Open in new window / Try shogun cloud
--- Log opened Thu Apr 18 00:00:28 2013
-!- _rohan [~rohan@] has quit [Quit: Leaving]00:18
-!- shogun-notifier- [] has quit [Quit: transmission timeout]00:48
-!- iglesiasg [] has quit [Quit: leaving]01:56
-!- shogun-notifier- [] has joined #shogun02:56
shogun-notifier-shogun: Viktor Gal :develop * 377fbcb / src/shogun/ui/SGInterface.cpp:
shogun-notifier-shogun: Fix SubgradSVM and SubgradLPM removal in SGInterface02:56
shogun-notifier-shogun: Viktor Gal :develop * 5bfeb95 / .travis.yml:
shogun-notifier-shogun: Travis: remove commas from package list for ruby03:00
-!- travis-ci [] has joined #shogun03:24
travis-ci[travis-ci] it's Viktor Gal's turn to pay the next round of drinks for the massacre he caused in shogun-toolbox/shogun:
-!- travis-ci [] has left #shogun []03:24
-!- abinash [1261415@gateway/web/cgi-irc/] has joined #shogun03:33
-!- travis-ci [] has joined #shogun03:52
travis-ci[travis-ci] it's Viktor Gal's turn to pay the next round of drinks for the massacre he caused in shogun-toolbox/shogun:
-!- travis-ci [] has left #shogun []03:52
-!- abinash [1261415@gateway/web/cgi-irc/] has quit [Quit: - A hand crafted IRC client]03:53
-!- abinash [1261415@gateway/web/cgi-irc/] has joined #shogun03:55
shogun-notifier-shogun: Viktor Gal :develop * d53a898 / .travis.yml,src/configure:
shogun-notifier-shogun: Fix ruby call in test_ruby_narray in configure script04:08
shogun-notifier-shogun: Try fixing ruby-narray error by enforcing to use ruby1.804:08
-!- abinash [1261415@gateway/web/cgi-irc/] has quit [Quit: - A hand crafted IRC client]04:11
-!- foulwall [] has joined #shogun04:21
foulwallsonney2k: gsomix: I've sent a pull request . Check it pls and I'll start the svm demo04:22
-!- Yanglittle [deb20afb@gateway/web/freenode/ip.] has joined #shogun04:52
Yanglittleis there any body?05:02
-!- foulwall [] has quit [Ping timeout: 256 seconds]05:03
-!- abinash [1261469@gateway/web/cgi-irc/] has joined #shogun05:06
-!- Yanglittle [deb20afb@gateway/web/freenode/ip.] has quit [Quit: Page closed]05:12
-!- abinash [1261469@gateway/web/cgi-irc/] has quit [Quit: - A hand crafted IRC client]05:30
-!- wiking [~wiking@huwico/staff/wiking] has quit [Read error: Connection reset by peer]05:59
-!- wiking [] has joined #shogun05:59
-!- vikram360 [~vikram360@] has quit [Ping timeout: 245 seconds]06:27
-!- sumit_ [73f91219@gateway/web/freenode/ip.] has joined #shogun06:41
-!- lambday [3ba21749@gateway/web/freenode/ip.] has joined #shogun06:53
-!- shogun-notifier- [] has quit [Quit: transmission timeout]07:08
-!- lambday [3ba21749@gateway/web/freenode/ip.] has quit []07:12
-!- saket [0e8b5206@gateway/web/freenode/ip.] has joined #shogun07:15
-!- sumit_ [73f91219@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]07:20
saketI am working on making a tree in shogun07:41
saketI would like to calculate the dimension in the (CFeatures*) data which has maximum variance07:42
saketis there any already function for doing so?07:42
saketanyone there?07:45
-!- vikram360 [~vikram360@] has joined #shogun07:47
-!- k_matsuda [] has quit [Remote host closed the connection]07:51
-!- hoijui [] has joined #shogun08:03
saketanyone there?08:05
sonne|worksaket: no there is no function like that08:49
sonne|workso this would be a first small welcome patch08:50
saketactaully, I needed that function for another patch ;p08:50
saketI see that load(CFile) is also not implemented ..Can you please guide me on where (which code) to read about accessing reals from CFeatures08:51
saketI see methods like "get_feature_matrix" and "get_feature_vector(index)" as functions in the python interface08:57
saketbut am i unable to locate them in the src folder08:57
saketplease help me out with this ..08:58
-!- iglesiasg [] has joined #shogun09:01
-!- mode/#shogun [+o iglesiasg] by ChanServ09:01
saketsonne|work: I will certainly have get_var() in the patch i submit :-)09:04
-!- Yanglittle [deb20af9@gateway/web/freenode/ip.] has joined #shogun09:06
sonne|worksaket: it is in CDenseFeatures and SGMatrix09:07
saketsonne|work: thanks :) I will have a look09:08
-!- vikram360 [~vikram360@] has quit [Ping timeout: 264 seconds]09:23
-!- vikram360 [~vikram360@] has joined #shogun09:23
-!- Foulwall [~vkmahfa@] has joined #shogun09:34
FoulwallHello guys09:35
-!- Foulwall [~vkmahfa@] has quit [Client Quit]09:36
@iglesiasgsonne|work: I guess it is fine if I send the mail of the workshop around in the department where I am doing my master thesis?10:11
sonne|workiglesiasg: encouraged even!10:13
