Open in new window / Try shogun cloud
--- Log opened Thu May 24 00:00:41 2012
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]02:03
-!- wiking [] has joined #shogun02:08
-!- wiking [] has quit [Changing host]02:08
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun02:08
-!- av3ngr [av3ngr@nat/redhat/x-dnxruybggcigxdmo] has joined #shogun02:16
-!- av3ngr [av3ngr@nat/redhat/x-dnxruybggcigxdmo] has quit [Client Quit]02:16
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]03:04
-!- vikram360 [~vikram360@] has quit [Ping timeout: 256 seconds]07:05
-!- vikram360 [~vikram360@] has joined #shogun07:05
-!- pluskid [] has joined #shogun07:37
-!- n4nd0 [] has joined #shogun08:11
CIA-113shogun: Soeren Sonnenburg master * r95c124d / (24 files): fix most of octave_modular examples for new labels -
-!- n4nd0 [] has quit [Quit: leaving]08:33
-!- n4nd0 [] has joined #shogun08:34
-!- uricamic [~uricamic@2001:718:2:1634:ac88:5a8e:28de:2e53] has joined #shogun08:53
-!- wiking [] has joined #shogun08:54
-!- wiking [] has quit [Changing host]08:54
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun08:54
-!- n4nd0 [] has quit [Ping timeout: 265 seconds]09:03
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]09:16
-!- wiking [] has joined #shogun09:23
-!- wiking [] has quit [Changing host]09:23
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun09:23
-!- blackburn [~blackburn@] has joined #shogun10:54
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]11:39
-!- vikram360 [~vikram360@] has quit [Ping timeout: 245 seconds]11:50
-!- n4nd0 [] has joined #shogun11:57
-!- wiking [] has joined #shogun11:58
-!- wiking [] has quit [Changing host]11:58
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun11:58
n4nd0pluskid: hey! did you find the version of the paper you talked about yesterday?11:58
n4nd0pluskid: I am in the university now so I can try to look for it otherwise11:58
pluskidn4nd0: no, please try11:59
pluskidI don't have easy access to the school network after graduation... :(11:59
n4nd0pluskid: all right, got it :)12:00
n4nd0pluskid: do you want me to mail it to you, dropbox, or?12:00
pluskidn4nd0: mail me, thanks!12:02
n4nd0all right, it should be in your mailbox soon12:02
n4nd0pluskid: let me know if it is the correct one12:02
pluskidn4nd0: awesome! It's the one I'm looking for12:03
n4nd0pluskid: nice, and I forgot to say, you are welcome ;)12:03
-!- imeanwow [46e7e626@gateway/web/freenode/ip.] has joined #shogun12:03
imeanwowI accidently typed #shogun when I tried to name myself Shogun12:04
imeanwowand that mistake brought me here.12:04
-!- pluskid is now known as Shogun12:04
blackburnn4nd0: Shogun: what is the paper?12:09
n4nd0blackburn: sensitive ECOC12:09
blackburnsome data-dependant ecoc scheme?12:09
blackburnhmm can you mail to me too? :)12:10
Shogunblackburn: don't know, haven't read yet :p12:11
-!- Shogun is now known as pluskid12:11
-!- wiking [~wiking@huwico/staff/wiking] has quit [Ping timeout: 245 seconds]12:18
blackburnn4nd0: when do you think you will be able to merge structured labels thing?12:20
n4nd0blackburn: now it is in compilable state but there are things that probably need to be completed12:21
n4nd0blackburn: what state do you want it to be in order to be merged?12:21
blackburnhmm no idea - what is uncompleted in SO labels?12:22
n4nd0I just have a basic interface now12:23
n4nd0once I start using it for specific examples, functionality needs will pop up12:23
n4nd0blackburn: for me, the sooner we merge it, the better. It feels safer12:25
blackburnyes - but can't you provide all the needed functionality in that interface?12:25
n4nd0blackburn: it could be a good idea to wait since I may not have a perfect idea of all we need there a priori12:26
blackburnthat's why you have mentor :)12:29
blackburnhowever yes hard to predict everything12:29
n4nd0what has it happened in github? the comments in the interface files seem to be wrong indented12:30
n4nd0I have opened some of them at random and they look so12:35
pluskidlooks fine here12:42
n4nd0pluskid: mmm really? it appears in the browser, it should be the same for all I guess12:43
pluskidall comments are of the same indent in my browser12:44
imeanwowWhat is right or what is wrong? When you realize that you possess a highly powerful set of skills12:44
