Open in new window / Try shogun cloud
--- Log opened Mon Aug 13 00:00:17 2012
wikingblackburn: yes00:18
wikingubuntu's package doesn't ship with gmock-config :(00:19
wikingso we need to do soemthing else in ./configure script to detect and add compile flags for gmock on ubuntu00:19
blackburnwhat is the solution? install by hand?00:19
wikingheiko solved it by installing gmock by hand00:19
wikingbut as i said in the mailing list00:19
wikingthat's not a really good solution ;)00:19
blackburnI think it is rather ok - we (developers) can install it and only we do need it00:21
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]02:37
shogun-buildbotbuild #53 of nightly_none is complete: Failure [failed compile]  Build details are at
-!- av3ngr [av3ngr@nat/redhat/x-xdzltetycxkppxhs] has joined #shogun04:04
-!- n4nd0 [] has joined #shogun08:47
-!- uricamic [~uricamic@2001:718:2:1634:7db4:863b:e25:317c] has joined #shogun08:55
-!- n4nd0 [] has quit [Ping timeout: 256 seconds]10:00
wikingbusy bees everywher? :)10:56
-!- av3ngr [av3ngr@nat/redhat/x-xdzltetycxkppxhs] has quit [Quit: That's all folks!]11:46
wikinganybody around?12:44
wikingahhahahahaha blackburn u've fcuked me :)12:47
-!- alexlovesdata [] has joined #shogun13:53
-!- blackburn1 [~blackburn@] has joined #shogun14:01
-!- alexlovesdata [] has left #shogun []14:02
-!- alexlovesdata [] has joined #shogun14:03
blackburn1uricamic: hey14:05
blackburn1uricamic: can I merge your PR now?14:05
uricamicblackburn1: hi, I hope so14:05
blackburn1uricamic: I mean did you work out things Soeren pointed?14:05
blackburn1okay I see14:06
uricamiceverything should be fixed according to comments14:06
CIA-39shogun: Sergey Lisitsyn master * r8c987fd / (18 files in 4 dirs): Merge pull request #699 from uricamic/BM_SOL_EXAMPLE (+8 more commits...) -
blackburn1uricamic: btw I had a question14:06
wikingblackburn1: mergmeme :D14:06
uricamicblackburn1: go on :)14:07
blackburn1wiking: yeah will glance it over once again in 5 mins14:07
blackburn1uricamic: in risk function there is float64_t* R14:07
blackburn1is it just a pointer to one float - risk value?14:07
uricamicwiking: there is one important change in risk function you should be aware of14:07
uricamicblackburn1: yep14:07
blackburn1uricamic: actually may be we could change it not to void but float64_t return type14:08
blackburn1to reduce num of parameters14:08
uricamicyou mean to directly return the value of risk from that function?14:09
uricamicyep, that sounds reasonable14:09
blackburn1uricamic: that would make director risk function easier to use14:09
wikinguricamic: ah cool then i'll wait for the merge14:09
wikingi mean your merge... so i change mine as well14:09
blackburn1now one have to do R[0] = ...14:10
uricamicblackburn1: ok, so I will change it now14:10
blackburn1heh I merged too fast14:10
uricamicok, next PR :)14:10
blackburn1uricamic: btw how fast it is for MNIST?14:10
shogun-buildbotbuild #319 of deb1 - libshogun is complete: Failure [failed compile]  Build details are at  blamelist: Michal Uricar <>14:11
shogun-buildbotbuild #320 of deb1 - libshogun is complete: Success [build successful]  Build details are at
uricamicwell, to be honest, the risk function could be written in much better way (some precomputations), but now it is like 1s per iteration on whole MNIST database (i.e. 60k examples, feature vector dimensionality 784)14:12
shogun-buildbotbuild #329 of deb3 - modular_interfaces is complete: Failure [failed compile libshogun]  Build details are at  blamelist: Michal Uricar <>14:13
blackburn1uricamic: how many iterations does it require to converge to some consistent solution?14:13
uricamiccomparing with exactly same implementation in MATLAB it is like 8x faster14:13
uricamicdepends on regularization constant lambda, but I think several hundreds14:13
blackburn1uricamic: hmm it sounds like it is not really slower than linear crammer-singer14:14
blackburn1I mean liblinear and libocas implementations (they take similar time on 40K examples/ 1K features dataset)14:15
uricamicblackburn1: but I think it can be made faster with better risk function, doing some clever precomputations14:16
blackburn1I see14:16
shogun-buildbotbuild #222 of bsd1 - libshogun is complete: Failure [failed compile]  Build details are at  blamelist: Michal Uricar <>14:16
blackburn1wiking: so shall I wait you to rebase against uricamic changes?14:18
shogun-buildbotbuild #223 of bsd1 - libshogun is complete: Success [build successful]  Build details are at
wikingblackburn1: let me try to rebase14:23
wikingmm rebase went fine i'll try to compule14:24
wiking*compile if that works then u can just merge it straight away14:24
wikinghehe yeah no i need to fix some shit14:26
blackburn1~150 warnings left14:26
blackburn1I will kill them all soon14:26
shogun-buildbotbuild #330 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Sergey Lisitsyn <>, Michal Uricar <>14:30
wikingblackburn1: what should be done about the libocas update?14:33
blackburn1wiking: I didn't check yet14:34
wikingblackburn1: there was this thing with the multiclass svm solver of libocas... the -1 or no -114:34
wikingi guess we shouldn't do -1 as in previous cases there was no -1 ;)14:34
wikingprevious cases = previous version of libocas14:35
* wiking still compiling... stupid modular interface ;)14:37
wikingi hate this14:39
uricamicIs it possible to get PyCharm license for shogun?14:39
blackburn1I had a crack somewhere14:40
wikingok i'll redo this pull request14:41
uricamic:) but it could be possible to obtain free license for shogun developers :)14:41
wikingotherwise it'll be just a mess14:41
uricamicwhich I guess is cleaner solution14:41
blackburn1wiking: redo??14:41
blackburn1uricamic: yeah just apply on their site as shogun developer14:41
wikingblackburn1: do a new PR14:42
wikingblackburn1: that'll be nicer14:42
uricamicblackburn1: ok, and should I apply it just for me, I mean who should be the contact person?14:43
blackburn1uricamic: I think it should be you14:44
uricamicok, thx14:44
blackburn1Please make sure that you meet the following criteria:14:44
blackburn1You are the project lead or a committer14:44
blackburn1You have been working on your open source project for a minimum of 3 months14:44
blackburn1Your community is active. This means that you have recent activity in your newsgroups or forums14:44
blackburn1You have an updated News section on your site14:44
blackburn1You release updated builds on a regular basis14:44
blackburn1uricamic: ^14:44
wikingblackburn1: this should work
uricamicblackburn1: I know, I just didn't know if it should be like the leader of the project asks for license for all developers14:45
blackburn1uricamic: first point is not satisfied probably but just try14:45
blackburn1uricamic: I don't get that too14:45
uricamicok, I will just try14:46
blackburn1if they deny lets just ask soeren14:46
uricamicblackburn1: sure :)14:47
blackburn1however I am a commiter so I could try too14:47
wikingok this compiled again14:48
wikingso all's good with that PR14:48
uricamicblackburn1: ok, I have sent it, in few days there should be some response14:51
blackburn1uricamic: I've used pycharm for a while but didn't really like it - do you like it?14:52
uricamicblackburn1: I don't know, I am a python newbie :D but so far it seems nice14:53
blackburn1uricamic: I see14:53
uricamicand one my college, who is a big fan of python likes it a lot14:53
blackburn1uricamic: what do you use to develop with C++?14:54
