of rest classes18:57
blackburnwhen choosing one18:57
heikobut its also possible to "fake" probabilities, as for example for multiclass SVMs18:57
heikothen you got two different numbers18:57
heikoah ok18:57
heikoblackburn, what were the other things?19:03
blackburnI can't really see how to handle that19:03
blackburnheiko: other other19:03
blackburnother is19:03
blackburnmodelselection output19:03
blackburnwhat do you think is the way to *store* these things it prints now19:04
heikowhy do you want to do that?19:04
heikofor producing plots etc?19:04
heikowhat about adding a method that collects all performance measures for all parameter combinations?19:05
heikoand then output and array with that19:05
blackburnheiko: hmm19:07
blackburnheiko: yes, for plots and everything19:07
heikowhat exactly do you want to do with that?19:07
blackburnokay currently it prints everything19:07
blackburnprobably it would be better to store it19:07
blackburnand not to print19:07
heikodo you also want to store all indices etc?19:08
heikoor just performance per point?19:08
blackburnno, indices19:08
blackburnI think it is needed sometimes for reproducibility19:09
blackburnpeople really often make they own xval and modelselection snippets19:09
heikoI cannot imagine why one would want that. The thing about x-validation is that is gives you a normal distributed estimate of the real performance19:10
heikoI agree, because shogun often does not work ;)19:10
heikowhen it does, one doesnt need the indices in my eyes, or I can't imagine a use-case for that19:10
heikoanother thought: SHOGUN's model selection/x-validation is more for end-user cases: to automagically select parameters or get a performance measure. Its not really designed for analysing what happens in x-val.19:14
heikoBut thats a wide field and its almost impossible to cover everything that you might want to do with x-val.19:15
heikoso if people want to analyse more subtle stuff, they will have to write their own routines19:15
heikoand if they just want the distribution of the performance measure, we can add that19:16
heikoblackburn, what about the testing PR from wiking?19:17
heikothat is something I am really interested in :)19:17
blackburnsorry got distracted19:17
heikono worries :)19:17
blackburnI never got back to that PR atually19:18
blackburnheiko: can't we put GPL to readme or so?19:30
blackburnit is so damn big19:30
heikoehm, where?19:31
heikoin the tutorial?19:31
heikoblackburn, whats wrong with having it in the appendix?19:31
blackburnheiko: yeah, why to have it where?19:31
heikoI had this book-style in mind, so if printed, its good to put it somewhere19:32
heikobut I dont insist on it19:32
heikowe can give an url :)19:32
heikofor this license for documentation19:33
heikoit sais that the document contains a section where the full license is printed19:33
heikothats why I added it19:33
blackburnI see19:33
blackburnokay let it stay then19:33
heikobut I dont know whether its compulsory19:33
CIA-18shogun: Sergey Lisitsyn master * ra51e803 / src/interfaces/modular/Machine.i : Added apply rename for libsvr -
blackburnheiko: btw there is a shogun book already :D19:35
heikolol :D19:36
heikois that actually written?19:37
heikolook fakey19:38
blackburnfake for sure19:38
blackburnheiko: did you visit something at games?19:39
heikoblackburn, not yet19:39
blackburnwill you? :)19:39
heikobut I cannot get the motivation for this, I find it rather boring -- but I feel bad for this19:39
heikoand so  much work currently19:40
blackburnheiko: probably it is not so special in your countries :D19:41
blackburnthere is a good old hype around this19:41
blackburn"we must secure our regime via a lot of gold medals!!"19:41
blackburnthat's how china acts actually and soviet union been acting :D19:41
heikowell, yes it is, its actually amazing, and also a great opportunity since its next door, thats why I feel bad for not being interested :D19:41
heikoI like that approach btw ;)19:41
heikobecause it seems to naive19:42
blackburnwell competition is good19:42
heikokind of odd19:42
blackburnbut with all these dopes china uses (I am sure they do)19:42
blackburnit stops to be a competition of athletes :D19:42
heikoeverybody does this19:43
heikothe chinese are just best at staying closely under the accepted limit19:43
blackburnthey just find it important to produce these roboathletes, normal countries are more relaxed I believe :)19:44
-!- flxb [] has quit [Ping timeout: 240 seconds]19:46
heikoone GSoC student is an athlete btw did you know that?19:47
blackburnno I didn't know20:05
blackburnthat's crazy20:06
-!- KMcQuisten [d8338942@gateway/web/freenode/ip.] has quit [Quit: Page closed]20:22
blackburnhow does he manage to do that :D20:26
octopinecan anyone help me understand what this dip is all about?
