Open in new window / Try shogun cloud
--- Log opened Mon Apr 23 00:00:19 2012
gsomixit's time to sleep00:03
gsomixgood night, guys00:03
n4nd0good night gsomix00:04
-!- puffin444 [230bf329@gateway/web/freenode/ip.] has quit [Quit: Page closed]00:06
-!- blackburn_ [d5578adc@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]00:41
harshit_good night guys.00:56
-!- harshit_ [~harshit@] has quit [Remote host closed the connection]00:57
-!- av3ngr [~av3ngr@] has quit [Ping timeout: 240 seconds]00:59
-!- av3ngr [~av3ngr@] has joined #shogun01:00
-!- PhilTillet [] has quit [Read error: Connection reset by peer]01:10
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]01:13
-!- n4nd0 [] has quit [Quit: leaving]02:23
-!- av3ngr [~av3ngr@] has quit [Ping timeout: 260 seconds]02:43
-!- av3ngr [~av3ngr@] has joined #shogun02:45
-!- pluskid [~pluskid@] has joined #shogun03:42
-!- av3ngr [~av3ngr@] has quit [Ping timeout: 260 seconds]04:02
-!- xiangwang [~chatzilla@] has joined #shogun04:04
-!- xiangwang [~chatzilla@] has quit [Client Quit]04:05
CIA-64shogun: Soeren Sonnenburg master * r719f903 / examples/undocumented/ruby_modular/modelselection_parameter_tree_modular.rb : disable printing of combinations -
-!- Netsplit *.net <-> *.split quits: pluskid06:57
-!- Netsplit over, joins: pluskid07:14
shogun-buildbotbuild #507 of ruby_modular is complete: Success [build successful]  Build details are at
-!- vikram360 [~vikram360@] has quit [Ping timeout: 250 seconds]07:37
-!- av3ngr [~av3ngr@] has joined #shogun07:54
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun08:26
-!- vikram360 [~vikram360@] has joined #shogun08:46
-!- vikram360 [~vikram360@] has quit [Ping timeout: 246 seconds]08:51
-!- vikram360 [~vikram360@] has joined #shogun08:52
sonne|workpluskid: I was just reading your patch thx!09:10
sonne|workpluskid: to get the idea - you introduce multiclassstrategy that can be attached to any multiclass classifier and will generate configurations to be learned by some binary classifier / same with evaluation09:16
-!- n4nd0 [] has joined #shogun09:17
pluskidsonne|work, yes09:18
sonne|workpluskid: that makes a lot of sense - only thing that might be confusing to users is that we have *true* multiclass machines09:19
pluskidsonne|work, the motivation is that blackburn suggest me to write a tree-structure training strategy09:19
sonne|worklike KNN that will directly do everything09:19
pluskidsonne|work, yes, like many MulticlassSVM09:19
pluskidthey handle training directly09:20
sonne|workI think we have only one multiclass svm - GMNP09:20
sonne|worklibsvmmulticlass is also doing OvO09:20
pluskidmaybe LaRank?09:20
-!- emrecelikten [~Anubis@] has joined #shogun09:20
pluskidI think either way (the old and the new) of handling multiclass strategy is OK, if we only have OvO and OvR09:21
pluskidbut if we want to introduce more, this way is more convenient09:21
pluskidespecially when we want to attach some more data to the strategy09:21
pluskidfor example, in a tree-structure training, we need the user to define a tree and attach that to the strategy object09:22
sonne|workpluskid: no your way is much better09:23
sonne|workI mean you should read the ECOC JMLR paper / code09:23
sonne|workthere are so many strategies to do so it is much nicer the new way09:23
pluskidhmm, you are right09:23
sonne|workpluskid: only thing I don't know is how to make clear in the class hierarchy that we have true multiclass machines09:24
sonne|worknot requiring any of this09:24
pluskidyes, and training and applying can also be separated09:24
sonne|workyeah only training is true multiclass for many 'true' multiclass machines09:25
pluskidsome machines use their own training, but use the generic strategy for applying09:25
sonne|worknot clear how to handle that yet09:25
sonne|workpluskid: anyway the best way to proceed is to implement other ECOC schemes from the paper - things will become clear then09:27
sonne|workpluskid: btw I added some (minor) comments to your patch - please address them and then it is ok to be merged09:27
pluskidsonne|work, yes, practicing is the best teacher :p09:27
pluskidOK, I'll look at them09:28
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]09:28
sonne|workpluskid: thanks and please continue the great work!09:32
-!- wiking [] has joined #shogun09:32
-!- wiking [] has quit [Changing host]09:32
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun09:32
pluskidsonne|work, thanks, I'll try my best :)09:32
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has joined #shogun10:00
-!- emrecelikten [~Anubis@] has quit [Quit: Leaving.]10:20
blackburnsonne|work: multiclassliblinear and multiclassocas are true too10:24
pluskidhaha, my laptop is *just* capable of compiling shogun, if I opened other big prog (like Matlab) simutaneously, the compiler will complain out-of-memory10:26
blackburnpluskid: oh you are here ;)10:27
blackburnyou did great work10:27
blackburnI haven't been thinking about it actually10:28
pluskidblackburn, thanks, I'm making some modifications according to sonney2k's feedback10:28
-!- emrecelikten [~Anubis@] has joined #shogun10:29
-!- pluskid [~pluskid@] has quit [Ping timeout: 260 seconds]10:34
-!- pluskid [~pluskid@] has joined #shogun10:52
pluskidsonne|work, blackburn I updated the pull request10:52
blackburnyes thanks10:52
blackburnpluskid: have you seen ECOC training code in scikits learn?10:54
pluskidblackburn, no10:54
blackburnlooks pretty easy actually10:55
pluskidI think I understand the basics10:57
pluskidafter reading its comments10:57
blackburnpluskid: yes rather easy - I am also having some long discussion with v. franc about training ECOC svm in true multiclass fashion10:58
pluskidwhat does that mean? "in true multiclass fashion"10:59
blackburnpluskid: in one optimization problem11:00
blackburnimplies better normalization11:00
blackburnpluskid: I mean in basic ECOC with given K bits11:01
blackburnyou would need to construct K classifiers11:01
pluskidso you train the K classifiers in one time?11:01
blackburnI've been thinking about SVM (looks like SO at some place)11:01
blackburnactually I'd like to construct K classifier for each class11:02
blackburnhowever K classifiers at the same time would work too may be11:02
pluskidK classifier for each class?11:02
blackburnhmm yes if I didn't get crazy11:03
pluskidI didn't quite get you, so e.g. we have C class, K bits, then we have CK classifiers11:03
pluskidwith a test point x, how do we classify it?11:04
blackburnmax_{c\in C} \sum_{k\in K} <w_c^k, x>11:05
blackburndoes it make sense for you?11:05
blackburnI am unsure already11:05
pluskidlet me think11:06
blackburnit is not ECOC :D11:06
blackburndaaamn I have to think about it11:06
pluskidno, this has nothing to do with ECOC11:07
pluskidordinary one-vs-rest linear machine, haha11:07
pluskidmove the sum inside the inner product11:07
blackburnnot ovr11:07
blackburncomposition of K OvR - useless11:08
pluskidhmm, but you can move the sum into the inner product, then \sum_{k\in K}w_c^k can be treated as a single w_c11:09
blackburnah right11:09
blackburnpluskid: no, decision rule is wrong11:10
pluskidblackburn, why?11:10
blackburnshould be11:10
pluskidI don't understand11:11
blackburnshould be different I mean11:11