uricamicnow I am using QtCreator14:54
uricamicsince it seems to cooperate with CMake nicely14:55
blackburn1I see14:55
wikingblackburn1: :)15:01
CIA-39shogun: Sergey Lisitsyn master * rd87c378 / (13 files in 8 dirs): Fixed a few documentation warnings -
blackburn1wiking: I am confused whether we should -1 in ocas15:22
shogun-buildbotbuild #331 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Sergey Lisitsyn <>15:34
wikingblackburn1: just writing the unittest15:37
wikingand we'll see15:37
blackburn1this probably maps 0 to -115:38
wikinglol shogun-unit-test(85744,0x7fff7c412180) malloc: *** error for object 0x7fb913c071b8: incorrect checksum for freed object - object was probably modified after being freed.15:42
CIA-39shogun: Sergey Lisitsyn master * r676cefa / (2 files): Fixed java modular compilation issue with BMRM -
wikingblackburn1: mmmm in multiclass labelling the labels should be: 0,1,2 etc?15:51
wikinghow fast multiclassocas is actually?15:58
wikingsuppose to be fast for a 3 class problem15:59
wiking3 guassians, each having 25 examples and the vector dim is 216:00
wikingmmm this doesn't work :)16:05
wikinganybody ever tested multiclassOCAS?16:07
wikingi see a python exmaple of it :)16:07
blackburn1wiking: yes it works16:08
blackburn1wiking: slower than liblinear but for your case should converge in a few milliseconds I believe16:09
wikingmmm i wonder what i do wrong16:15
wikingas it really doesn't find anything16:15
wikingi mean everything is mapped to 0.016:15
wikingmost probably my bad...16:17
shogun-buildbotbuild #332 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Sergey Lisitsyn <>16:19
wikingblackburn1: ok i've fixed the (c) header... see the last patch in PR16:22
CIA-39shogun: Viktor Gal master * rc87bf5e / (12 files in 4 dirs): [WIP] restructure latent solvers -
CIA-39shogun: Viktor Gal master * rbbb7d9b / (23 files in 8 dirs): Various changes for latent variable learning -
CIA-39shogun: Viktor Gal master * rd3e9250 / src/interfaces/modular/Latent.i : Fix author in copyright header -
CIA-39shogun: Sergey Lisitsyn master * r97b8217 / (27 files in 9 dirs): Merge pull request #709 from vigsterkr/latent -
wikinglet's see if it fails :)16:24
shogun-buildbotbuild #323 of deb1 - libshogun is complete: Failure [failed compile]  Build details are at  blamelist: Viktor Gal <>16:26
wikingah i guess i know how16:26
shogun-buildbotbuild #324 of deb1 - libshogun is complete: Success [build successful]  Build details are at
wikingi've missed the --amend16:27
wikingblackburn1: i've tried now with only 2 labels (0/1) using multiclass ocas, basically the same data set that works for svmocas16:30
wikingand it fails16:30
blackburn1wiking: fails like?16:31
wikingall example is classified to 016:32
wikingso i have 2 gaussians16:33
wikingi mean samples of 2 different gaussians16:33
blackburn1how many iterations?16:33
wikingone is labeled as 0 the other as 116:33
wikingblackburn1: haven't debugged it, just a sec i'll print it16:34
CIA-39shogun: Sergey Lisitsyn master * rdf80e5c / (15 files in 7 dirs): Removed SLEP machine and redundant L1LogisticRegression, rebased FeatureBlockLogisticRegression and fixed missed weight initialization -
wiking CMulticlassOCAS* mocas = new CMulticlassOCAS(C, train_feats, ground_truth);16:36
wiking  mocas->io->set_loglevel(MSG_DEBUG);16:36
wikingmmm doesn't really work :(16:36
blackburn1it outputs information just after training16:37
blackburn1with debug level16:37
wikingyes i know16:37
wikingbut if i do this16:37
wikingand then call mocas->train()16:37
wikingit doesn't print out anything :S16:37
blackburn1then something is broken in your example :D16:38
blackburn1I do not undrestand how can it skip debug print16:40
wiking?![DEBUG] Number of iterations [nIter] = 516:41
wikingheheh ok i've fixed it16:41
wikingshogun wasn't init-ed in a right way16:41
wiking[DEBUG] Number of training errors [trn_err] = 2516:41
wikingthat is funny16:41
wikingout of 50 samples :)))16:41
wikingwhere 25 is 0 and 25 is 116:42
blackburn1what is epsilon?16:42
wikinglemme check what is it16:43
blackburn1no, not really relevant16:43
blackburn1I mean if it was 1e-1 I would believe heh16:43
blackburn1okay then compare16:43
blackburn1with liblinear16:43
blackburn1I am totally sure liblinear works16:43
shogun-buildbotbuild #333 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Viktor Gal <>16:44
wikingoptimization finished, #iter = 16416:44
wikingsame error with liblienar16:45
blackburn1then something is wrong with your features and labels16:45
wikingblackburn1: but it works with SVMOcas16:45
wikingthe same binary case16:45
blackburn1so you set labels to16:46
blackburn10 and 116:46
wikinginstead of -1 and 116:46
wikingeverything is the same16:46
wikingjust the labels are 0 and 1 instead of -1 and 116:46
wikingthe only difference is16:48
wikingline 2916:48
wikingline 3716:48
wikingand line 4016:48
wikingand obviously 4216:49
wikingeverything else is the same16:49
wikingeven the random seed is the same16:49
wikingthus the gaussians are the same16:49
blackburn1that's strange16:50
wikingespecially that not only libocas fails16:51
wikingbut liblinear16:51
blackburn1I just checked on my rsr stuff - liblinear converges to right solution16:51
wikingso i really wonder16:51
blackburn1did you change it to multiclasslabels?16:51
wikingthis is the full test code16:52
wikingand yeah16:52
wikingi'm using CMulticlassLabels16:52
blackburn1wiking: I have one suspect - copy subset..16:57
wikingblackburn1: but that works in case of SVMOcas :)16:57
blackburn1yes because svmocas would work with subset16:58
wikingany fast way to divide the whole set then in another way?16:59
blackburn1I checked code and it should be ok16:59
blackburn1if (i % 2 == 0)      train_idx[j] = i;    else      test_idx[j++] = i;17:00
blackburn1is it correct?17:00
shogun-buildbotbuild #334 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Sergey Lisitsyn <>, Viktor Gal <>17:01
blackburn1wiking: why do you set half of labels and increase j only on test indices?17:02
wikingas u can see j increases every 2nd time in the iter17:04
wikingas you want to divide the whole matrix17:04
wikinginto 2 different sets17:04
blackburn1I don't know - that all sounds strange17:14
blackburn1I wouldn't believe something is wrong both in liblinear and ocas17:15
wikingheheh me neither17:24
wikingi guess something is wrong on my end17:24
wikingit's just weird that it works with svmocas17:25
shogun-buildbotbuild #335 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Sergey Lisitsyn <>17:26
-!- uricamic [~uricamic@2001:718:2:1634:7db4:863b:e25:317c] has quit [Quit: Leaving.]17:44
-!- heiko [] has joined #shogun17:47
wikingheiko: is there a fast way to test whether the features in a matrix are from 2 diff gauss distrib?17:50
heikoyes :)17:51
heikoif you are one-dimensional17:51
heikoand if not: My statistics framework :)17:51
heikowiking, what kind of data do you got?17:51
wikingheiko: the SGMatrix that is created by that function i wrote in DataGenerator17:52
heikoIll take a look17:52
wikingheiko: for some reason i'm having trouble using that and trying to learn a multiclass classifier with the generated data :S17:52
heikooh :)17:52
wikingso now i'm trying to find out what could be the trouble :S17:52
heikomaybe plot it for dim=2?17:52
heikoyou could also fit a Gaussian with ML17:53
heikowhat do you want to create in the method i dont get the comments17:54
wikingheiko: create a matrix17:55