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun12:54
-!- blackburn [~blackburn@] has quit [Ping timeout: 248 seconds]13:16
sonne|workn4nd0: my take on this is that I would rather merge early and often as long as your code compiles and doesn't break anything14:00
sonne|workand as you have seen - nothing in shogun is set in stone...14:00
sonne|workbut some things are harder to correct than others14:00
-!- vikram360 [~vikram360@] has joined #shogun14:01
n4nd0sonne|work: it compiles well and I have nothed nothing breaking here :)14:01
sonne|workthen pull request and I merge14:01
n4nd0sonne|work: ok, I will rebase first to introduce changes from yesterday midday14:02
n4nd0sonne|work: I think it is a good idea to merge since in that way uricamic, wiking and I know how each other is working14:05
n4nd0what will be good since our projects are somewhat related14:05
wikinganybody seen alex around today?14:07
sonne|workwiking: no ...14:08
wikingwe suppose to have a meeting now14:08
wikingor at least today14:08
sonne|workwiking: hmmmh?14:08
sonne|workemail him :)14:08
wikingsonne|work: yeah we've agreed it via email ;P14:09
sonne|workno email him now again reminding him14:09
n4nd0by the way, does any of you see this warning when compiling?14:12
n4nd0./../shogun/lib/DynamicArray.h:296: Warning 516: Overloaded method shogun::CDynamicArray< char >::element(int32_t,int32_t,int32_t) ignored,14:12
n4nd0../../shogun/lib/DynamicArray.h:284: Warning 516: using shogun::CDynamicArray< char >::element(int32_t,int32_t,int32_t) const instead.14:12
n4nd0it appears quite a lot of times here14:12
sonne|workn4nd0: with swig you mean?14:35
sonne|workyeah no idea why we need the const one...14:35
n4nd0sonne|work: I don't know if it is a problem with SWIG or something different14:35
n4nd0and I am not sure if it is something that happens to me locally14:36
sonne|workn4nd0: well if you compile with ./configure --interface=14:36
sonne|workthis warning shouldn't be there14:36
n4nd0I am compiling with ./configure --interface=python_modular --disable-optimization --disable-doxygen14:36
n4nd0mmmm maybe --integerface=pyton_modular was not there the last time14:37
n4nd0sonne|work: ok, I've just compiled with --interfaces= and no warnings :) thank you!14:43
sonne|workn4nd0: well you usually (before doing a pull request) should compile it with one static and one modular interface14:44
sonne|workotherwise lots of errors slip through14:44
sonne|workn4nd0: can we have a better name than VanillaStructuredOutputMachine?14:48
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]14:48
sonne|workwhat is a vanilla machine?14:48
sonne|workice cream :)14:49
n4nd0we are using there the optimization algorithm that we have designated as "vanilla"14:49
n4nd0I guess because it is the most straighforward or naive one, not because of its taste :P14:49
n4nd0mmm let me see14:50
sonne|workn4nd0: I first thought you meant the SOMachine14:51
sonne|workbut then I realized that there is SOMachine -> linearsomachien -> vanillasomachine14:51
n4nd0sonne|work: check here
n4nd0there is a class diagram in the last post that will make you see the design faster14:52
n4nd0sonne|work: now I realize that it probably makes sense to have the weight vector in LinearSOMachine instead of in VanillaSOMachine14:54
sonne|workyeah ok but we need some better term than vanilla14:57
sonne|workn4nd0: anyway I am merging right now but please fix both the 'vanilla' and w -> linearsomachine14:58
n4nd0sonne|work: any suggestion instead of vanilla?14:58
CIA-113shogun: Soeren Sonnenburg master * r8b0279c / (28 files in 5 dirs): Merge pull request #551 from iglesias/so (+18 more commits...) -
sonne|workn4nd0: what does 'vanilla' do ?14:59
sonne|workn4nd0: nice blog15:00
sonne|workI have to add it to planet tonight15:00
n4nd0sonne|work: it is the algorithm Nico wrote in the pdf he sent around15:01
sonne|workn4nd0: which one15:01
sonne|workthere were many :)15:02
n4nd0in the second page of the document15:03
n4nd0Algorithm 1 (the only algorithm there)15:03
sonne|workalright the primal solver15:03
sonne|workunconstrained optimzation problem ?15:04
n4nd0there are constraints15:05
n4nd0although they are not written in the algorithm15:05
n4nd0they appear in equation (1), using the output of init_op15:05