blackburnpluskid: nevermind this time11:11
CIA-64shogun: Soeren Sonnenburg master * r1ccb9e4 / (21 files in 5 dirs):11:17
CIA-64shogun: Merge pull request #479 from pluskid/multiclass-strategy11:17
CIA-64shogun: Multiclass strategy refactoring (+14 more commits...) -
sonne|workpluskid: thanks for the quick updates!11:17
pluskidsonne|work, :)11:17
pluskidhope build-bot is happy with my patches, though I have already checked locally11:18
sonne|workpluskid: I hope so too :D11:18
sonne|workblackburn: how many exams left?11:18
blackburnsonne|work: 311:18
blackburnsonne|work: but they are not related with optics11:18
blackburnso I am available11:19
blackburnsonne|work: do you think I should ask SLEP guys to GPL their code under their copyright?11:20
sonne|workblackburn: I think we should do the intrusive renaming changes rather soon11:20
blackburnsonne|work: which ones?11:20
sonne|workblackburn: like moving around multiclass classifiers11:20
sonne|workand multitask / domainadaptation stuff11:20
blackburnsonne|work: I still think multitask normalizers are better suited to be in transfer/multitask..11:21
sonne|workthen the only remaining intrusive stuff for this year is IMHO sgvector11:21
sonne|workmaybe but the should be totally hidden from the user11:22
blackburnsonne|work: but what to rename?11:24
-!- pluskid [~pluskid@] has left #shogun ["for supper"]11:26
-!- pluskid [~pluskid@] has joined #shogun11:27
shogun-buildbotbuild #726 of cmdline_static is complete: Failure [failed test_1]  Build details are at  blamelist: pluskid@gmail.com11:33
sonne|workblackburn: rename includes moving things around :)11:34
sonne|workblackburn: so?11:35
blackburnsonne|work: so what?11:36
blackburnactually you are right we need to move knn, etc to multiclass11:36
shogun-buildbotbuild #704 of r_static is complete: Failure [failed test_1]  Build details are at  blamelist: pluskid@gmail.com11:36
sonne|workI am saying better do all the rename / move now than later when we run on full steam11:37
blackburnah yeah I'll do something about it11:39
shogun-buildbotbuild #705 of octave_static is complete: Failure [failed test_1]  Build details are at  blamelist: pluskid@gmail.com11:41
shogun-buildbotbuild #692 of python_static is complete: Failure [failed test_1]  Build details are at  blamelist: pluskid@gmail.com11:45
n4nd0see you later people!11:46
blackburnsee yopu11:47
-!- n4nd0 [] has quit [Quit: leaving]11:48
shogun-buildbotbuild #727 of cmdline_static is complete: Success [build successful]  Build details are at
shogun-buildbotbuild #705 of r_static is complete: Success [build successful]  Build details are at
shogun-buildbotbuild #706 of octave_static is complete: Success [build successful]  Build details are at
shogun-buildbotbuild #693 of python_static is complete: Success [build successful]  Build details are at
shogun-buildbotbuild #517 of java_modular is complete: Failure [failed compile]  Build details are at  blamelist: pluskid@gmail.com12:04
shogun-buildbotbuild #540 of octave_modular is complete: Failure [failed compile]  Build details are at  blamelist: pluskid@gmail.com12:05
shogun-buildbotbuild #528 of csharp_modular is complete: Failure [failed compile]  Build details are at  blamelist: pluskid@gmail.com12:06
shogun-buildbotbuild #516 of lua_modular is complete: Failure [failed compile]  Build details are at  blamelist: pluskid@gmail.com12:07
shogun-buildbotbuild #513 of python_modular is complete: Failure [failed test_1]  Build details are at  blamelist: pluskid@gmail.com12:14
shogun-buildbotbuild #508 of ruby_modular is complete: Failure [failed test_1]  Build details are at  blamelist: pluskid@gmail.com12:20
shogun-buildbotbuild #512 of r_modular is complete: Failure [failed compile]  Build details are at  blamelist: pluskid@gmail.com12:21
shogun-buildbotbuild #518 of java_modular is complete: Success [build successful]  Build details are at
-!- pluskid [~pluskid@] has quit [Ping timeout: 246 seconds]12:35
-!- pluskid [~pluskid@] has joined #shogun12:35
shogun-buildbotbuild #541 of octave_modular is complete: Success [build successful]  Build details are at
-!- PhilTillet [] has joined #shogun12:41
PhilTillethello everybodyh12:44
shogun-buildbotbuild #529 of csharp_modular is complete: Success [build successful]  Build details are at
shogun-buildbotbuild #517 of lua_modular is complete: Success [build successful]  Build details are at
gsomix>> Welcome again to the Machine learning class.12:59
gsomixhi all12:59
PhilTillethi gsomix12:59
PhilTilletgsomix, yes, a lot of coursera courses start today13:04
shogun-buildbotbuild #509 of ruby_modular is complete: Success [build successful]  Build details are at
shogun-buildbotbuild #513 of r_modular is complete: Success [build successful]  Build details are at
blackburnhappiness again13:22
sonne|workblackburn: look at python_modular13:27
blackburnsonne|work: yes13:30
blackburnI'll fix once I get home13:30
blackburnsonne|work: I think that's ok to patch true multiclass machines13:35
blackburnto avoid this error we can add trained flag there13:35
blackburnso once we train machine we store number of classes13:35
sonne|workblackburn: just think of doing x-validation with Mc machines13:36
blackburnwhat is the problem?13:36
sonne|workfor some subsets you observer x for some x-1 classes or less13:36
sonne|worksame problem with sparse features btw13:36
sonne|workthere could be 1000 dims or 10013:36
blackburnthat is the reason to train on different number of classes?13:36
blackburnno sense for me at all13:37
-!- wiking [~wiking@huwico/staff/wiking] has quit [Remote host closed the connection]13:37
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun13:37
sonne|workblackburn: you said you want to *infer* number of classes based on labels13:37
sonne|workthat is dangerous13:38
blackburnyes train labels should contain all possible classes13:38
sonne|workI think specifying nr classes is better13:38
sonne|workblackburn: why should they?13:38
sonne|workif you do x-val you might not have some classes13:38
blackburnbad x-val13:38
sonne|worknot bad xval13:38
blackburnbut you will get wrong results!13:39
blackburnsonne|work: how can you ignore some classes?13:40
blackburnso if you skip class C you will have zero w_c, what makes you able to use it?13:42
blackburnwiking: hey13:43
pluskidI support specifying nr classes when constructing a classifier13:43
pluskidjust a matter of taste, however13:43
blackburnpluskid: both ways are ok13:43
blackburnhowever crossvalidation with class ignores makes no sense to me13:44
blackburnthings would be totally wrong if you skip some classes13:44
pluskidI mean it would be easier to implement the code if we are given nr classes at first13:45
pluskidthen no need for a flag like m_trained or similar13:45
blackburnpluskid: where is it set?13:45
pluskidno in current code13:45
pluskidbut I think it might be needed13:45
blackburnpluskid: yes that's I am asking13:45
blackburnon which step?13:46
blackburnnot in current?13:46
pluskidfor example, you do not know the number of machines, until it is trained13:46
blackburnhow does it work now13:46
pluskidfor example, in MulticlassMachine13:47
pluskidwill return "wrong" value, until the machine is actually trained13:47