heikowhat kind of gaussians should the samples come from17:55
wikingwhere there are m samples for each of the n gaussians17:55
heikoyou only say that there are multiple ones17:55
heikowhat are their parameters?17:55
wikingheiko: random17:55
wikingheiko: you mean for mu, right?17:56
wikingyes that is randomly generated17:56
heikoso random mean17:56
heikoand random variance17:56
wikingvariance is a diagonal17:56
heikoand from each of these a fixed number of samples?17:56
heikodo all of them have the same probability to generate a sample or are they weighted?17:56
wikingif dim = 217:56
heikoso variance is not random17:57
wikingyep that's not random17:57
heikoI suggest the following: first create a method (or check whether it exists): gaussian_sample(SGVector mean, SGMatrix variance)17:58
heikoand then just call that with random means17:58
wikingheiko: well that's more or less what's happening ;017:58
heikoah thats the Gaussian class right?17:58
wikingGaussian class generates the samples for the given gaussian17:58
wikingok maybe the problem is not linearly separable? :)17:59
heikobtw you dont need to call this get_allocated_matrix stuff if you dont have a target matrix as parameter17:59
wikingi mean the mean is just not good17:59
heikogaussians mixtures are per definition not linearly separable18:00
heikoyou might be lucky but not in general18:00
heikoyou  need a large distance  at least e*std-dev18:00
wikingheiko i guess i got lucky in case of svmocas then18:00
heikoI looked at the method18:01
heikothis mean.random()18:01
heikobreaks it18:01
heikoyou need to ensure that the means are far enough apart from each other18:01
wikinghtat's what i thought now18:01
wikingany ideas18:01
wikinghow can i generate good enough means if i know the dim and the number of gaussians i have18:02
wikingsince i guess all of them should be equally far apart from each other18:02
heikoyeah, dont use random means18:02
heikothe samples are random anyway18:02
blackburn1one must be careful with 0.01 probability left with 3 sigmas18:03
heikoor you could order them in a grid18:03
heikothanks blackburn :)18:03
blackburn1heiko: do I remember that correctly? 0.995 is percentile or so?18:04
heikoyeah :)18:04
heikowiking, just create hypercubes of gaussians18:04
heikoIll send you a picture18:04
wikingheiko: great18:04
blackburn14d cube18:04
wikingheiko: so depending on the dim18:04
wikingi take the 'corners' of the hypercube?18:05
wikingi.e. if dim = 218:05
blackburn1just generate all combinations of vector of size dim consisting +-d18:06
wiking0,0; 0,1; 1,0; 1118:06
blackburn1or (-1,1) (1,1) (1,-1) (-1,-1)18:06
heikowiking, sent18:06
heikowiking, dont need the corners18:06
heikojust do n gaussians per dim18:06
heikoand set distance between them to a large value18:07
heikothe picture I sent is for d=218:07
wikingah ok18:07
wikingyeah because i was about to say what if one wants 10 gaussians but with d=218:07
wikingok so what do you reckon the distance should be?18:07
heiko5 should be enough18:08
blackburn1heiko: I found a way to evaluate OvR things18:08
heikobut dimension wise18:08
heikonot euclidean18:08
heikoblackburn, tell me18:08
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has joined #shogun18:08
wikingheiko: i.e. 0,0; 0,5; 5,0; 5,5; 10,0 etc18:08
heikoyeah for d=218:08
wikingfor d=3 5 would be good i.e. 0,0,0; 0,5,0 etc?18:09
blackburn1heiko: I am adding an options to store all scores in multiclass labels18:09
heikodim=3 would be a 3d cube " filled with" gaussians18:09
blackburn1the multiclassovrevaluation computes things via binaryclassevaluation18:09
wikingdamn i need a grid generator code18:20
heikoblackburn1, custom fonts with python latex are a pain is the a**18:23
heikocomputer modern is quite easy18:23
heikowhat do you think?18:23
blackburn1heiko: python latex?18:23
heikoyeah, latex typo in pytohn plots18:23
heikofor tutorial18:23
blackburn1I used custom fonts there18:24
blackburn1without much pain18:24
heikocould you send me a script with that?18:24
blackburn1pain is to use cyrillic18:24
heikolike a small example where you use the pathpazo18:24
blackburn1you would never imagine how much pain cyrillic brings18:24
heikolol :) ok I stop complaining18:24
blackburn1okay let me find something18:24
blackburn1heiko: all you need is to put a few things into rc18:25
heikodo you have a script?18:25
blackburn1wait a min18:26
wikingheiko: this shoudl work afaik18:26
wiking-               mean.random(0.0, 10.0);18:26
wiking+     ;18:26
wiking+               for (index_t k = 0; k < dim; ++k)18:26
wiking+                       mean[k] = i*grid_distance;18:26
wikingwhere i is the i-th gaussian18:26
wikingso basically it's 0,0... 5,5... 10,10...18:26
blackburn1lines 83,84,8518:27
heikowiking, plot in 2d and 3d then youll see :)18:27
blackburn1for 4d case use the force!18:27
wikingheiko: display_matrix is not working :D18:27
blackburn1heiko: but show won't work18:28
wikingor it does? :)18:28
heikoit does18:28
wikingoh yeah it does i'm just looking at it on a wrong way18:28
blackburn1heiko:  show doesn't work here with latex at all18:28
blackburn1heiko: neither with custom nor default latex fonts18:28
heikoblackburn, this is all annoying18:28
heikowhy isnt there an export button?18:28
blackburn1heiko: export?18:29
heikoi hate this messy stuff! argh18:29
heikosave-as you know18:29
heikoits 201218:29
blackburn1heiko: it is in matplotlib18:29
heikobut it doesnt integrate into latex18:29
heikoIt would be great to let latex do the typo18:29
heikothen everything would look much nicer18:29
blackburn1heiko: in my script latex is used with custom font and saves to .pdf18:30
blackburn1what else do you need? :)18:30
heikoyeah this works, but is it comfortable? :)18:30
heikothats my point18:30
heikojust complaining in general :D18:30
blackburn1comfortable in which way?18:30
blackburn1I think there is a way to make show work too18:31
blackburn1with latex18:31
blackburn1heiko: nicholas wants to normalize train and test data with preprocessor inited with train data18:34
heikoblackburn1, what does that mean?18:34
blackburn1heiko: for example you compute mean of train vectors18:35
blackburn1and subtract it both from train and test vector18:35
blackburn1during the process of evaluation18:35
blackburn1got it?18:35
heikoso he wants no normalise training data for each fold?18:35
blackburn1heiko: it is next on my list18:36
heikopreprocessors cannot do that?18:36
heikoah ok18:36
blackburn1heiko: they can18:36
blackburn1but we do not handle it18:36
blackburn1heiko: we should call init in cross validation I think18:36
blackburn1before training18:37
heikoI forgot that18:37
heikoits very important for some kernels18:37
blackburn1you could do that too if you have time18:37
blackburn1while I am messing with ovr multiclass :)18:37
heikoI am currently getting organised with the rest for my gsoc18:37
heikothis sigmoid fitting still has to come18:37
heikoand mkl weights per fold output18:37
blackburn1heiko: I spent some time on pre-release stuff already btw18:38
blackburn1there are ~100 doc warnings left and I checked tests18:38
heikonice, yeah I also fixed a few warning on the weekend18:39
heikothen there is still the migration stuff18:39
heikoand reference counts for sgvec/mat18:39
blackburn1we need to release on 1st of September18:39
blackburn1I like that date :D18:40
-!- zxtx [] has joined #shogun18:42
wikingfucking hell how do you plot this shit in matlab fast?18:42
wikingor python18:42