n4nd0but I think that, even if the formulation that appears there is for the primal, this same algorithm can be used for the dual15:06
sonne|workn4nd0: I think you should change then name completely then15:06
sonne|workI mean drop the SOMachine part15:06
n4nd0in the paper they do it with the dual IIRC15:06
sonne|workand invent something like libsvm :)15:07
n4nd0I would like to have the SO part though15:08
sonne|workn4nd0: will you be using mosek for it?15:08
n4nd0this is something that works for SO since we are using this Delta loss function and the combined featuers15:08
n4nd0sonne|work: hopefully15:08
n4nd0I need some help with the formulation of the optimization problem, I am supposed to talk to Nico about it now around here ...15:09
sonne|workn4nd0: yeah but suffixing with the machine name unecessary makes the name longer15:09
-!- karlnapf [] has joined #shogun15:09
sonne|workI mean we could have LiblinearLinearClassifierMachine otherwise15:10
n4nd0sonne|work: haha ok you convinced me15:10
-!- karlnapf1 [] has joined #shogun15:12
-!- karlnapf [] has quit [Ping timeout: 260 seconds]15:14
-!- n4nd0 [] has quit [Ping timeout: 244 seconds]15:31
-!- n4nd0 [] has joined #shogun15:41
-!- nicococo [] has joined #shogun15:58
nicococohi n4nd016:00
n4nd0hey nicococo, how is it going?16:01
nicococomedium to well :)16:01
nicococohow is the toolbox progressing?16:01
n4nd0so so+16:02
n4nd0I am kind of stuck with the optimization part16:02
nicococohehe.. what can i do to help speed it up?16:02
n4nd0so I don't know how to translate a couple of terms in our optimization problem to mosek16:03
n4nd0the first one I already asked you by mail, but I think I didn't get that much the answer16:03
n4nd0so we have this16:03
n4nd0min_w 1/2 w'?C?w + sum_i l(...)16:04
n4nd0where l is loss function evaluated in this expression max_y16:04
n4nd0you know which one I mean right?16:04
nicococookay and the optimizer has the following form:16:04
nicococomin_x 0.5 x'Hx + f'x .. right16:05
n4nd0with a constant term too16:05
n4nd0but yes, that's it16:05
nicococothe trick here is: 1. we don't use the unconstraint version16:05
nicococowhich means we have:  0.5w'Cw + sum_i slack_i16:05
nicococoall you have to do know is to add one constraint per iteration per example to the Ax <= b16:06
nicococonow, we have to setup the optimization variable x which is a block variable16:07
nicococox = [w, slacks]16:07
nicococothen f = ones(N,1)16:08
sonne|worknicococo: hey there - do you have an idea for a short but more reasonable name than VanillaSOMachine?16:08
nicococoi like vanilla... don't you16:08
sonne|workno the name is too long16:08
sonne|workwe have CPLEXSvm :)16:09
n4nd0or VanillaSOSVM16:09
nicococofine with me16:09
sonne|worknot vanilla though16:09
sonne|workit doesn't say anything16:09
n4nd0it says that it is basic16:10
sonne|worknicococo: don't you have a name for that problem?16:10
sonne|worka lot better to digest16:10
sonne|workdoes it need SVM?16:10
sonne|workin the name?16:10
sonne|workotherwise we take this16:11
nicococoBasicPrimal is not a good name without SOSVM, right?16:11
sonne|workerr if yes16:11
sonne|worknicococo: you can also think up names like svmlight, torch, libsvm or whatever16:12
n4nd0nicococo: I don't understand why f turns out to be a vector of ones16:12
nicocococool, what abaout VanillaSOSVM ;)16:12
sonne|worknicococo: when we meet again next time ....DANGER16:13
sonne|worknicococo: will this use mosek?16:13
nicococoi'm not really scared :)   and yes, mosek it is16:13
sonne|workPrimalMosekSOSVM then16:13
nicococookay back to ones vector?16:15
n4nd0nicococo: so how it turns out that f = ones(N,1)?16:15
n4nd0for me that looks like16:15
nicococoyoure right it is a zero ones vector16:16
n4nd0min w'Cw + w16:16
sonne|worknicococo: did you see Alexander today? wiking was looking for him said that they wanted to meet in chat..16:16
nicococono, we have x = [w;slacks]16:16
nicococosonney, i can give him a hint16:16
n4nd0nicococo: zero ones vector?16:16
nicococookay, so x = [w;slacks] and min_x 0.5 x'Hx + f'x16:17
n4nd0slacks is empty at the beginning or?16:17
nicococoif f =[zeros(len(w),1); ones(N,1)] then we have sum_i slacks_i16:17
n4nd0I understand what you mean :)16:18
nicocococool ;)16:18
n4nd0so this vector x has always the same dimensions16:19