pluskidWe need to know the state of the machine to see whether some "knowledge" is already available13:48
pluskidhere the number of classes13:48
pluskidthe same problem happens in MulticlassStrategy13:48
pluskidMulticlassStrategy::get_num_machines(int32_t num_classes)13:48
pluskidone has to pass num_classes as an argument when calling the function13:49
pluskidbut if we know num_classes from the begining, there's no need for this argument13:49
pluskidanyway, I think it is just a matter of taste, coz the current solution also works13:49
blackburnpluskid: I think we can assume no ties in labels while training13:50
blackburnhowever it looks like it would be better to store number of classes on training13:50
blackburnand obtain it while applying13:50
blackburnthis way you can be sure it works with ties in test set13:51
wikingblackburn: ho :)13:51
blackburnwiking: does it make sense to you to skip some classes in train set?13:52
blackburn(that are not skipped in test one)13:52
pluskidblackburn, I'm OK with this solution13:52
wikingblackburn: why would one do that?13:52
blackburnwiking: x-val13:52
pluskidthough still prefer specifying nr classes in advance :p (my taste...)13:53
blackburnpluskid: in advance?13:53
pluskidI mean when constructing the classifier13:53
blackburnwhen training or constructing?13:53
blackburnpluskid: problem is that if you have ties in labels13:54
blackburnlike test set contains only 1,2, 4 instances13:54
blackburnof classes 1,2,413:54
blackburnI mean13:54
blackburnso get_num_classes would return 313:54
blackburnand you will run only 3 of 4 machines13:55
blackburnskipping last one - it looks incorrect13:55
pluskidI mean in the constructor13:58
pluskidbefore training13:58
pluskidwhat do you mean by "ties in labels"?13:58
blackburnno 313:58
blackburnpluskid: but what if you set different labels?13:58
pluskidbut the number of labels won't change, will it?13:59
pluskidnumber of classes13:59
blackburnwhy not?13:59
blackburnyou can set features, kernel, labels - everything14:00
pluskidYou mean 1,2,4 or 2,3,4, or 1,2,3, etc?14:00
blackburnjust some missing class14:00
pluskidso we need set_num_classes, haha14:00
blackburnpluskid: yes we need to save num of classes on training14:00
blackburnwe can't rely only on constructor - set_labels() would make no sense then14:01
pluskidblackburn, OK, that makes sense14:01
blackburnhowever this requires to set number of classes in strategy14:02
blackburnin every true svm as well14:02
blackburntrue multiclass*14:02
pluskidyes, that's the problem14:02
blackburnnot really14:02
sonne|workblackburn: I think that is the same issue like having a linear classifier operating on sparse features - sometimes not the last dim in any of the sparse feature vectors is non-zero so you need to specify length of w of the sparse classifier or so14:03
blackburnsonne|work: are you still sure train set can contain a few classes?14:03
blackburn(not all classes)14:04
blackburnsonne|work: and how would you classify class X if you haven't trained a classifier for that?14:05
sonne|workit will never predict class X but operate correctly otherwise14:06
blackburnsonne|work: it broke other class' results14:06
blackburnsonne|work: hmm I assume it would be w=014:06
blackburnsonne|work: it can override some other class scores14:07
blackburnah damn14:07
blackburn0 always14:07
blackburnokay sorry14:07
blackburnsonne|work: it complicates things pretty much14:08
pluskidI think if the samples for a class is missing, the problem is not with predicting, but with training. You need a one-class algorithm to train a "classifier" :D14:09
blackburnpluskid: I am not even sure that would work with current algorithms14:09
blackburnlike gmnp/etc14:09
pluskidhowever, if this happens, it should be an error14:10
pluskidso if we are specified nr classes, we might be able to detect this kind of error14:10
pluskidbe back in a while14:10
blackburnwe need map there14:11
blackburnto avoid any errors14:11
blackburnwe can't keep things like that anymore14:11
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has quit [Quit: Page closed]14:57
-!- gsomix [~gsomix@] has quit [Ping timeout: 265 seconds]15:00
-!- PhilTillet [] has quit [Read error: Connection reset by peer]15:09
-!- harshit_ [~harshit@] has joined #shogun15:27
-!- ted__ [2e1fd566@gateway/web/freenode/ip.] has joined #shogun15:28
ted__hi all15:28
ted__what type of Features do I need to use (c++) if I want to train multiclass svm using string kernel with float array samples of different size ?15:30
ted__thanks for help15:30
ted__and which string kernel is suitable for samples of different sizes ?15:31
-!- PhilTillet [] has joined #shogun15:40
-!- gsomix [~gsomix@] has joined #shogun15:50
ted__RE: hi there, what type of Features do I need to use (c++) if I want to train multiclass svm using string kernel with float array samples of different size ? and which string kernel is suitable for samples of different sizes ? Thanks in advance. T.15:54
harshit_ted__, hi ,I havn't had much experience with strings in shogun but i think you need to use CStringFeatures<char> as features .16:04
harshit_ted__, have a look at examples/documented/libshogun/features_copy_subset_string_features16:04
harshit_ted__, Also wait for a while till other members are online [blackburn, soeren] they will help you better16:05
-!- blackburn [~qdrgsm@] has joined #shogun16:18
PhilTillet4hours40 :S16:20
harshit_blackburn, hi , I think ted__  needs ur help :)16:20
blackburnharshit_: yes I just read logs16:21
-!- vikram360 [~vikram360@] has quit [Ping timeout: 265 seconds]16:21
harshit_PhilTillet, really nervous moments :/16:21
sonne|workted__: StringCharFeatures ?16:23
PhilTilletharshit_, let's just get our body frozen for 5 hours16:26
blackburnpluskid: I don't like dependency of multiclassmachine and multiclassstrategy16:30
harshit_PhilTillet, haha ..16:30
-!- pluskid [~pluskid@] has quit [Ping timeout: 246 seconds]16:33
harshit_blackburn, I think there might me some issue with svmlin.cpp present in shogun, I have compared the results of original svmlin and that of shogun's, they are close but not exactly similar .16:39
harshit_blackburn, have a look :
blackburnare all parameters equal?16:40
harshit_blackburn, In svmlin original library we can set lambda so I set it to 116:41
ted__sonne|work, harshit_ : sorry, was away16:41
gsomixharshit_, I think I'll drink vodka in the 17-18UTC and go to bed, aha16:41
harshit_blackburn, but in shogun internaly lambda = 1/2c so I make value of C as 0.516:41
blackburnharshit_: but other params? are there any?16:42
harshit_blackburn, only other parameter is epsilon which is 1e-5 in both16:42
ted__Let's resume: I have samples of different sizes as vectors of float , I am looking for best shogun feature structure and kernel to train multiclass svm.16:43
ted__I'll take a look at StringCharFeatures16:43
ted__what about kernels ?16:43
blackburnted__: vectors of float?16:44
blackburnof variable size?16:44
harshit_gsomix, I think vodka is for blackburn, we should go for tequila !16:44
blackburnted__: non-trivial.. how can you compare it?16:44
ted__I know, I'm guessing stringkernel could do the work as they implement smthg like dtw distance ,16:45
gsomixharshit_, huh, I have never tried tequila.16:46