blackburnuh interesting20:39
blackburnheiko: may be you?20:40
blackburnI am out of expertise there20:40
-!- cronor [] has joined #shogun20:42
heikooctopine, blackburn, sorry just read this, got to go,  may be able to help later20:55
-!- heiko [] has left #shogun []20:55
blackburnsonney2k: then you :)20:55
-!- cronor [] has quit [Ping timeout: 252 seconds]21:19
-!- cronor [] has joined #shogun21:20
-!- cronor [] has quit [Read error: Connection reset by peer]21:47
-!- cronor [] has joined #shogun21:48
@sonney2koctopine, well can happen... prc curve is not convex monotone or anything...22:12
blackburnsonney2k: but doesn't it mean something?22:18
blackburnthat dip22:18
blackburnsonney2k: locking is really helpful with multitask - that's really good we introduced it22:56
-!- cronor [] has quit [Read error: Connection reset by peer]22:59
-!- cronor [] has joined #shogun22:59
-!- zxtx [] has joined #shogun23:36
-!- zxtx [] has quit [Read error: Connection reset by peer]23:44
@sonney2kblackburn, thank heiko for that...23:48
-!- cronor [] has quit [Read error: Connection reset by peer]23:49
blackburnsonney2k: do we want to have bayesian networks?23:49
-!- cronor [] has joined #shogun23:50
-!- n4nd0 [] has joined #shogun23:51
n4nd0hey blackburn23:53
blackburnhey n4nd023:53
n4nd0how is it going?23:53
blackburnwell pretty fine, finishing up yet another refactoring of multiclass23:54
blackburnand you?23:54
@sonney2kn4nd0, you are right about the label stuff - if you swap +1 <-> -1 POIMs will be inverted...23:54
n4nd0I am fine too23:54
blackburnn4nd0: bioinformatics expert now? ;)23:54
n4nd0sonney2k: hey! I didn't expect you around here now :)23:54
n4nd0blackburn: haha not at all23:54
blackburnno matter how long I support shogun I NEVER damn understand these POIMs and stuff23:55
n4nd0sonney2k: I was going to consult with blackburn, tomorrow we have a meeting right?23:55
@sonney2kblackburn, well MKL and POIMs are not exactly trivial stuff23:55
blackburnsonney2k: mind send us a reminder?23:55
n4nd0sonney2k: unfortunately I won't be able to make it23:56
n4nd0I will be in a plane to Spain tomorrow at that time23:56
n4nd0I thought about sending to the mailing list an email with the summary of the state of the project, as we normally do in these meetings23:57
blackburnn4nd0: what you wanted to consult me about?23:57
n4nd0blackburn: if there was going to be a meeting or was it just a mistake in my agenda23:57
@sonney2kn4nd0, hmmhh so neither you nor nico will attend?23:57
@sonney2kblackburn, argh so let me send some reminder ...23:58
n4nd0sonney2k: I am not sure about Nico, but he has not been answering my mails for a long while ... so my guess is that we won't make it23:58
n4nd0I am sorry23:58
blackburnI bet it will be low population23:58
n4nd0sonney2k: if you let me know about other topics that were going to be discussed in the sesion I can include them in the mail23:58
n4nd0if that is ok for you of course23:59
--- Log closed Fri Aug 03 00:00:02 2012