wikingi have a matrix of 2 rows and n columns18:42
blackburn1from pylab import *18:43
blackburn1X = array([yourmatrix])18:43
wikingwy this shit only shows 1 points :))018:48
blackburn1I expect your X is transposed then18:51
wikingnoup :)18:51
wikingi'm switching to python18:51
blackburn1print X.shape18:51
wikingi've tried scatter in matlab18:51
wikingshould do the same18:51
wikinginstalling scipy.. forgot to reinstall it after os upgrade :X18:52
wikingi should be able to copypaste the output of display_matrix into python :S18:55
wikingheiko: SHIIITFACK19:01
wikingwhy it no working:)19:01
heikohey wiking, calm down19:01
heikowhats happening?19:01
heikoone has to patient with plotting19:02
wikingi've plotted19:02
wikingbut i think it should be ok19:02
wikingas a lot of them are actually 0,019:02
wikingi mean for the mean =[0,0]19:02
wikingthe samples are (0,0)19:02
wikingthus when i plot i don't see it that there's equal amount there19:03
wikingas around the other grid point19:03
wikingbut as far as i can see from the matrix19:03
wikingit should be alright19:03
wikingit cannot separate it19:03
heikoit also has to be ok from the plot19:05
heikomatrix reading is left for keanu reeves :D19:05
wikingthis is weird19:07
wikinglet me pastebin it19:07
wikingthis is the matrix19:07
wikingand 2 gaussians19:07
wikingwhere mean_1=(5,5)19:07
wikingah no i'm sorry19:08
wikingmean_1= (15,15) and mean_2=(30,30)19:09
wikingand variance is always [1,0;0,1]19:09
heikohey man I cant read that19:09
heikoneed images19:10
wikingok here's the image... (sending in email)19:10
heikoI see19:13
heikobut what happened to the hypercubes?19:13
wikingit's just dim=219:13
wikingso no hypercube19:13
wikingi think the problem is the way i set the matrix elements19:14
heikobut in two dimension you should get this grid in the image i sent you19:19
heikosorry phone was ringing19:20
wikingthe problem is the filling of the matrix19:20
wikingi've printed separately the generated samples19:20
wikingand it's good19:20
wikingthis is with 2 gaussians with 2 dim19:20
wikingthe problem is that how i fill up the matrix19:21
heikoyou are adding the value to each dimension19:25
heikothats why you get this diagonal structure19:25
heikowhy not make this grid?19:26
wikingheiko: yeah i get that diagonal struct19:26
wikingbut mainly that's not the problem19:26
wikingi mean it should work with this diag struct as well19:26
wikingbut the problem is that how i fill up the matrix with these vectors19:27
wikingi think it was a miracle that it was working with SVMOcas at the first place19:27
heikoyes it should work19:27
heikoreally, oh no :)19:27
heikoisnt there a reference dataset that you could use?19:28
wikingi have data!19:29
wikingand it actually looks good19:29
wikingalthough multiclass couldn't separate it still19:29
wikingheiko: sent u an email19:31
wikingwith a graph where the means of the gaussians: 15,15 and 30,3019:31
heikowiking these should be fine19:31
wikingmulticlass cannot separate it19:31
wiking :D19:31
wikingneither multiclass liblinear19:33
wikingnor multiclass ocas19:33
heikoare you sure you did not switch dimensions or something?19:33
wikingwell now i'm going to print out the matrices separately19:34
wikingafter i devide this big matrix of samples19:34
wikinginto train and test dataset19:34
CIA-39shogun: Sergey Lisitsyn master * rc49b8cf / (8 files in 5 dirs): Added multiclass OVR evaluation -
wikingthe plot is very similar obviously19:35
wikingi'll print out the labeling as well next to the vector values19:36
wikingas i really cannot see wtf is happening19:36
heikoblackburn, matplotlib is so bad I can't believe it19:39
heikoNow 1h trying to get along with these plots19:40
blackburn1heiko: what's up?19:40
blackburn1what is so wrong with it?19:40
heikoeverything is so complicated19:40
wikingi'm right19:41
blackburn1heiko: yes it is very flexible so complex19:41
wikingthis is the train set19:41
wikingand the test set is very similar19:42
heikoblackburn, how to change the image size19:42
heikoI fail to do so19:42
heikowiking, and?19:42
wikingheiko: and everything is classified after the training as 019:42
blackburn1heiko: image size?19:43
blackburn1size of what exactly?19:43
wikingor no sorry19:43
wikingno everything is classified as 119:43
heikoblackburn, something in order to have a latex compatible size19:43
blackburn1heiko: why to do that?19:43
heikobecause scaling looks bad19:44
wikingheiko: and basically what it does according to the debug that it takes the 0 examples as training errors19:44
heikothats why I want to produce plots that fit without scaling19:44
blackburn1heiko: scaling of fonts?19:44
heikoblackburn1, yes19:44
heikowiking, so theres an error in multiclass?19:45
blackburn1heiko: okay I'd suggest to tune fontsize19:45
wikingheiko: i have no idea... neither multiclassocas nor multiclassliblinear can separate this problem19:45
heikoblackburn, no Just figure size, I set the font size in the figure to the same as in tutorial, which is 1119:45
heikowiking, try libsvm19:46
heikowithout multiclass19:46
wikingheiko: svmocas works smoothly19:46
heikoblackburn you joker, I have tried the first 10 google entries19:46
blackburn1heiko: and?19:46
heikothey dont work thats the problem19:46
blackburn1no way, why?19:46
wikingheiko: w/o any missclassification19:46
wikingheiko: w/o any missclassification19:46
wikingwrong window19:47
heikoblackburn, and I refuse to type in more that 10 lines for that problem btw ::D19:47
blackburn1heiko: so rc('figure', figsize=(5, 10)) doesn't work?19:47
heikoit looks like, but then when I save, things are discarded19:48
heikoand when I call savefig19:48
heikothe resulting pdf is empty19:48
heikoand thats when I started complaining :)19:48
blackburn1hmm it doesn't work okay19:49
wikingi've used not liblinear instead of libsvm19:49
wikingWARN] reaching max number of iterations19:49
wikingUsing -s 2 may be faster(also see liblinear FAQ)19:49
blackburn1figure(figsize=(size in inches, size in inches))19:50
blackburn1just do that19:50
blackburn1it works here19:50
wikingcomeeooon wtf is with this :)19:50
-!- blackburn1 [~blackburn@] has left #shogun []19:50
* wiking doesnt get it!19:50
heikoyou get this warning when you should optimise primal instead of dual19:50
heikotry libsvm19:51
-!- blackburn1 [~blackburn@] has joined #shogun19:51
heikoblackburn, I just figured out the problem19:51
wikingheiko: with linear kernel?19:51
heikowiking yes19:51
heikoand not multiclass19:51
blackburn1arghh I got why I have headache19:51
blackburn1I forgot to eat today19:51
blackburn1heiko: so now it works?19:53
heikoblackburn, not really, still looks horrible, this causes me a headache19:53
heikofont size is wrong and stuff19:53
heikosince this scales things19:54
shogun-buildbotbuild #336 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Sergey Lisitsyn <>19:54
heikoits so stupid, since it is an easy problem that every scientist has to approach19:54
blackburn1heiko: font size of what?19:54
heikoI dont get why there are no good solutions19:54
heikono all19:54
blackburn1wait but is size of plot correct?19:55
wikingheiko: works just fine19:56
wikingheiko: as with liblinear and svmocas19:56
wikingwhen it's binary19:56
CIA-39shogun: Sergey Lisitsyn master * ra6cdaf5 / src/shogun/evaluation/CrossValidation.cpp : Added preprocessor init calls in crossvalidation -
wikingbinary classification just works fine19:58
wikingheiko: i've tried changing dimension20:01
wikingbut that did not confuse at all the binary classifiers20:01