nicococothen H = [C,zeros(N);zeros(N),zeros(N)] right?16:19
nicococox has always dim = len(w) + num_examples16:19
nicococowhen you generate a new constraint then you add a line to A and b16:20
n4nd0mmm, ok wait a moment there16:22
n4nd0we have two types of constraints16:22
nicococowhere A = [-psi_i +psi_i(y_argmax) -delta_i(y_argmax), ...]16:22
nicococotwo types?16:22
n4nd0Aw <= a16:22
n4nd0Bw = b16:22
nicococono, we have Ax <= b16:23
nicococoand thats enough we don't need the equality constraint16:23
nicococoand we have a lower bound for slacks16:23
n4nd0so is it like Ax <= b includes also the equality constraint?16:24
n4nd0or it shouldn't be there in the pdf16:24
nicococowhich equality constraint you mean?16:24
n4nd0in the pdf you wrote16:24
nicococooh..  in the pdf was a very general version .. all constraints there are for applications only..16:25
nicococohere: 1. we drop all the constaints i wrote in the pdf16:25
nicococoand 2. we make the unconstraint version constraint again by introducing slacks16:25
nicococohence, we have inequality constraints again16:26
nicococois that clear?16:26
n4nd0I don't understand why to introduce the slacks as in x = [w; slacks]16:27
n4nd0makes that inequality contraints like Ax <= b appear16:27
nicococowe introduce the slacks because otherwise we cannot solve it with the QP solver16:27
n4nd0yes, I understand that16:27
n4nd0but I don't understand why if we have something without contraints16:28
n4nd0why then Ax <= b appears16:28
n4nd0I have clear the min x'Hx + f'x part16:28
n4nd0how we have get there from min w'Cw + sum_i ...16:29
nicococosometimes in some application you may want to introduce some restrictions to your variables in w.. like the w(20) <= 10 or whatever16:29
nicococowas that your question?16:29
n4nd0but those are lb <= w <= ub constraints16:30
nicococo..or how do we get from the unconstraint to the constraint version..16:30
nicococo(ps. what about w(20)-w(19)<=10)16:31
n4nd0so what about A and b in Ax <= b16:32
n4nd0you said something about adding rows and cols16:32
nicococow'psi_i - w'psi_yhat >= delta_yi_yhat - slack_i16:32
nicococowell, this is one line in Ax <= b, right?16:38
nicococookay, so is everything clear?16:38
n4nd0nicococo: yeah ... kind of :)16:39
nicococoif you have more questions, please ask :)16:39
n4nd0nicococo: I will try to formulate the problem for mosek16:40
n4nd0I think I should be able to do it with this new info you have given to me16:40
n4nd0although I don't know where some of the terms given by init_op, like A, a, B and b are used16:41
nicococothe new constraint would be:  -delta >= w'psi_y_hat - w'psi_i - slack_i16:41
n4nd0and this is a bit dangerous now since there are two different things called A (and the same for b)16:41
nicococoohh.. right. Ax <= a16:42
nicococoin many applications A,a,B,b are empty in the beginning and C=eye(len(w))16:43
n4nd0nicococo: ok, I think I can forget about them for the moment16:44
nicococoand in PrimalMosekSOSVM we don't handle equality constraints Bx = b16:44
n4nd0I don't have clear one point16:44
n4nd0about Ax <= b16:44
n4nd0what are the training examples that have to be included there?16:45
nicococopsi(x_i,y_i) ?16:46
n4nd0let's put it in another way16:46
n4nd0Ax <= b represents constraints that we use like w'psi_y_Hat - w'psi_i <= -delta + slack_i16:47
n4nd0each i is a training example16:47
n4nd0what are the training examples that we have to include?16:47
n4nd0mmm I don't think I am making it very clear :S16:47
n4nd0A and b in Ax <=b, do they always have the same size?16:48
nicococoi think i know what you mean16:48
nicococoA is growing one line per iteration and b is growing too16:48
n4nd0in the first iteration, the first time we solve the QP, how are they?16:49
nicococoin the first iteration, you first add num_examples constraints therefor A has num_examples lines16:49
nicococothen you solve it16:50
n4nd0how does A and b look like at the beginning?16:50
n4nd0if we have no w initially, how do we get y_hat?16:50
nicococow =zeros16:51
nicococoor rand or something else we get from the pre-solver16:52
n4nd0ok, and we just predict with that16:52
nicococoyepp, the first iteration will give us really bad estimates certainly16:52
n4nd0and then, what are the constraints that are added, one per iteration?16:53