ted__It can be vectors of int but sizes are different.16:46
harshit_g2g, blackburn i'll do some more debugging and will let you know about the status !16:46
blackburnharshit_: yeah thanks16:46
blackburnted__: yes makes sense but they are for more 'finite' features like char16:47
harshit_gsomix, you should, maybe today !!16:47
ted__blackburn: and I dont want to use simple bag-of-words representation16:47
blackburnted__: what is the nature of your features?16:47
ted__blackburn: [w1,w2,w3,..] with w_iEUR[0..N] N could be 10, 100, 1000, or 1000016:48
ted__and size of the "string" is not defined16:48
blackburnted__: have you any idea of similarility function? you can create custom kernel actually16:49
harshit_ted__, what I think is that you should first use string to manipulate the words and then prepare its corresponding VSM model which you can store in a normal CSimpleFeatures16:49
-!- pluskid [] has joined #shogun16:50
ted__harshit_: VSM ?16:50
harshit_vector space model16:50
harshit_in which you have words as features and documents as examples/datapoints16:51
-!- harshit_ [~harshit@] has left #shogun ["Leaving"]16:52
ted__I thought that "string kernel" could handle features of different size ?16:52
blackburnted__: yes if they are contained with chars16:53
ted__then clustered by [0,255] ?16:54
blackburnyes would work16:54
ted__Then it could be same algo for string of cluster indices (indices btw 0 and 1000..)16:55
ted__Well, which shogun features/cluster could I overload if I want to do that ?16:56
blackburnted__: could you please clarify what exactly you need?16:56
ted__Well, which shogun features/kernel could I overload if I want to do that ?16:57
blackburnrepeating sentence is not a clarification I'd say :D16:57
ted__blackburn: sry, I just type cluster and meant kernel .16:58
blackburnted__: you don't really need to overload features for that16:59
ted__blackburn: So. I have a set of indices vector of different size. Let's say Indices EUR [0,100]. Each vector is associated with a label. let's say we have 10 labels.16:59
-!- pluskid [] has quit [Quit: Leaving]17:00
ted__blackburn: I want to train multiclass svm of this data. Problem is that features have different size.17:00
-!- pluskid [] has joined #shogun17:00
ted__blackburn: indeed I have read about string kernel apply to string of char. I want to do the same with my problem.17:01
blackburnted__: you may try to construct StringFeatures<char> with your features17:01
blackburnand use any kernel17:01
blackburnstring kernel I mean17:01
ted__ok. And if I have more than 256 indices ?17:02
blackburnted__: you have vecotr x = [x1,x2,...] where xi \in [0,255]?17:03
blackburnwhat do you mean more that 256 indices? wider range of xi values?17:03
ted__blackburn: I have vector x = [x1,x2, ..] where xi \in [0,1000]17:05
-!- vikram360 [~vikram360@] has joined #shogun17:05
blackburnted__: ok and what is the similarity measure you want to use?17:06
blackburnI am asking because string kernel are mainly used by bioinformaticians to estimate similarities between some ACTCGCT and TCGCGA - not a big alphabet17:07
ted__I know17:08
@sonney2kwiking,gsomix,pluskid,blackburn, do you have some homepage I could link to?17:08
blackburnI haven't17:08
wikingsonney2k: mmm not really :(17:08
wikingbut i'm planning to keep a weekly posting on maeth.com17:08
pluskidsonney2k: http://pluskid.org17:08
wikingbut currently it's personal rampage :)17:08
gsomixshogun-buildbot, nope17:09
gsomixsonney2k, nope17:09
ted__Dunno exactly yet what type of  similarity measure.17:09
ted__Maybe for the beginning one you used with your already implemented string kernels which support features of diff size17:10
ted__LocalAlignmentString ?17:10
ted__Any advice ? :p17:10
blackburnted__: it is based on alignment - I am not sure it would work for your features17:11
blackburnted__: where these features come from?17:11
gsomixpluskid, cool pictures :]17:12
ted__spatial arrangement of my vectors are important: features extracted from video sequence, then I would like to keep the temporal info (=as in ACTCGCT)17:13
pluskidgsomix: you like them? :D17:13
gsomixpluskid, yep17:13
pluskidgsomix: thanks! haha, I like drawing17:14
blackburnted__: really non-trivial thing you do17:14
ted__I know but it seems to be a good way to keep temporal information which is lost in a bag-of-frame algo.17:22
ted__maybe I could try with a mapping to [0,255] and see the results17:22
-!- pluskid [] has quit [Quit: Konversation terminated!]17:23
ted__thanks for helping , I anythg else come in mind, I stay here for a couple hours more17:23
-!- pluskid [] has joined #shogun17:23
blackburnpluskid: you forgot to ref multiclass strategy ;)17:24
pluskidblackburn, oh, my fault!17:25
blackburnpluskid:  I fixed17:25
pluskidcould you please help me to fix this?17:25
blackburnwill commit in a min17:25
blackburnpluskid: I am currently working on getting examples to work17:25
blackburnbecause of that enum -> class change17:25
ted__blackburn: do you have an example using string kernel and multiclass svm ?17:26
pluskidblackburn, I see, I fixed libshogun, but it seems I forgot python_modular17:26
pluskidblackburn, thanks again!17:26
-!- av3ngr [~av3ngr@] has quit [Ping timeout: 260 seconds]17:26
blackburnted_: in C++?17:26
ted__blackburn:  yes please, with features of diff size17:27
blackburnted__: check examples/undocumented/libshogun/ for some examples17:30
blackburnI don't think we have any complex C++ example of string features and kernel learning17:30
ted__blackburn: I will try EAlphabet with RAW BYTE - [0-255]17:32
pluskidsonney2k: why did you collected the time zones for all the candidates? Is there any online conferences?17:32
blackburnpluskid: yes for kickoff meeting17:32
ted__I already checked those examples, was hoping you had somthg more specific on your own machines.17:32
blackburnted__: I am sorry, I haven't worked with such features (yet)17:33
ted__blackburn: thanks anyway for all those info. Maybe anyone else have ?17:34
blackburnted__: yes sonney2k has some experience with strings but he is pretty busy17:34
blackburnpluskid: got error :(17:35
blackburn  File "", line 25, in classifier_multiclassmachine_modular17:35
blackburn    mc_classifier.train()17:35
blackburnSystemError: [ERROR] Not a two class labeling - no positively labeled examples found17:35
pluskidblackburn: leave that for me17:37
blackburnpluskid: ok I'll push some fixes17:37
pluskidblackburn: thanks!17:37
CIA-64shogun: Sergey Lisitsyn master * reeaa056 / (5 files in 3 dirs): A bunch of fixes for multiclass machine -
pluskidblackburn: BTW, what is the kickoff meeting for?17:38
blackburnpluskid: hmm I think we need to discuss our plans17:39
blackburnit is already late there you are..17:39
blackburnpluskid: will you be available at 19 UTC?17:39
pluskidI mean, we don't need to be there when Google announce the result, do we?17:39
blackburnyes that's not required17:39
pluskidOK, cool17:40
pluskidif it's 15:40 UTC now17:40
pluskidI don't think I will be available at 19 UTC17:40
pluskidunless required17:40
pluskidhave to sleep :p17:40
blackburnpluskid: when will you wake up?17:41
pluskidlet me calculate17:41
pluskidaround UTC 00:30 maybe17:41
blackburnthere is a chance I will not be sleeping yet17:41