heikothen there is a bug somewhere20:01
wikingheiko: in case of multiclass if i put higher dimension20:03
wikingthen the error gets smaller20:03
wikingbut still nothing near to the binary case20:03
wikingi just need a confirmation20:09
wikingok i'm gonna feed this train/test pair20:11
wikingdirectly to liblinear20:11
wikingjust to see wtf20:11
heikowiking I dont know20:12
heikoit should give the same answer when you hand in the same (easy) data20:13
wikingheiko: heheh let's see i'm very very curious now20:13
wikingAccuracy = 36% (27/75)20:16
wikingif i give the train set as a test set20:16
shogun-buildbotbuild #337 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Sergey Lisitsyn <>20:19
wikingok this si just not separable20:20
wikingheiko: any ideas how we could solve this that there's a datagenerator function that creates a dataset that is separable in a way that i can do a unittest for it?20:21
heikodo the unit test on a dataset that is definately lin. separable20:21
heikoyou could use the mean dataset20:22
heikobut actually20:22
heikoI would use a fixed one20:22
heikoso that you can assert outputs20:22
wikingheiko: heheh so add a header file that has the dataset?20:22
wikingand one could include it when he wants to test a classifier?20:23
heikojust produce it in your code20:23
wikingheiko: ok how ?:)20:23
wikingas the gaussian doesn't work20:23
heikowiking, just via signs or so20:23
-!- n4nd0 [] has joined #shogun20:31
n4nd0blackburn1: hey!20:31
blackburn1n4nd0: hey there20:31
n4nd0blackburn1: how are you?20:31
wiking!@#$ :)20:31
blackburn1n4nd0: fine, watching wiking playing around with gaussians :D20:32
n4nd0wiking: how is it going with them?20:32
wikingn4nd0: shitfuck20:32
wikingwasted half a day20:32
wikingand it's not working at all20:32
heikoblackburn, I got it finally with the plots20:32
n4nd0wiking: I wish you luck!20:33
heikon4nd0, you were also interested in that right?20:33
n4nd0blackburn1: I am going to rebase the branch now20:33
n4nd0heiko: yes20:33
n4nd0blackburn1: does that mean that we'll be ready to merge afterwards??20:33
n4nd0I am really looking forward to it!20:33
blackburn1heiko: yeah I think so20:33
n4nd0blackburn1: was that last message regarding my branch?20:35
blackburn1about rebase? yeah20:36
blackburn1wow wiking you did weekly report in time :)20:36
wikingit was already in draft20:37
wikingbut the the hell with the multiclassocas unittesting20:38
n4nd0wiking: was the change CStructuredData -> CData finally merged?20:40
n4nd0I took a look to your latent branch but didn't see it20:40
wikingn4nd0: i haven't changed your code20:41
wikingn4nd0: just did the latent part20:41
wikingn4nd0: and introduced lib/CData...20:41
@sonney2kheiko, did you discuss the 'e.V.' issue?20:41
blackburn1master sonney2k in da hous20:41
@sonney2kheiko, about now we have to start to organize formalities w/ google20:42
heikosonney2k, not really, but I got a friend who studies law visiting, he hinted me some reference to check stuff out, Ill do it later this week20:42
@sonney2kblackburn1, bow before the leader ;-)20:42
heikobusy with plotting currently20:42
heikohope to sort that out soon20:42
@sonney2kheiko, ok - it is starting to be a bit pressing though20:43
heikoyeah, sorry that this took so long20:43
@sonney2kwiking, n4nd0 what is CData?20:43
blackburn1we need an offshore to keep money we get from stealing people and oil20:43
@sonney2kwiking, n4nd0 it sounds like a very generic name doing what CFeatures was invented for - no?20:44
alexlovesdatawhy not an island close to the pacific trash vortex??20:44
* sonney2k also prefers an island20:44
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]20:45
blackburn1well cyprus is an island I heard it is ok to keep stolen money20:45
n4nd0sonney2k: it is what I called CStructuredData20:45
n4nd0wiking prefers to use CData for that20:45
n4nd0to be honest, I am ok with any of them...20:45
@sonney2kn4nd0, structured data is CFeatures + CLabels right?20:46
n4nd0sonney2k: no20:46
wikingheiko: !!!!20:46
wikingheiko: i think i did it20:46
n4nd0CStructuredData is the building block of CStructuredLabels20:47
n4nd0CStructuredLabels is a DynamicObjectArray of CStructuredData20:47
heikoblawiking, what happened? :)20:47
CIA-39shogun: Heiko Strathmann master * rbe9da8e / (5 files): Merge pull request #710 from karlnapf/master -
CIA-39shogun: Heiko Strathmann master * rdaabd3f / (2 files): updated graphical examples -
CIA-39shogun: Heiko Strathmann master * r15cb24b / (2 files): renamed file -
CIA-39shogun: Heiko Strathmann master * ra99f6fe / examples/undocumented/python_modular/graphical/ : added import to make matlibplot plots look nice -
CIA-39shogun: Heiko Strathmann master * r9b66a85 / (2 files): minor updates -
blackburn1blawiking? my hamming based reception thought you are calling me and then kaboom20:48
n4nd0sonney2k: CStructuredData can be e.g. a sequence (for HM-SVM), a real number (for the multiclass example) a tree, etc20:48
heikon4nd0, blackburn, sonney2k, check out nice looking plots in tutorial!20:48
n4nd0heiko: let's see :)20:48
@sonney2kn4nd0, ok then the name is a bit misleading maybe ... in fact it is a part of the label20:52
n4nd0sonney2k: yeah, it could be20:54
n4nd0sonney2k, wiking: what name would you like to have instead?20:54
@sonney2kheiko, nice figures :D20:54
heikosonney2k, thanks :) they depict the tests very nice, take them for the new webpage20:55
blackburn1heiko: what the heck is plotted?20:56
n4nd0heiko: I am checking shogun-tutorial/fig/statistical-learning, is that the correct place?20:56
heikoyeah, but better look at the compiled pdf to the how the fit in20:57
heikoblackburn, read the tutorial :D20:57
blackburn1heiko: I sneeze when I read statistics20:57
blackburn1all the time20:58
n4nd0heiko: nice!20:58
@sonney2kbtw heiko this reminds me I haven't received any figures (apart the ones from you)20:59
n4nd0sonney2k: I won't be able to make it for the meeting this time :(20:59
wikingheiko: here?20:59
@sonney2kn4nd0, well you have been around last time - btw nico should be back so update him :D21:00
heikowiking, yes21:00
n4nd0sonney2k: yeah he sent us a mail, I am writing back to him now21:00
wikingheiko: it works :)))21:00
wikingheiko: both multiclass and binary21:00
heikowiking, well done :)!21:01
CIA-39shogun: Heiko Strathmann master * rf1dc712 / (4 files in 4 dirs): Merge pull request #711 from vigsterkr/utest -
CIA-39shogun: Viktor Gal master * r6cbc4ee / (4 files in 4 dirs): Fix gaussian_generator in DataGenertor -
heikowiking, gmock is still not workin ghere21:01
heikothats my plan for this weelk21:01
n4nd0sonney2k: btw, I am spending quite a bit of time lately reading about speech recognition. I'd like to apply our HM-SVM to some problems in this area. What do you think about it?21:02
heikosonney2, should we write an email about the figures?21:02
wikingn4nd0 sonney2k anything is good for me :) but since it's we use the same concept in latentdata and structureddata... i thoguth that it would be nice to have only one class like that21:02
@sonney2kheiko, yes21:02
@sonney2kheiko, please do21:02
heikosonney2k, which time is the meeting again?21:02
heikosonney2k, Ill do21:02
wikingheiko blackburn1 now i can use this unit test to see what happens with the update of libocas to 0.96 \o/21:02
heikosonney2k, sorry for the question just saw your mail21:02
shogun-buildbotbuild #338 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Heiko Strathmann <>21:04