nicococoone per example per iteration16:53
n4nd0which example is added? the one with max loss?16:54
nicococofor each example add constraint if it produces more slack than all the constraints added for the current example16:55
n4nd0so in the pdf, Vanilla primal-SO-SVM16:56
n4nd0this is what is done in lines 7-916:57
nicococoyes, looks cryptic but is actually easy16:58
-!- uricamic [~uricamic@2001:718:2:1634:ac88:5a8e:28de:2e53] has quit [Quit: Leaving.]16:58
-!- pluskid [] has quit [Quit: Leaving]16:58
nicococo(there might be some signs that are wrong ;) )16:59
n4nd0nicococo: ok, let me try to give it a go then17:00
nicococoand a w'psi_i term is missing :)17:00
n4nd0nicococo: and I will get back to you so we can correct mistakes ;)17:00
CIA-113shogun: Soeren Sonnenburg master * rd4a6596 / (6 files): fix lua examples -
nicococolooking forward to17:00
n4nd0nicococo: have you used mosek before?17:00
nicococoonly matlab interface17:01
n4nd0nicococo: ok, I am fairly new to it though17:01
n4nd0nicococo: did you use matlab interface to solve this same problem?17:01
nicococoyes, indeed ;)17:01
n4nd0nicococo: it would be awesome if you can share that code17:02
n4nd0is there any problem with that?17:02
nicococommhh.. i guess i can sent it to you via email17:03
@sonney2knice way to compare results...17:03
n4nd0nicococo: cool17:03
n4nd0nicococo: ok I have enough info to start working on this with more possibilities now :)17:04
n4nd0thank you very much!17:04
nicococoi will send the code to you and we can chat on monday again? (or earlier if theere are problems)17:05
n4nd0nicococo: I have my exam on Wednesday so let me better mail to you and see how I am doing with it ok?17:06
nicococoalright ;)17:06
nicococobye and good luck17:06
-!- nicococo [] has left #shogun []17:06
@sonney2kn4nd0, last one then?17:08
n4nd0sonney2k: yes17:08
n4nd0sonney2k: last one this year17:09
@sonney2kwhich subject?17:09
@sonney2kvery good17:09
n4nd0software engineering17:09
n4nd0it is very boring ... I am not studying that much :(17:09
n4nd0way too much theory17:09
CIA-113shogun: Soeren Sonnenburg master * r29b7613 / examples/undocumented/java_modular/ : fix minimal libsvm java_modular example -
CIA-113shogun: Soeren Sonnenburg master * r1ff3824 / examples/undocumented/csharp_modular/classifier_libsvm_minimal_modular.cs : fix minimal libsvm csharp example -
CIA-113shogun: Soeren Sonnenburg master * r2be463b / (4 files): fix label get_name() -
-!- n4nd0 [] has quit [Quit: leaving]17:13
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has joined #shogun17:42
-!- gsomix [~gsomix@] has joined #shogun18:01
gsomixhi all18:01
puffin444how is it going?18:07
puffin444gsomix: how's your first week going?18:09
gsomixpuffin444, hey. exams :)18:10
puffin444gsomix: Well that's not fun :) Are these the final exams? Does your university run through the summer?18:10
gsomixbut I'm working with director classes... slowly18:11
gsomixpuffin444, not final, but most hard. I have exams until the end of June.18:12
puffin444gsomix: I see.18:12
-!- karlnapf [] has joined #shogun18:28
-!- karlnapf1 [] has quit [Ping timeout: 260 seconds]18:29
-!- karlnapf [] has quit [Ping timeout: 260 seconds]18:33
@sonney2kgsomix, where was your blog again?18:58
* sonney2k is lost18:58
@sonney2kdid karlnapf have one?18:58
@sonney2kpuffin444, do you have a blog19:03
-!- blackburn [~blackburn@] has joined #shogun19:06
puffin444sonney2k, no, I do not.19:12
blackburnhey there19:13
@sonney2kpuffin444, ok19:16
@sonney2kblackburn, do you remember whether karlnapf has a blog19:16
blackburnsonney2k: most of python modular fail19:20
blackburnis it the same on your machine19:20
blackburnTraceback (most recent call last):19:23
blackburn  File "", line 32, in <module>19:23
blackburn    classifier_averaged_perceptron_modular(*parameter_list[0])19:23
blackburn  File "", line 26, in classifier_averaged_perceptron_modular19:23
blackburn    out_labels = perceptron.apply().get_labels()19:23
blackburnTypeError: LinearMachine_apply expected 1 arguments, got 019:23
blackburnbe back in 30 mins19:23
@sonney2kblackburn, ahh yes that is him :)19:26
gsomixsonney2k, I don't have blog. :(19:42
gsomixsonney2k, I will give the results for director classes tomorrow.19:45
gsomixbut I cannot be at meeting. labs, labs, labs. last labs =____=19:46