blackburna lot of things to do today17:42
pluskida regular sleeping time is good for a programmer!17:42
pluskidto live longer! :D17:42
blackburnpluskid: it would be painful to not complete these things today :)17:43
blackburnpluskid: okay just read logs once you wake up17:43
pluskidblackburn: hmm, then wish you finish them soon17:43
ted__sonney2k: I am looking for C++ example of string features (with different size) and (multiclass) kernel learning. I know you are pretty busy, I you have time to share an example it would be great. Thank you in advance.17:52
-!- blackburn [~qdrgsm@] has quit [Ping timeout: 272 seconds]17:52
-!- harshit_ [~harshit@] has joined #shogun17:57
-!- n4nd0 [] has joined #shogun17:57
gsomixn4nd0, yo17:58
n4nd0gsomix: hey!17:58
gsomixn4nd0, how are you?17:58
n4nd0gsomix: I am fine, what about you?17:58
n4nd0and hey to all people :)17:59
gsomixn4nd0, fine. I'm waiting for 19UTC.18:00
n4nd0gsomix: I think we all are!18:01
n4nd0gsomix: are you nervous?18:01
PhilTilletgsomix, join the club :p18:01
n4nd0I am getting more and more nervous :O18:01
gsomixn4nd0, a little.18:01
pluskidboys, keep relax, since no matter what we do *within the several hours*, we can not change the (currently unknown) result :p18:03
pluskidgood night all, hope to see good news when I wake up18:09
-!- pluskid [] has quit [Quit: Leaving]18:09
n4nd0pluskid: good night and good luck!18:10
ted__sonney2k: Is there any example of using string features with Alphabet [0,255] ?18:23
@sonney2kgsomix, dont' wait code :)18:52
gsomixsonney2k, moin. what do you mean?18:54
@sonney2kgsomix, SGVector is waiting for you :)18:55
gsomixsonney2k, ah, I know.  I'm doing now a small patches for wiking and blackburn.18:56
CIA-64shogun: Soeren Sonnenburg master * r9cd129d / src/shogun/multiclass/MulticlassOneVsRestStrategy.cpp :18:56
CIA-64shogun: Merge pull request #481 from pluskid/multiclass-fix18:56
CIA-64shogun: Fix a bug in OvR strategy: inc counter in wrong point -
CIA-64shogun: Chiyuan Zhang master * r0814c83 / src/shogun/multiclass/MulticlassOneVsRestStrategy.cpp : Fix a bug in OvR strategy: inc counter in wrong point -
@sonney2kgsomix, SGVector is more important I think - lets hope blackburn knows that18:57
wikinggsomix: \o/18:57
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has joined #shogun19:06
blackburnsonney2k: yes but it needs clarification19:08
blackburnhey guys less than two hours left19:10
PhilTilletyes ^^'19:11
-!- wiking_ [~wiking@huwico/staff/wiking] has joined #shogun19:12
-!- wiking [~wiking@huwico/staff/wiking] has quit [Ping timeout: 246 seconds]19:15
-!- wiking_ is now known as wiking19:15
@sonney2kblackburn, clarification in which respect?19:15
blackburnsonney2k: explain to gsomix more what to do19:17
@sonney2kgsomix, I don't know what to explain actually ... is there anything in particular that is unclear?19:18
@sonney2kI mean just continue to use SGVector& in arguments19:18
@sonney2kblackburn, but the SGVector& stuff is not yet complete or gsomix?19:19
blackburnah this transition should be pretty clear19:19
gsomixhuh, I had planned to discuss it tomorrow. :)19:19
gsomixbut ok19:19
blackburnsonney2k: ok then19:19
@sonney2kgsomix, so please explain :)19:20
gsomixsonney2k, hmm, I already did almost work for SGVector -> SGVector& in arguments, isn't it?19:21
gsomixI am still wondering how to return SGVector from methods.19:22
@sonney2kgsomix, SGVector& except if you return a new sgvectro19:22
@sonney2kthen just ^19:22
@sonney2kgsomix, ok?19:23
gsomixsonney2k, it's clear for me. But I worry about how it is then transferred to ref/unref.19:24
-!- harshit_ [~harshit@] has quit [Quit: Leaving]19:24
gsomixsonney2k, ok, I'll do it in the next few days.19:25
@sonney2kgsomix, we do the ref/unref business after this is done19:25
@sonney2kgsomix, note that this will likely cause some errors that we cannot fix before we have ref/unref stuff19:26
@sonney2kgsomix, if you want to know about ref/unref stuff - let us just copy the functions we have in SGObject for now19:26
gsomixsonney2k, well I think step-by-step strategy is correct.19:27
@sonney2klater on we should investigate whether we add another base class to derive SGVector / SGObject from19:27
@sonney2kgsomix, btw, the same task then for SGMatrix/SGStringList/SGSparseVector ...19:27
@sonney2kSGNDArray .. you name19:27
n4nd0sonney2k: a question, why to return SGVector and not a reference?19:32
n4nd0I mean what you said to gsomix ^19:32
@sonney2kn4nd0, if he allocates a new vector within a function (as local variable) - he cannot return a reference to it19:33
@sonney2ksince the reference will be invalid19:33
@sonney2k(object de-allocated when the function returns)19:34
n4nd0wouldn't it better to create the vector to return in the heap?19:34
gsomixblackburn, check a git.19:34
n4nd0some stuff ... = new SGVector<...more stuff19:35
blackburngsomix: nice19:35
gsomixn4nd0, I think it's not good at now (without ref/unref).19:37
n4nd0sonney2k: well actually SGVector is kind of light, a pointer and a couple of more things so probably the return by value shouldn't be much worse than returning a reference19:37
n4nd0gsomix: do you prefer return by value as well?19:37
CIA-64shogun: Soeren Sonnenburg master * r3007eb5 / src/shogun/io/StreamingAsciiFile.cpp :19:37
CIA-64shogun: pure whitespace changes19:37
CIA-64shogun: - convert file to use tabsize 4 -
CIA-64shogun: Soeren Sonnenburg master * r098c0cc / src/shogun/lib/v_array.h :19:37
CIA-64shogun: delete v_array's begin on exit19:37
CIA-64shogun: This hopefully fixes the memory leak observed in19:37
CIA-64shogun: -
gsomixn4nd0, it is a necessity, I mean19:38
n4nd0gsomix: I wondered if return by reference was an option19:43
@sonney2kn4nd0, yes the object is alloc'd on heap however just *once* since RVO is used
@sonney2kn4nd0, there is no option19:44
@sonney2kn4nd0, see and the first answer19:45
n4nd0sonney2k: ok, thank you19:45
blackburnsonney2k: any preliminary date for kickoff?19:47
@sonney2kn4nd0, I tried to fix that leak in streaming features - can you check if it still occurs?19:48
@sonney2kblackburn, some people have ECML deadline tomorrow morning so not before that :D19:48
@sonney2kmaybe wednesday or some other day this week19:49
blackburnsonney2k: yeah chris has19:49
@sonney2kyeah he is using shogun :)19:49
blackburnwhat is the relation?19:49
n4nd0sonney2k: I'll check it19:52
CIA-64shogun: Evgeniy Andreev master * r15afb6f / (14 files in 5 dirs): knn and conjugate to multiclass -
CIA-64shogun: Soeren Sonnenburg master * rc46d1c0 / (14 files in 5 dirs):19:59
CIA-64shogun: Merge pull request #482 from gsomix/to_multiclass19:59
CIA-64shogun: move KNN and ConjugateIndex to shogun/multiclass -
n4nd0I think qda could me moved to shogun/multiclass too ^20:00
n4nd0be moved, sorry20:00
-!- harshit_ [~harshit@] has joined #shogun20:01
@sonney2kn4nd0, send a pull request20:02
@sonney2kn4nd0, btw didn't you also convert LDA to multiclass?20:02
@sonney2kblackburn, that reminds me - didn't we want to have CMulticlassLabels?20:03
n4nd0sonney2k: no, I didn't do anything to LDA20:03
n4nd0sonney2k: once, even before QDA was done, it was commented though20:03
blackburnsonney2k: yes20:04
blackburnsonney2k: we will do it once20:04