wikingFAILURE I AM? :)21:04
heikothese errors are strange21:05
heikowiking no not you21:05
heikoand not me21:05
wikinghah fuck just a sec i need to send in a quickfix21:06
wikinggood that bots are not running automatically make unit-tests :D21:06
heikoah so thats whats happening?21:06
heikoI gotta go, wiking Ill  merge tomorrow21:08
-!- heiko [] has quit [Quit: Leaving.]21:08
CIA-39shogun: Sergey Lisitsyn master * r5df0dcd / src/interfaces/modular/modshogun_ignores.i : Ignored bmrm stuff in modular -
n4nd0sonney2k, blackburn1: I have just rebased a few moments ago21:16
blackburn1sonney2k: are you ok to merge?21:16
@sonney2kblackburn1, sec21:17
@sonney2kn4nd0, is you PR updated - so show I re-review it?21:17
n4nd0sonney2k: I did the changes you told me in your last revision21:19
n4nd0and now I have just rebased solving a couple of minor conflicts21:19
@sonney2kn4nd0, ok then let me have a quick look21:20
@sonney2kn4nd0, what I still don't understand is why you didn't replace the constructor21:21
@sonney2kSGMatrix< ST >* feats, int32_t num_vecs, int32_t num_feats21:21
@sonney2kwith the SGMatrixStringList<ST> one21:21
@sonney2kthen we could create typemaps for that type21:21
@sonney2kn4nd0, also I think the SGMatrix< ST > feats, int32_t feat_length, int32_t num_vecs constructor21:22
shogun-buildbotbuild #339 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Heiko Strathmann <>21:22
@sonney2kshould rather be one that takes an SGNDArray as input21:22
n4nd0sonney2k: ok with the first thing about introducing SGMatrixList21:25
n4nd0sonney2k: but the second change, it is another different constructor21:25
@sonney2kn4nd0, please explain - what it is supposed to do then21:26
n4nd0sonney2k: have you seen the explanation in the header file?21:26
wikingblackburn1: can u merge this:
CIA-39shogun: Sergey Lisitsyn master * r60278ef / (2 files): Merge pull request #712 from vigsterkr/utest -
CIA-39shogun: Viktor Gal master * r2e361a5 / (2 files): Fix unit testing of LatentSVM -
@sonney2kn4nd0, it is the one where I get *parse error*21:28
n4nd0sonney2k: hehe where?21:28
n4nd0so you construct a MatrixFeatures (a list of matrices)21:28
n4nd0from a big matrix21:28
n4nd0you give it a matrix [ ........ ]21:29
@sonney2kn4nd0, yeah but why not use SGNDArray then? you can have a tensor that is NxMxL big21:29
@sonney2kwhere L is the number of matrices21:29
@sonney2kand NxM the matrix size21:29
n4nd0and partition it in a list of [...][...][...]21:29
n4nd0sonney2k: ok, but why not to have also one constructor like this?21:30
n4nd0I mean, it can be useful to do it from a matrix21:30
n4nd0I just did it to use it from CTwoStateModel::simulate_data21:30
n4nd0I ported a function from the hmsvm toolbox that simulates two data and stores it in a matrix21:31
@sonney2kn4nd0, I understand that constructor now though I don't like it - it is probably the same as having an SGVector as input for a constructor of SGMatrix21:33
n4nd0sonney2k: yes21:33
@sonney2kanyways this one certainly is not a showstopper21:33
n4nd0I can modify the simulate data function so this constructor is not required though21:34
n4nd0in any case I don't like the idea of getting rid of this constructor ...21:35
n4nd0I can use data generated from the hmsvm toolbox in matlab from python scripts thanks to it21:36
n4nd0see examples python_modular/structure_hmsvm_mosek.py21:36
@sonney2kn4nd0, well you could easily do one reshape to get this to work then21:37
n4nd0sonney2k: ok :)21:37
n4nd0sonney2k: should I change the other constructor now and ammend?21:38
@sonney2kn4nd0, yes do that now and then let blackburn1 or so merge :D21:38
n4nd0I will stay here to feel guilty if a disaster happens :-O21:39
shogun-buildbotbuild #340 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Heiko Strathmann <>, Sergey Lisitsyn <>, Viktor Gal <>21:40
n4nd0I'll not be the only one in the blamelist at least :D21:41
wiking RUN      ] MulticlassOCASTest.train21:41
wikingmake: *** [build] Segmentation fault: 1121:41
blackburn1wiking: -1?21:41
wikingnot exactly the way i've imagined buuuut it indicates that it doesn't work!21:41
wikingblackburn1: hehe yes21:44
wikingok i'm sending in the patch21:44
wikingfor that pr21:45
-!- alexlovesdata [] has left #shogun []21:51
wikingblackburn1: done21:53
wikingthe unit tests for svmocas and multiclassocas runs as expected21:53
CIA-39shogun: Viktor Gal master * r11c1730 / src/shogun/lib/external/libocas.cpp : Add minor fixes for libocas 0.96 update -
CIA-39shogun: Viktor Gal master * r3548686 / src/shogun/lib/external/libocas.cpp : Fix labeling in msvm_ocas_solver -
CIA-39shogun: Sergey Lisitsyn master * r892a159 / (2 files): Merge pull request #697 from vigsterkr/master -
CIA-39shogun: Viktor Gal master * r0ffe237 / (2 files): Update libocas to version 0.96 -
wikingso now we'll only need to create a wrapper for svm_ocas_solver_nnw ;)21:55
blackburn1wiking: why to use it?21:56
wikinghahah good question... but if anybody who would need it...21:57
wikinganyhow i guess after this doing train unittests for other classifiers should be straightforward ;P21:57
blackburn1vojta said it was developed for some computer vision guys21:57
blackburn1I am curious what is the case to require positive weights21:58
blackburn1with only positive alphas we get better understanding of class representatives21:58
blackburn1but what do we get with positive weights?21:59
shogun-buildbotbuild #341 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Sergey Lisitsyn <>, Viktor Gal <>21:59
wikinghah i have to do there something22:00
wikingas it's my code that breaks java22:00
-!- KMcQuisten [d8338942@gateway/web/freenode/ip.] has joined #shogun22:01
blackburn1wiking: why do you inherit from dense labels?22:01
blackburn1KMcQuisten: hey there, what is the issue today? ;)22:01
wikingthat i have to redesign now22:02
KMcQuistenNot so much an issue.22:02
KMcQuistenI just want to take a chainsaw to part of your object model.  Specifically, the machine evaluation code as well as the result code.22:03
@sonney2kKMcQuisten, I would very much prefer if you do these things publicly on the mailinglist22:03
KMcQuistenThat's fine, sonney.  I can certainly do that.  I was just hoping to have a conversation first.22:04
KMcQuistenI don't have any bugs to report.  everything seems to be working as it should22:04
wikinghahah cache and pipelining kicked in in my cpu22:04
blackburn1sonney2k: we failed to discuss some issues at mailing list :)22:04
wikingcompiling is getting faster and faster :)22:04
blackburn1KMcQuisten: what do you mean?22:04
wikingblackburn1: i guess he means that all's good ;)22:05
blackburn1no about chainsaw and code22:05
blackburn1sounds scary22:05
KMcQuistenHehehe.  Not so much scary22:05
blackburn1anyway please elaborate22:05
KMcQuistenHere's my take:  When constructing cross-validation, you must pass the evaluationCriterion object to the crossValidation object, which calculates the statistics, averages them, and then passes them to the EvaluationResult object.22:07
KMcQuistenI think thats backwards.  The crossValidation object should pass the observed/predicted pairs to the evaluationResult object, and the evaluationCriterion object should be passed to the evaluationResult using a method.  This way one run of crossvalidation could be used, and then multiple evaluationCriterion could be thrown at it to look at CV results in multiple ways without having to rerun a full CV for each evaluationcriterion on22:09