gsomixI'm very sorry for that.19:46
@sonney2khmmhh too bad :(19:50
@sonney2kblackburn, yeah same here19:51
@sonney2kblackburn, I finally understand the error though - but no idea how to fix it currently19:51
gsomixsonney2k, crazy russians universities :(19:52
blackburnsonney2k: well one thing I think is wrong is rename for apply()20:04
blackburnall machines have apply() and apply(CFeatures* data)20:04
blackburnhowever i tried that and it didn't help20:04
blackburnbut I still believe it is something wrong20:04
CIA-113shogun: Soeren Sonnenburg master * r168604a / src/interfaces/modular/Machine.i : fix %renames to not remove %extend'ed applies -
@sonney2kblackburn, please check if you multiclass liblinear works now20:13
@sonney2kI mean x.apply_multiclass() gives same result as x.apply()20:13
blackburnsonney2k: does apply work now?20:13
@sonney2kvery good20:17
* sonney2k is off for a swim20:17
blackburnsonney2k: yes all is ok now20:18
blackburnlets check.sh20:18
blackburnI am impressed20:18
blackburnbattleshogun operational20:19
blackburnok I think we need apply_structured20:21
blackburnfits perfectly20:21
-!- imeanwow [46e7e626@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]20:26
-!- vikram360 [~vikram360@] has quit [Read error: Connection reset by peer]20:27
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]20:27
-!- vikram360 [~vikram360@] has joined #shogun20:27
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has joined #shogun20:36
@sonney2kall good, fresh and relaxed I am back :)21:19
@sonney2kblackburn, we have to make a decision about static interfaces21:19
puffin444by the way, I got to present at Google Chicago last night.21:19
blackburnsonney2k: which decision?21:19
@sonney2kcurrently the command there is set_labels', 'TRAIN', lab)21:20
-!- n4nd0 [] has joined #shogun21:20
blackburnpuffin444: and how it was?21:20
puffin444blackburn: It was fun. The talks only lasted five minutes so I didn't get to say much21:20
@sonney2know we need to specify the label type21:20
@sonney2kblackburn, question is how do we do it21:20
@sonney2kI mean we have to break the interface I am afraid21:21
blackburnsonney2k: automagically!21:21
@sonney2kblackburn, can we do that?21:21
blackburnsonney2k: why not - just add some function that dispatches: if (-1,+1) then create binary21:21
blackburnelse regression21:22
@sonney2kblackburn, I didn't see the beginning21:22
blackburndo we support multiclass at all?21:22
@sonney2kgot disconnected ...21:22
blackburnsonney2k: we can add function there21:22
blackburnthat dispatches21:22
blackburnif {-1,1} then binary21:22
blackburnif {0,1,..} then multiclass21:22
blackburnelse regression21:22
@sonney2kahh ok... I thought we do it more clever21:23
blackburnok we can add some factory actually21:23
@sonney2kby checking if the machine needs a certain label type21:23
@sonney2kwe create the label only then21:23
blackburnno that's too complicated I think21:23
@sonney2kwith correct type21:23
blackburnhowever can be done too21:24
blackburnI like both21:24
blackburnpuffin444: so do they liked shogun? :D21:24
blackburnn4nd0: what do you think about apply_structured21:25
n4nd0blackburn: yeah, it should be done like that indeed21:25
-!- vikram360 [~vikram360@] has quit [Ping timeout: 250 seconds]21:25
puffin444blackburn: Yes, they were impressed by such an international team (We have 7+ countries represented)21:25
blackburnhmm really?21:25
n4nd0blackburn: to follow the same structure there21:26
n4nd0puffin444: oh true, how did you presentation go?21:26
puffin444n4nd0: It went well. I was actually able to sort of explain generally GPs in only five mins21:26
puffin444Their office in Chicago has a beautiful view21:27
@sonney2kpuffin444, do you have the slides somewhere?21:27
@sonney2kblackburn, there is one problem21:27
@sonney2kwe have a labels load function21:28
puffin444sonney2k: Yes, I do.21:28
puffin444sonne2k: All four of them :)21:28
blackburnload like load from file?21:28
@sonney2kso we need the label type before21:28
@sonney2kblackburn, yes21:28
@sonney2kpuffin444, care to share?21:28
puffin444sonney2k: sure, Do I put them on the mailinglist or somehow transfer them through irc?21:29
@sonney2kpuffin444, some official post on the mailinglist would be nice - but you can also put them somewhere and link to them21:29