@sonney2kn4nd0, too bad - that's really something we should have if we have multiclass enabled QDA20:05
n4nd0sonney2k: just for completeness or is there something else?20:06
@sonney2kn4nd0, multiclass LDA is certainly much more robust than QDA :)20:07
@sonney2kso it is really a very useful baseline method20:07
@sonney2kmuch faster than SVM on low-dim data sets and even optimal when you can assume gaussian distribs20:07
blackburnsonney2k: btw have you seen that SLEP supports some logistic regression even?20:09
blackburnyes that package I am going to integrate20:10
@sonney2kbut we have logistic regression in liblinear already?!20:10
blackburnsonney2k: why not to provide alternative?20:11
@sonney2kI just don't understand your excitement20:11
blackburnsonney2k: it is library containing a lot of L1 and L1/L2 things20:12
n4nd0sonney2k: I quite don't understand why QDA may fail somewhere LDA successes20:14
n4nd0I thought of QDA as a generalization, within class covariances need not to be equal, so I assumed it would give better results20:15
n4nd0even if teh covariances are the same for all the classes, case where LDA would perform fine too20:15
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]20:16
@sonney2kn4nd0, estimating more parameters is difficult - so for noisy / not a lot of data LDA will work better20:25
n4nd0sonney2k: ok, intuituively I understand too that QDA may be more prone to overfitting20:26
* sonney2k is bored20:31
gsomix27min :]20:32
n4nd0sonney2k: double free or corruption in muticlasslinearmachine example, pretty ugly stack trace :S20:33
@sonney2kn4nd0, paste :)20:33
-!- nickon [] has joined #shogun20:36
-!- vikram360 [~vikram360@] has quit [Ping timeout: 260 seconds]20:39
-!- blackburn [~qdrgsm@] has joined #shogun20:41
n4nd0sonney2k: gdb trace is pretty cool too20:41
n4nd0full of20:41
n4nd0[New Thread 0xb6345b70 (LWP 22769)]20:41
n4nd0[Thread 0xb6345b70 (LWP 22769) exited]20:41
blackburn15 in room20:41
n4nd0blackburn: do you mean 15 people there?20:41
n4nd0blackburn: oooh! you have followers man20:42
n4nd0yeah, I don't know :P20:42
n4nd0I thought you were all waiting for GSoC's answer20:43
PhilTilletn4nd0, GSoC answer? What are you talking about? :p20:43
wikinghahah the crowd gets excited ... ;)20:43
PhilTilletwhat is GSoC? :o20:43
wikingi bet gsoc site will be flooded20:43
PhilTilletlike last year20:43
PhilTilletbut as far as I remember the result is also send by mail20:44
wikinghehehe i wonder why don't they do a caching20:44
wikingand just do a simple precompiled html and that's it20:44
* emrecelikten starts biting nails20:44
n4nd0sonney2k: oh wait, maybe I was handling wrong SG_UNREF in my example20:45
blackburnare you fskin' ready? ;)20:49
PhilTilletf5 spree20:50
blackburnno way to see that even after 5 minutes of deadline20:51
PhilTilletyes I know20:52
PhilTilletat first not all the names are loaded anyway20:52
harshit_Is it gonna be same as the organizations names were released ?20:53
PhilTilletin principle there's also an e-mail20:53
blackburnemail is pretty late20:54
PhilTilletlast year I had the e-mail 5mins before the deadline20:54
PhilTilletbut I guess it was some bug20:55
@sonney2kn4nd0, can you be more specific?20:55
@sonney2kor attach a gdb bt20:55
@sonney2kI cannot make sense of that20:55
n4nd0sonney2k: I think the double free was caused by SG_UNREF I am doing and shouldn't be done20:56
@sonney2kwhich one?20:56
n4nd0in any case I think that to use more SG_UNREF shouldn't cause that20:56
n4nd0in the example classifier_multiclasslinearmachine20:57
n4nd0the one with which I detected those memory leaks20:57
emreceliktenThe emails are going out20:57
n4nd0well good luck to all guys!21:00
harshit_good luck guys21:00
PhilTilletgood luck :)21:01
blackburnsonney2k: you are free to announce I think ;)21:01
harshit_site is down :(21:02
wiking"Error retrieving data: please refresh the list or the whole page to try again"21:02
wikingyeey for google :)21:02
emreceliktenI'm in too from another project21:03
emreceliktenBest of luck to you guys21:03
n4nd0emrecelikten: congrats!21:03
harshit_emrecelikten, congrats !21:04
emreceliktenThanks guys :)21:04
blackburnany of you received email?21:04
emreceliktenI did21:05
PhilTilletI didn't21:05
blackburnheheh melange is really SLOW21:05
emreceliktenThe servers experience something like delta(t) * e^(-at) now I guess21:06
PhilTilletit loaded about 500 students (no shogun :p) and then crashed21:06
harshit_PhilTillet, same here !21:06
PhilTilletsame again21:09
PhilTilletafter 400, crash!21:09
wikingthey have only the first 100 loaded21:12
blackburnoh mailing list will be full of crazy stuff again21:13
-!- karlnapf [] has joined #shogun21:13
harshit_blackburn, why ?21:14
karlnapfhi all21:14
PhilTillethi karlnapf  :p21:15
blackburnharshit_: gsoc mailing list is full of stuff every year21:15
karlnapfPhilTillet, hi there21:15
blackburnkarlnapf: hey21:15
karlnapfblackburn, hi21:15
blackburnkarlnapf: received email? ;)21:15
karlnapfblackburn, no not yet, but just turned on computer21:16
karlnapfhave been vomiting all day :(21:16
blackburnare you ok?21:16
karlnapfnot really21:16
-!- vikram360 [~vikram360@] has joined #shogun21:16
blackburnI hope you will recover soon21:17
karlnapfblackburn, thanks21:17
blackburnI can recall last year Soeren disappeared on the announcement too hehe21:17
n4nd0karlnapf: yeah, take care ;)21:17
karlnapfn4nd0, thanks21:18
emreceliktenIt might a common thing, because my mentor disappeared too21:18
karlnapfI only got a mail fro carol21:18
blackburnkarlnapf: which mail?21:19
karlnapfHi GSoC Students,21:19
karlnapfAs with every year around this time, we are about to add new students to this mailing list in a....21:19
karlnapfthe melange site seems to be down :)21:19
blackburnyes tends to21:19
PhilTillet774 loaded and no shogun o_o21:20
wikingthe first 500 does not contain shogun related proj21:20
wikingPhilTillet: :D21:20
blackburnno shogun at all!21:20
wikingwell there were 4k applications21:20
wikingi don't know how many of them are going to get accepted21:21
n4nd0PhilTillet: you can load quite a lot man, are you using GPU in the browser too? :P21:21
blackburnwiking: ~1.1K21:21
PhilTilletn4nd0, XD21:21
PhilTilletbut it stopped at 77421:21
harshit_I have a different case all together, cant even open melange site :(21:21
PhilTilletno more loading21:21
n4nd0harshit_: I am mostly there too21:22
@sonney2kwhich site are you all trying to load?21:22
shogun-buildbotbuild #516 of python_modular is complete: Success [build successful]  Build details are at
blackburnit should be a nice DDoS attack - to tell gsoc students accepted project are announced on site X21:24
PhilTilletGSoC accepted projects are on universalmusic.com21:25
PhilTilletonce again, stopped at 77421:25
PhilTilletand no shogun!21:26
blackburnI have 100 still21:26
emreceliktenI'm trying too, for you guys :)21:26
wikingPhilTillet: i guess they are inserting lines by lines21:26
-!- puffin444 [230bf329@gateway/web/freenode/ip.] has joined #shogun21:27
PhilTilletwiking, well, seems like i've loaded everything though21:27
wikingPhilTillet: yeah 774 and it gets stuck21:27
PhilTilletI think we'll have to wait some additional time21:27
blackburnlets have some vodka instead21:27
blackburnwho's in?21:28