KMcQuistenAs it stands, It's impossible for me to actually look at the obs/pred pairs generated by the builting cv, so I can't do more in depth analysis of my CV results if I use the builtin methods22:10
KMcQuistenI'd like the opportunity to refactor the code to make it more flexible22:11
KMcQuistenIf there's a formal proposal process for such things I'd like to follow it, of course.22:11
blackburn1if you want to improve anything just do and sent a PR22:12
blackburn1(pull request)22:12
KMcQuistenYep Yep.22:12
KMcQuistenI know that.  I have my own branch forked off.22:12
KMcQuistenI'd rather tell folks what i'm planning to do first.  Common courtesy and all, since it would involve some big changes.22:13
KMcQuistensonney, would you like me to write something up for the mailing list before I get started?22:13
blackburn1sonney2k: ^22:14
blackburn1KMcQuisten: so you want to use multiple22:14
@sonney2kblackburn1, any idea where gsomix is22:15
blackburn1sonney2k: will appear soon22:16
blackburn1sonney2k: I am surprised to see kind of machine leraning here22:16
-!- gsomix [~gsomix@] has joined #shogun22:17
gsomixgood evening22:17
KMcQuistenYes.  Without having to completely rerun CV for each criterion.  Assuming the splitting criterion doesn't change, each time CV is run on a dataset with a fixed set of parameters, it would produce the same observed/predicted label pairs, so why generate them over and over for different evaluations?22:17
blackburn1KMcQuisten: but you optimize only one criterion right?22:18
KMcQuistenThis would also allow output for evaluations to return different types, and return them as values to wrapped languages, instead of having to rely only on stdout22:18
@sonney2kKMcQuisten, I understand this desire - but what do you do if you have test data sets? then you would need to store all outputs for all evaluations in memory right?22:18
@sonney2kgsomix, hey - so what's going on? did you manage to do the example fixes?22:19
blackburn1KMcQuisten: but why no to refactor modelselectionoutput a little to store things?22:19
@sonney2kgsomix, (or did I miss anything?)22:19
blackburn1not to print but to store I mean22:19
blackburn1it already receives all you need22:19
blackburn1and it supports custom additional evaluations too22:19
KMcQuistenSee, this is why I wanted to talk about this before I did anything :) .   I'm still familiarizing myself with the modelselection code, so I am unsure.  Does it allow for the comparison of arbitrary sets of features and kernels, or must those be fixed and only the parameters varied?22:21
gsomixsonney2k, hey. what do you mean? I plan to debug my DirectorDotFeatures example and solve problem with SGVector in directors. But little later. At now I'm working on zero-copying mechanism. At next week plan to start work on new typemaps for modelselection (it seems, that it's little easy that sg in director or zero-copy).22:22
blackburn1KMcQuisten: why to do that?22:22
gsomixah, and I forgot about examples for Labels.22:23
gsomixtoo bad22:23
blackburn1I am a bit lost, comparison of what?22:23
@sonney2kgsomix, that is quite a bit of work - not sure how you want to do that within 2 weeks...22:23
blackburn1sonney2k: why to +1?22:23
shogun-buildbotbuild #342 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Viktor Gal <>22:24
@sonney2kblackburn1, if I understand correctly ocas uses labels 1... <nr_classes>22:24
@sonney2kin shogun we use 0...<nr_classes-1>22:24
blackburn1sonney2k: there is no +1 in multiclass ocas we had before22:24
@sonney2kso we should +1 shogun's labels for ocas to work22:24
gsomixsonney2k, 2 weeks? when is the release of new version?22:25
@sonney2khopefully 1st of september22:25
KMcQuistenComparison of the CV accuracies for SVMs applied to my data under various choices of feature expression, kernel selection, and kernel parameter.22:25
KMcQuistenblackburn1: ^22:25
@sonney2kbut we need at least 1 week to just do testing / bugfixing / documentation an no more feature additions22:25
blackburn1KMcQuisten: okay lets specify things here22:26
gsomixzero-copying mechanism - 4-5 hours22:26
@sonney2kI would actually say 2 weeks ...22:26
gsomixsolve problem with SGVector  - 1-2 day22:26
@sonney2kgsomix, well do what you can...22:26
KMcQuistenblackburn1: Yes, let's.22:26
@sonney2kgsomix, what is the new SGVector problem?22:26
@sonney2kgsomix, I thought you resolved it?22:26
gsomixsonney2k, nope.22:27
blackburn1KMcQuisten: cross validation splits to train/test subsets22:27
blackburn1you need these indices, right?22:27
blackburn1when you construct parameter tree you may specify everything you want to vary22:28
KMcQuistenblackburn1:  No, I don't .  I don't want to change anything about how the CV code does splitting, training, or testing of the untrained portion of the data.  All that's fine, and I don't want to reinvent the wheel.22:28
gsomixsonney2k, LibLinear's train in my example says that 'not a numpy or... bla-bla'. I need to debug it.22:28
wikingsonney2k blackburn1 it's now good at is... ;)22:28
gsomixnot many time, I think22:29
blackburn1KMcQuisten: okay22:29
blackburn1KMcQuisten: so you need a few more evaluation to see22:29
blackburn1are you ok with optimizing by only one evaluation parameter?22:29
-!- sr___ [u5548@gateway/web/] has joined #shogun22:29
n4nd0wrong tab...22:30
@sonney2kblackburn1, wiking I think the best thing to do would be to compare mc ocas output (of stand alone ocas) against shogun once22:32
@sonney2kand then feel better22:32
wikingsonney2k: :DDDD22:32
wikingsonney2k: i've compared libocas v0.95 and how that was ported in shogun22:32
KMcQuistenYes.  I don't want to pass multiple evaluation criterion to the crossvalidation object.  My point is that structurally I don't think the evaluationcriterion belongs in the CV code.  The CV code should just spit out obs/pred pairs to the result object, and the evaluationCriterions can be passed to the result object.22:33
wikingsonney2k: and there the -1 was removed basically22:33
@sonney2kgsomix, well then do 0-copy now ...22:33
KMcQuistenThe question is, as sonney said, how that will work for very very large datasets22:33
@sonney2kKMcQuisten, make it optional...22:33
@sonney2kthen it is ok22:34
blackburn1KMcQuisten: but why not to use modelselectionoutput?22:34
@sonney2kwiking, what where the class labels you used?22:34
blackburn1I don't get that point yet22:34
KMcQuistenI have to look more into the modelselection code before I can give you a good answer to that.22:34
-!- n4nd0 [] has quit [Ping timeout: 244 seconds]22:34
blackburn1KMcQuisten: you may add a few more evaluation instances to modelselectionoutput22:34
blackburn1and it will be printed22:35
KMcQuistenIf that's a way I can get what I want without messing with code, then it's fine,22:35
wikingsonney2k: yeps22:35
blackburn1if you want to make it not printed but stored - we can do that22:35
KMcQuistenAha.  That's the other problem.  I'd rather just get that value and keep in internally rather than having to scrape from stdout22:35
@sonney2kblackburn1, would that enable us to also store trained models?22:36
KMcQuistenBut I can conceive of eval criteria that aren't simple floats, but rather arrays of floats, i.e. a list of binned accuracies across the entire prediction range of a regression problem22:36
blackburn1sonney2k: it prints trained model right now22:37
-!- n4nd0 [] has joined #shogun22:37
@sonney2kwiking, I would suggest to email vojtech (after reaing ocas doc)22:37