@sonney2kpuffin444, or even blog about it on shogun planet :D21:29
blackburnI actually would like to present shogun somewhere too :D21:30
blackburn(okay I admit it can be an easy way to visit some country for free)21:30
puffin444sonney2k: I could try blogging about it. How do I get subscription access to shogun planet?21:36
@sonney2kpuffin444, create a blog/rss feed somewhere and tell me the feed url21:39
@sonney2kblackburn, most certainly welcome :D21:39
blackburnsonney2k: yeah just need to find some conference I think?21:44
@sonney2kand get travel money from uni or so21:45
@sonney2kblackburn, my proposal wrt loading labels is21:47
@sonney2kI load regression labels21:47
@sonney2kand then we do the automagic conversion again21:48
blackburnhmm yes makes sense21:50
-!- ckwidmer [] has joined #shogun21:56
@sonney2kthe chris21:57
ckwidmerhi all21:57
puffin444hey chris21:57
puffin444eating lunch. be back later.21:57
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has quit [Quit: Page closed]21:58
@sonney2kblackburn, there is an issue with the automagic label stuff21:59
@sonney2kcurrently our exmaples use +1/-1 labels for regression22:00
@sonney2kblackburn, so what do we do now?22:03
@sonney2kcreate examples with real valued labels?22:04
blackburnsonney2k: hmm or act in the way you suggested22:04
blackburndepending on machine22:04
-!- ckwidmer [] has quit [Remote host closed the connection]22:18
@sonney2kblackburn, any idea how to add anything to data?22:26
@sonney2kI did add regression labels22:26
blackburnsonney2k: to data like?22:26
@sonney2kbut git push sais everything up-to-date22:27
blackburnah shogun-data?22:27
blackburnsonney2k: hmm probably you didn't 'activate' it as git repo22:28
blackburnI tried to remove toy here22:29
blackburnsonney2k: git submodule update --init?22:30
CIA-113shogun: Soeren Sonnenburg master * r99a35be / (13 files in 6 dirs): fix static interfaces -
@sonney2kblackburn, no I mean I want to add the data set22:34
@sonney2kcommit & push it to shogun-data22:34
@sonney2kbut it claims that this repo is up-to-date22:34
@sonney2keven though I have the new data set in there now22:34
blackburnsonney2k: yes - because it considers data as main git22:34
shogun-buildbotbuild #990 of libshogun is complete: Failure [failed git]  Build details are at  blamelist: sonne@debian.org22:35
blackburnis it fresh clone? did you set up submodules?22:35
@sonney2kblackburn, I go to shogun/data22:35
@sonney2khow do I convince git that this is not part of shogun/ ?22:36
blackburnsonney2k: yes but did you 'git submodule update --init' in shogun/?22:36
blackburnor somehow22:36
@sonney2kblackburn, yeah to fetch the data22:36
@sonney2kthen I modified the data22:37
blackburnhmm that's strange then22:37
@sonney2kadded a new remote22:37
@sonney2kand push now says f... you22:37
blackburnsonney2k: one idea22:38
blackburnsonney2k: git status?22:38
blackburn# Not currently on any branch.22:38
blackburnsave your data and recommit if so22:39
@sonney2kyeah indeed22:41
CIA-113shogun: Soeren Sonnenburg master * r23dfd86 / data : add new data -
blackburnsonney2k: interesting fact it all goes to valhalla if not on branch22:44
@sonney2kblackburn, yeah sometimes git is dangerous22:44
CIA-113shogun: Soeren Sonnenburg master * r7866158 / (2 files): forgot .dat suffix -
blackburnsonney2k: I am afraid you put labels creation into the loop22:52
blackburnah right22:52
shogun-buildbotbuild #991 of libshogun is complete: Success [build successful]  Build details are at
@sonney2kblackburn, heh22:53
@sonney2kclever me :D22:53
blackburnsonney2k: any idea how to boost accuracy for 0.5-1%?22:54
blackburnmulticlass, you know22:54
@sonney2kgsomix, in case you get the director kernels to work there is another cool python feature (python buffer protocol) that we all want :)22:54
@sonney2kblackburn, what?22:54
blackburnsonney2k: I desperately need to get 0.98 instead of 0.97 :D22:55
blackburnmaybe some tricks?22:55
@sonney2kno clue22:55
gsomixsonney2k, aha, aha.22:56
@sonney2kblackburn, could you fix these warnings
blackburnsonney2k: any idea why liblinear computing time -> infty after prunevarsubmean?22:56
shogun-buildbotbuild #893 of cmdline_static is complete: Success [build successful]  Build details are at