karlnapflol, not me :)21:28
PhilTilletI don't drink :p21:28
blackburnPhilTillet: bad way21:28
puffin444Well this is quite suspensful21:30
PhilTilletit's beyond suspens21:31
karlnapfhey guys, i will check back later in an hour or so21:31
puffin444Yeah it's sort of torturous :)21:31
puffin444I wonder how google updates these. Do you think my dashboard/email would tell me first?21:32
blackburnpuffin444: IIRC that list is first21:32
wiking774 is the magic number :D21:33
n4nd0400 here21:33
PhilTilletagain 77421:33
puffin444How many in total is it this year? 1000?21:33
PhilTillet774 + 921:33
blackburnpuffin444: should be ~1.1K again I think21:34
harshit_774 , no shogun :(21:34
emreceliktenYep, I'm stuck at 774 too21:34
puffin444I suspect they haven't released all of them at the same time for some reason.21:35
puffin444They did the same thing with the sponsoring organizations21:35
puffin444Some didn't pop up until days after the release data :o21:35
blackburnpuffin444: no with orgs the reason was different21:35
blackburnorganizations had to complete their profiles21:35
PhilTilletI think it's just a bug of theirs21:35
blackburnnot a bug21:36
PhilTilleta feature?21:36
blackburnwith orgs - yes21:36
PhilTilleti was talking about the students :p21:37
blackburnas for students no idea what is going on hehe21:37
PhilTilletmaybe it's a feature too21:37
PhilTilletthe number is evolving21:38
blackburnterribly slow thing21:39
puffin444Did anyone receive that "none" email a few days back?21:39
blackburnI didn't21:39
PhilTilletI didn't either21:39
emreceliktenI did21:40
emreceliktenIt was a positive one21:40
blackburn811 loaded21:40
blackburnno shogun ;)21:40
PhilTillethaha blackburn, new record!21:40
PhilTilleti'm refreshing again21:40
PhilTilletmaybe i'll get a better score21:40
-!- n4nd0 [] has quit [Quit: Lost terminal]21:40
-!- n4nd0 [] has joined #shogun21:41
puffin444emrecelikten: what do you mean be "positive"?21:41
PhilTillet827 :p21:42
PhilTilletno shogun though21:42
puffin444Do you know how many slots shogun had this year?21:42
emreceliktenpuffin444: There were two types of emails as far as I know. The "thank you" ones that corresponded to negative ones and the "congratulations" ones. Some guy claimed that he got "pick one of the projects" email21:42
puffin444This happened a few days ago?21:44
emreceliktenpuffin444: Yes, Friday I think21:44
blackburnhehe it becomes faster21:45
emreceliktenIt's like an auction21:45
@sonney2klook 1212 is the number21:45
wikingok so 1212 :>21:45
blackburnbig #21:46
puffin444I just got a rejection email from another organization21:46
wikingfyi: 5 500 * 1 212 = 6 666 00021:46
wikingfunny number... i wonder whether it was intentional ;)21:47
PhilTilletno shogun21:47
blackburnwiking: in russia it would be 300-400 - you have to stole 60% :D21:47
emrecelikten950 here21:47
emreceliktenNo shogun21:47
wikingblackburn: :DDD21:48
PhilTillet982 :p21:48
emreceliktenYou guys seems to be the last :P21:48
* wiking refreshing :)21:49
puffin444Maybe I'll never see the decision. I'll just die from a heart attack first :P21:49
PhilTilletsame for me21:49
emreceliktenThis is cruel21:49
n4nd0emrecelikten: yeah it is21:50
PhilTilletno shogun21:50
emreceliktenPhilTillet: Same21:50
harshit_1000 , no sign of shogun :(21:50
blackburnwe are elite21:50
wikingthey are just at R21:51
wikingso soon21:51
PhilTillethow do you know? XD21:51
PhilTilleti've got some T loaded21:51
PhilTilletand it will be hell cause Shogun's list will progressively appear21:51
wikingPhilTillet: yeah21:51
wikingPhilTillet: but then now suddenly between 900-100021:51
wikingnew P and R projects got in21:51
blackburnno shogun21:51
emreceliktenn4nd0: You're in21:52
emreceliktenIf you're Fernando Garcia21:52
n4nd0emrecelikten: really?21:52
blackburnhehe congrats21:52
wikingcongrats n4nd021:52
emreceliktenpluskid in21:52
n4nd0well Iglesias should be in between21:52
karlnapfnice :), congrats21:52
emreceliktenSergey Lisitsyn in21:52
emreceliktenMichal in21:52
wikingplsukid is in21:52
n4nd0I wan to see that guys!21:52
emreceliktenJacob Walker in21:52
emreceliktenvigsterkr in21:52
puffin444Oh wow!21:52
n4nd0still 1024 here21:53
PhilTilletsergey in21:53
emreceliktenJust filter it ;)21:53
wikinggot a mail :D21:53
n4nd0fuck I was not selected for multiclass :O21:53
karlnapfjust got the mail :)21:53
blackburnhehe got the mail21:53
wikingn4nd0: heheheh i'm more happy that you'll do that21:54
n4nd0GOT SO!21:54
emreceliktenCongratulations to you guys!21:54
PhilTilletgot a negative mail21:54
blackburnPhilTillet: I am sorry21:54
harshit_-ve :(21:54
wikingPhilTillet: :((((21:54
puffin444PhilTillet: Sorry PhilTillet :( :(21:54
emreceliktenPhilTillet: Next year mate :) I was rejected last year too, I know how it feels21:54
n4nd0PhilTillet: sorry man21:54
PhilTilletemrecelikten, I was rejected last year too :p21:54
wikingn4nd0: when are you going to start the SO?21:55
emreceliktenPhilTillet: I was being rejected for two years :P21:55
n4nd0n4nd0: asap :)21:55
n4nd0ups, again talking to me21:55
wikingn4nd0: hahahah21:55
wikingn4nd0: i'll need it21:55
wikingasap :D21:56
n4nd0wiking: cool, we can probably exchange some thoughts about it :)21:56
wikingyeah sure21:56
wikingare you going to use libqp?21:56
blackburnPhilTillet: okay actually we are just not ready for gpuing..21:56
n4nd0wiking: yes, that was the first option I thought of21:57
wikingso then you are the third one21:57
PhilTilletblackburn, why so?21:57
wikingsince i'm going to use it and michal as well21:57
blackburnharshit_: I am very sorry - the problem is that Soeren is not capable to mentor two students21:57
wikingso i've started to talk with sonney2k about how to refactor some stuff21:57
blackburnPhilTillet: no real mentor for that21:57
blackburnand no base for that actually21:57
wikingn4nd0: i'll push this week something so your comments are more than welcome21:58
PhilTillethmmm ok21:58
n4nd0wiking: thank you21:58
harshit_blackburn, can understand :|21:58
blackburnI really hope you are not upset with anything21:58
n4nd0wiking: I'll start to be much more active again after Saturday21:58
n4nd0damned exams21:58
PhilTilletblackburn, I'm not ^^21:58
wikingn4nd0: no worries21:58
blackburnPhilTillet: that's nice :)21:59
blackburnPhilTillet: we are still welcome your contributions22:00
blackburnharshit_: your contributions are welcome too I hope you are not very frustrated!22:01
PhilTilletblackburn, I'll probably do the project on my own :) that's why I sent only one application, there's no point in working in something i'm not fully interested in ;22:02
gsomixhey, congrats guys22:02
harshit_blackburn, no, actually a lil. but still can work on NN for shogun , and my research22:02
@sonney2kPhilTillet, harshit_ let me explain a bit22:03
blackburnyou all have done really a lot of work22:03
harshit_blackburn, but summer could be better with gsoc22:03
@sonney2kI am/was the only mentor for your projects and the one we selected gsomix for22:03
@sonney2kI would have loved to take you both too but I simply cannot handle >1 student22:04
@sonney2kso we decided for the task we needed most - and that is some restructuring / accessibility cleanups22:04
@sonney2kthere was nothing wrong with your proposal or involvement or anything22:04