blackburn1if you pass model selection output to the instance22:37
@sonney2kblackburn1, *store*22:37
blackburn1sonney2k: yes, if we change print to store - it would enable to do that22:37
blackburn1but I planned to do that optionally of course22:37
@sonney2kwiking, I am pretty sure he uses labels 1...nr_class22:37
@sonney2kmakes sense22:37
blackburn1so now you may put22:39
blackburn1ModelSelectionOutput instance22:39
blackburn1to ModelSelection instance22:39
blackburn1and it will print stuff you need22:39
blackburn1you may add anything to ModelSelectionOutput22:39
blackburn1like more evaluations22:39
blackburn1if you add some new complex evaluation I'd suggest to add its handling to the ModelSelectionOutput22:40
blackburn1I will change it to store things today/tomorrow22:40
KMcQuistenThat sounds fine to me.  I will take a look and see what I can do.22:40
shogun-buildbotbuild #343 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Sergey Lisitsyn <>, Viktor Gal <>22:41
blackburn1KMcQuisten: I just have to think how can we add this kind of storage here22:42
KMcQuistenWell, hold up on that.  It might not need to be stored if it can just be calculated and returned22:43
@sonney2kKMcQuisten, well it can take quite a while...22:44
* sonney2k Zzzz...22:44
blackburn1KMcQuisten: I hardly have an API in mind for that22:45
KMcQuistenblackburn1:  Ok, cool.  I'll familiarize myself a bit more with the modelselection code and hopefully I will be able to see how it can help me.22:46
blackburn1KMcQuisten: remind me please - are you using python?22:47
blackburn1the problem is that storage should be quite generic22:47
KMcQuistenwith the SO machines and all being put together it has to work for them too22:48
KMcQuistennot everyon'es outputs are nice floats. :)22:48
blackburn1well evaluation is always float - a score22:49
KMcQuistenYes, I know, but they predictions that the machines make are not always nice floats22:49
n4nd0sonney2k, blackburn1: I updated the branch with the new constructor22:49
KMcQuistenand I can think of ways that the score is actualyl an array of scores, as opposedto a single float, so even that assumption isn't quite right wither22:50
blackburn1KMcQuisten: would be nice to have map: ParameterCombination -> some bunch of results22:51
blackburn1but it is very inconvenient for user22:51
KMcQuistenPerhaps.  I'm not entirely sure we're talking about the same thing here.22:51
KMcQuistenI want to dig into the modelselection code before we discuss this further.22:52
blackburn1KMcQuisten: okay22:52
blackburn1you are not the only user that wants that stuff btw22:52
blackburn1second one already :)22:52
n4nd0blackburn1: should we merge or do you prefer to wait for Soeren until tom?22:54
blackburn1n4nd0: did you have to include .cpp?22:55
blackburn1that sounds to be a blocker for you lets just merge22:55
n4nd0yes, it looks like this now22:55
n4nd0from SGMatrixList.h it is included SGMatrix.h22:56
blackburn1however I do not like that that much22:56
blackburn1is .cpp included?22:56
blackburn1n4nd0: do you mind to make it static method of sgmatrixlist?22:56
n4nd0in SGMatrix.h it is SGMatrixList forward declared22:57
n4nd0and from SGMatrix.cpp, SGMatrixList.cpp is included22:58
n4nd0blackburn1: yeah, but I can tell you why it is done like that22:58
n4nd0blackburn1: ok22:58
blackburn1n4nd0: including .cpp could be bad22:58
n4nd0blackburn1: it is just because of the style in which templated classes are defined in SHOGUN22:59
n4nd0blackburn1: because it is in the .cpp where the types are specified22:59
blackburn1i see22:59
blackburn1n4nd0: anyway .cpp should not be included - it can become a mystery problem later23:00
n4nd0I agree23:00
n4nd0I am changing it to static right now :)23:01
blackburn1KMcQuisten: there are now 3 guys wanting to check xval results in a fancy way23:02
KMcQuistenHehe.  Fancy23:02
KMcQuistenI feel like i need a parasol now LOL23:03
* blackburn1 never knew what is parasol23:05
KMcQuistenfancy umbrella for protecting fair-skinned ladies form the sun23:06
blackburn1yeah I googled23:06
KMcQuistenAh yes23:06
KMcQuistenSHould have known ;)23:06
blackburn1the only word for that I knew is umbrella23:06
blackburn1but there are quite a few like rainshade and others23:06
KMcQuistenHahaha!  I'd not hear that one23:07
KMcQuistenOk, i'm going to study modelselection, and then I'll know better what can and cannot be easily fiddled with :)23:07
KMcQuistenthanks for your help again23:07
-!- KMcQuisten [d8338942@gateway/web/freenode/ip.] has quit [Quit: Page closed]23:08
n4nd0blackburn1: you know, parasol is a funny word23:10
blackburn1yeah sounds funny at least23:10
n4nd0we use it in Spanish too23:11
n4nd0it means like this23:11
n4nd0para = stop23:11
n4nd0sol = sun23:11
blackburn1I remember song or so23:11
blackburn1dia de sol23:11
n4nd0which one?23:11
blackburn1sunny day?23:11
blackburn1huh I know spanish hurray23:11
blackburn1I am watching participants list in your PR23:12
blackburn1two are normal guys23:13
blackburn1and only I am a cat23:13
blackburn1that cat is so damn awesome you shall not resist23:13
CIA-39shogun: Sergey Lisitsyn master * r31528b4 / (2 files in 2 dirs): A few more warnings fixes -
* wiking cannot keep up with rebasing :)23:19
wikingblackburn1: here? have u managed to get unit testing working? :)23:30
-!- n4nd0 [] has quit [Ping timeout: 248 seconds]23:30
blackburn1wiking: no I didn't try :)23:30
wikingi've just made a simple pre-commit hook for git23:30
wikingand it works great with the unit testing23:30
wikingif it fails, it does not let you commit :)23:30
-!- n4nd0 [] has joined #shogun23:31
wikingblackburn1: should i add that into README.developer?23:33
blackburn1I am not a big fan of that solution to be honest23:33
wikingblackburn1: ok i'll add it to the README.developer to kindly ask developers to add that to their pre-commit hook23:34
wikingbut of course we cannot force it23:34
blackburn1yeah sure23:34
wikingso it'll be just a suggestion in a way23:34
shogun-buildbotbuild #344 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Sergey Lisitsyn <>23:39
blackburn1wiking: can you change base of latentlabels?23:41
wikingblackburn1: yep i'll change now23:42
wikingjust a sec23:42
n4nd0blackburn1, wiking: hey guys23:47
n4nd0something I don't really understand entered in a commit23:47
-!- blackburn1 is now known as blackburn23:47
blackburnuh finally I got that back23:47
n4nd0it is related to the data submodule23:48
n4nd0take a look here please
n4nd0I ignore whether this is or not relevant23:48
n4nd0did you get my messages about the problem ^?23:48
n4nd0my connection sucks...23:49
blackburnn4nd0: no please remove that23:50
blackburnyour data can be outdated and you trying to restore it to your state23:50
blackburnif it appears in your changes just update data folder , i.e. enter to shogun/data and git pull origin master23:51
wikingor git submodule update23:51
CIA-39shogun: Viktor Gal master * rdabe375 / src/README.developer : Add unit testing description to README.developer -
wikingblackburn: deep enough description?23:54
CIA-39shogun: Sergey Lisitsyn master * r2250325 / src/README.developer : Merge pull request #713 from vigsterkr/latent -
blackburnmay be I should read it before23:54
blackburnyeah deep enough23:54
wikingok let's see that label thingy23:55
wikingas a quickfix23:55
n4nd0blackburn: I do not manage to take out that change from the past commit though23:58
blackburnn4nd0: which change?23:59
--- Log closed Tue Aug 14 00:00:14 2012