blackburnhmm my issues are all strange22:57
@sonney2kgsomix, ahh and I think we need you to turn on the tests afterwards ...23:03
blackburnsonney2k: which normalization is better for linear SVM? 1/2?23:06
gsomixsonney2k, which tests?23:07
shogun-buildbotbuild #782 of octave_static is complete: Success [build successful]  Build details are at
@sonney2kblackburn, depends on problem sqrtdiag quite often23:11
@sonney2kbut it also helps to use binneddotfeatures23:12
blackburnsonney2k: what is advantage of using binned features?23:12
@sonney2kyou can learn piece wise linear functions23:12
@sonney2kso with enough bins basically any function23:13
@sonney2kworks pretty well23:13
@sonney2kand no normalization issues23:13
shogun-buildbotbuild #859 of python_static is complete: Success [build successful]  Build details are at
blackburnsonney2k: with which features?23:13
blackburnanything special?23:13
-!- PolarBearAttack [46e7e626@gateway/web/freenode/ip.] has joined #shogun23:13
PolarBearAttackWhat is shotgun toolbox about?23:13
blackburnshotgun :D23:13
PolarBearAttackshogun, sorry23:14
blackburnI'd say its name is shogun23:14
blackburnmachine learning23:14
@sonney2kworld domination23:14
@sonney2kego trips23:14
PolarBearAttackHow do you acquire such practice?23:14
PolarBearAttackrequirements? how do you know23:14
blackburnPolarBearAttack: for example I use it to construct road sign recognition 'system'23:15
-!- PolarBearAttack was kicked from #shogun by sonney2k [PolarBearAttack]23:15
-!- PolarBearAttack [46e7e626@gateway/web/freenode/ip.] has joined #shogun23:15
blackburnwhy :)23:15
@sonney2khow do we disable that bot?23:15
@sonney2kblackburn, is that your grisly?23:15
PolarBearAttacklet me be part of the society?23:15
blackburnhmm I did not recognize it as bot :D23:15
blackburnPolarBearAttack: are you?23:16
n4nd0sonney2k: it might dangerous to kick polar bears :D23:16
@sonney2kthey can bite23:16
@sonney2kPolarBearAttack, can you?23:16
PolarBearAttackI'm coherent, I will contribute23:16
blackburnsonney2k: can you please describe how can I appoach anything with binned features?23:17
@sonney2kblackburn, are you coherent and will contribute?23:18
blackburnsonney2k: I promise I will23:18
@sonney2kPolarBearAttack, see23:18
@sonney2kblackburn, has a grisly at his home23:18
@sonney2kwe are not afraid...23:19
@sonney2kblackburn, you have to give it the bins23:19
PolarBearAttackCan people talk in terms of mathematics?23:19
blackburnPolarBearAttack: f(x)23:19
@sonney2kthere is an example23:19
@sonney2khave a look23:20
blackburnsonney2k: bins like [0,1] (1,2]23:20
@sonney2kI think so23:20
@sonney2kblackburn, how many dims do you have?23:20
blackburnsonney2k: depends how many hog cells I have but generally varying from 200 to 90023:20
@sonney2kPolarBearAttack, f(x)=\sum_i \alpha_i y_i K(x_i, x)23:20
@sonney2kblackburn, so you can easily use 100 bins per feature23:21
@sonney2kblackburn, how many examples?23:21
PolarBearAttacksonney2k: , how did you formulate that thought?23:21
blackburnsonney2k: 39K23:21
@sonney2knot so many23:21
blackburnhmm I could generate some with rotations, etc :D23:21
@sonney2kPolarBearAttack, I used a brain computer interface23:22
-!- mode/#shogun [+o blackburn] by sonney2k23:22
-!- mode/#shogun [-o blackburn] by ChanServ23:22
PolarBearAttackdid you subvocalized?23:22
blackburnnot registered23:22
PolarBearAttackFormulating thought in terms of symbol is pretty powerful23:23
n4nd0this is getting awkward ...23:23
blackburnPolarBearAttack: all the machine learning can be formulated as searching f such that f: X->Y!23:23
-!- mode/#shogun [+b *!*46e7e626@*.] by sonney2k23:23
-!- PolarBearAttack was kicked from #shogun by sonney2k [PolarBearAttack]23:23
blackburnwhy that was pretty funny23:23
@sonney2kblack /query him23:25
@sonney2kif you want to talk to him/it23:25
blackburnsonney2k: argh I forgot with HKM I get *3 features23:25
@sonney2kbed time!23:25
n4nd0sonney2k: gute Nacht23:26
shogun-buildbotbuild #873 of r_static is complete: Success [build successful]  Build details are at
gsomixgood night guys23:37
n4nd0good night gsomix23:39
shogun-buildbotbuild #563 of csharp_modular is complete: Failure [failed test_1]  Build details are at  blamelist: sonne@debian.org23:56
--- Log closed Fri May 25 00:00:41 2012