PhilTilletsonney2k, we know :)22:05
@sonney2kit really is about being at our capacity limits22:05
PhilTilletsonney2k, nothing prevents us from contributing without being paid I guess :)22:05
harshit_sonney2k, no worries :)22:05
@sonney2kand there is nothing I could do about it so late22:05
@sonney2kin the end I *never* expected that we get 9 slots22:05
@sonney2kif I knew that beforehand I would have asked a lot more people :)22:06
@sonney2kanyways whole wikimedia foundation has 8 (or 9?) slots22:06
@sonney2kso just to give you some impression how many we received22:06
blackburnwe keep receiving strong gsoc support :)22:07
@sonney2kPhilTillet, harshit_ of course we welcome your contributions22:07
harshit_sonney2k, sure ..!22:07
@sonney2kkarlnapf, do you have a homepage I can link to22:08
@sonney2kpuffin444, ^ ?22:08
@sonney2kand does any of the students here blog? I can put you on shogun planet if you want...22:09
puffin444sonney2k: Not at the current moment. I might be able to make one if you need one though.22:09
blackburnsonney2k: do you think introduction @ mailing list is a good idea?22:09
@sonney2knot required - I just wanted to give it some traffic - hehe22:09
@sonney2kblackburn, I think so22:09
n4nd0sonney2k: I want to take advantage of gsoc and start doing it22:09
blackburnhey guys22:10
@sonney2kblackburn, I totally forgot how we did it last year22:10
blackburnsonney2k: just some introduction mail nothing special I think22:10
@sonney2kblackburn: every student had to send some short email introducing himself + project right?22:10
blackburnn4nd0: puffin444: wiking: karlnapf: gsomix: could you please do this ^?22:10
blackburnkarlnapf: however you are well-known22:10
@sonney2kbut still22:11
blackburnyes there is a need to introduce your project22:11
@sonney2kyou can look up in the mailinglist archive how these things look like22:11
n4nd0all right22:11
blackburnjsut some short introduction22:11
puffin444blackburn: Yeah sure.22:11
@sonney2kthe other more annoying job is to send a weekly update (from the official beginning to end of gsoc)22:11
blackburnand some feature plan would be nice22:11
@sonney2kso another short email22:12
@sonney2kcannot be said often enough22:12
blackburnattach pics!22:12
wikingblackburn: on shogun mailing list?22:13
blackburnwiking: right22:13
wikingalrighty i'll do it tomo morning22:13
@sonney2kwiking, thanks22:13
karlnapfsonney2k, yes I just created a hp, but its not completely ready, will send you the link this week22:13
karlnapfblackburn, I will send an introducionary mail :)22:14
@sonney2kkarlnapf, your old blog is dead too it seems22:14
blackburnsonney2k: do you understand things will go even more crazy next days?22:14
blackburnI'll send too22:14
@sonney2kblackburn, why should they?22:14
blackburnsonney2k: once active development begins it would become crazy I am sure :)22:14
karlnapfsonney2k, yes i took it offline to replace it by something nicer22:15
* wiking wonders about the average number of users on #shogun after a month ;)22:15
@sonney2kwiking, let me guess 8 :D22:15
wikingsonney2k: + the bots!22:15
blackburnn4nd0: your task is really important and you will definitely collaborate with wiking and michal!22:16
n4nd0blackburn: sounds very exciting22:16
wikingyeah i'll most def will need a SO asap :>>22:16
@sonney2kn4nd0, that is why you were chosen ... we have high trust in you22:16
wikingn4nd0: SO WORK!!!22:16
n4nd0wiking: :D22:16
n4nd0sonney2k: thank you!!22:16
blackburnhowever I have high trust in pluskid too ;)22:16
wikingn4nd0: but anyhow when u start do define the new labels class let me know22:17
n4nd0wiking: ok22:17
@sonney2kwell actually we know that you all can do the task - that is why you got selected ;)22:17
blackburnokay congrats you guys22:18
blackburnI have to get back to my shity deals22:18
n4nd0blackburn: congrats to you too man22:18
@sonney2kblackburn, vodka?22:18
blackburnsonney2k: not yet22:18
blackburnsonney2k: yes I deal vodka22:18
wikingblackburn: :DDD what's your poison ?:)22:18
@sonney2kvodka blackburn the true soul of shogun :D22:19
blackburnwiking: my poison is water22:19
gsomixblackburn, `prostavlyaisya'. :]22:20
blackburngsomix: you should22:20
puffin444So just to clarify: An introductory email about me and my project tomorrow on the Shogun Mailing List, and possibly a link to a homepage?22:21
gsomixblackburn, you should in the past year.  so where? :]22:21
blackburnpuffin444: yes22:21
blackburngsomix: heh ok22:22
blackburnokay guys congrats again and see you later22:23
gsomixI should review accepted projects.22:23
gsomixit's very interesting22:23
blackburnall 1212?22:23
blackburnwill take a night22:23
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]22:23
gsomixblackburn, aha22:23
gsomixwow, `Category Theory Module' for SymPy.22:27
emreceliktengsomix: Nice22:29
n4nd0see you later guys22:29
emreceliktenn4nd0: Bye!22:29
shogun-buildbotbuild #731 of cmdline_static is complete: Failure [failed test_1]  Build details are at  blamelist:, gsomix@gmail.com22:29
puffin444ok guys, you'll be hearing from me tomorrow. See you later.22:32
gsomixshogun-buildbot, aww, I forget about it.22:32
shogun-buildbotbuild #709 of r_static is complete: Failure [failed test_1]  Build details are at  blamelist:, gsomix@gmail.com22:33
-!- puffin444 [230bf329@gateway/web/freenode/ip.] has left #shogun []22:34
-!- n4nd0 [] has quit [Ping timeout: 246 seconds]22:34
wikingshogun-buildbot what's up :)22:36
gsomixwiking, my bag22:36
shogun-buildbotbuild #710 of octave_static is complete: Failure [failed test_1]  Build details are at  blamelist:, gsomix@gmail.com22:37
@sonney2kgsomix, what goes wrong?22:41
shogun-buildbotbuild #697 of python_static is complete: Failure [failed test_1]  Build details are at  blamelist:, gsomix@gmail.com22:41
gsomixsonney2k, I forget about libshogun tests.22:42
@sonney2kgsomix, these are not even tests but just examples22:42
gsomixbut I do not know what with classifier_multiclasslinearmachine22:42
gsomixsonney2k, aha22:42
@sonney2kthey are supposed to be run as tests later but we didn't do it yet22:43
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]22:54
gsomixsonney2k, I corrected conjugate example. check github22:58
-!- harshit_ [~harshit@] has quit [Ping timeout: 244 seconds]22:58
-!- PhilTillet [] has left #shogun ["Leaving"]22:58
CIA-64shogun: Evgeniy Andreev master * r51e0223 / examples/undocumented/libshogun/classifier_conjugateindex.cpp : small fix in example -
CIA-64shogun: Soeren Sonnenburg master * r6ab1607 / examples/undocumented/libshogun/classifier_conjugateindex.cpp :22:59
CIA-64shogun: Merge pull request #483 from gsomix/to_multiclass22:59
CIA-64shogun: Small fix in example -
gsomixsonney2k, I need to go. Thanks a lot for your choose.23:01
gsomixgood night guys23:01
-!- wiking [] has joined #shogun23:04
-!- wiking [] has quit [Changing host]23:04
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun23:04
-!- harshit_ [~harshit@] has joined #shogun23:10
-!- harshit_ [~harshit@] has quit [Client Quit]23:15
@sonney2knite everyone23:19
-!- karlnapf [] has quit [Quit: Leaving.]23:26
-!- av3ngr [~av3ngr@] has joined #shogun23:50
-!- nickon [] has quit [Quit: ( :: NoNameScript 4.22 :: )]23:54
--- Log closed Tue Apr 24 00:00:19 2012