@iglesiasgcool, I will do it soon then10:14
sonne|workwiking: so what do we do about the failing 1/2 ?10:16
sonne|workI have absolutely no idea how to address that10:16
sonne|workthe only thing we can do is print the objects...10:17
@iglesiasgsonne|work: let me know if I can help with this somehow, I don't really know how the serialization works but maybe I could prepare some tests or sth10:18
-!- saket [0e8b5206@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]10:19
sonne|workiglesiasg: the problem really is that this test works on all our machines but it does not on the travis buildbot10:21
sonne|workabsolutely no idea why not10:21
@iglesiasgaham I see10:21
Yanglittleis there any body?10:23
Yanglittle I have a problem , mkl.getsubkernelweights(numkernel), and it always say type error, and what is "int32_t &"  is there any body can help me10:27
-!- abinash [75ef5e6e@gateway/web/cgi-irc/] has joined #shogun10:29
@lisitsynYanglittle: as I said in my mail try it without any parameter10:30
@lisitsynYanglittle: depends on what mkl here is though10:31
Yanglittlei tried , it expected a argument.10:33
@iglesiasgYanglittle: so you are using python_modular, don't you?10:36
@iglesiasgYanglittle: I think it can be a problem with the type, that you have to say explicitily that what you are giving is an int32_t10:37
@iglesiasgbut it is hard to tell if I cannot see the code you are using and the exact error10:37
Yanglittleyeah ,ok  I run the mkl_multiclass_modualr and add the code "mkl.getsubkernelweights(3)" after mkl.train()10:40
Yanglittleit says:  The debugged program raised the exception TypeError "in method 'MKLMulticlass_getsubkernelweights', argument 2 of type 'int32_t &'" File: /home/liyang/shogun/examples/documented/python_modular/, Line: 7510:40
@iglesiasgYanglittle: ok let me try it here10:40
@lisitsynYanglittle: try mkl.get_kernel().get_subkernel_weights()10:43
@lisitsynYanglittle: or rather10:44
@lisitsynjust call get_subkernel_weights() on your combined kernel10:44
@iglesiasglisitsyn: it is because of the & thing, no way to do that in python I think10:44
@lisitsyniglesiasg: it is legacy from the days we had no sgvector10:45
@iglesiasglisitsyn: mmm why to use & when there is no sgvector?10:45
@iglesiasgit is not  like passing a pointer, int32_t*10:45
@iglesiasgprobably I am getting something wrong10:45
@lisitsyniglesiasg: yes there should be a pointer too but the spirit is the same :)10:46
@iglesiasglisitsyn: mmm I don't understand10:46
@lisitsyniglesiasg: before sgvector getters were like10:47
@lisitsyn(float64_t**, int32_t&)10:47
@iglesiasgI understood the int32_t numweights as an int by reference10:47
Yanglittleok, now it can works, thanks very much!10:47
@iglesiasgah ok I see it now10:47
@iglesiasgthe int32_t& is to output the size of the float64_t* returned10:47
-!- abinash [75ef5e6e@gateway/web/cgi-irc/] has quit [Quit: - A hand crafted IRC client]10:51
-!- abinash [75ef5e6e@gateway/web/cgi-irc/] has joined #shogun11:00
-!- iglesiasg [] has quit [Quit: leaving]11:11
-!- iglesiasg [] has joined #shogun11:12
-!- mode/#shogun [+o iglesiasg] by ChanServ11:12
-!- sumit_ [73f91219@gateway/web/freenode/ip.] has joined #shogun11:14
-!- abinash [75ef5e6e@gateway/web/cgi-irc/] has quit [Quit: - A hand crafted IRC client]11:15
-!- abinash [75ef5e6e@gateway/web/cgi-irc/] has joined #shogun11:17
-!- abinash [75ef5e6e@gateway/web/cgi-irc/] has quit [Client Quit]11:21
-!- abinash [75ef5e6e@gateway/web/cgi-irc/] has joined #shogun11:24
-!- abinash [75ef5e6e@gateway/web/cgi-irc/] has quit [Client Quit]11:25
-!- abinash [75ef5e6e@gateway/web/cgi-irc/] has joined #shogun11:26
-!- abinash [75ef5e6e@gateway/web/cgi-irc/] has quit [Client Quit]11:28
-!- shogun-toolbox [] has quit [Ping timeout: 256 seconds]11:31
--- Log closed Thu Apr 18 11:31:44 2013
--- Log opened Thu Apr 18 11:33:03 2013
-!- shogun-toolbox [] has joined #shogun11:33
-!- Irssi: #shogun: Total of 19 nicks [3 ops, 0 halfops, 0 voices, 16 normal]11:33
-!- Irssi: Join to #shogun was synced in 66 secs11:33
-!- vikram360 [~vikram360@] has quit [Ping timeout: 252 seconds]11:36
-!- HeikoS [] has joined #shogun11:39
sonne|workhei HeikoS - please send around all the CfPs to whoever you can think of :D11:42
-!- shogun-toolbox [] has quit [Ping timeout: 245 seconds]11:42
--- Log closed Thu Apr 18 11:42:24 2013
--- Log opened Thu Apr 18 11:42:30 2013
-!- shogun-toolbox [] has joined #shogun11:42
-!- Irssi: #shogun: Total of 19 nicks [3 ops, 0 halfops, 0 voices, 16 normal]11:42
-!- Irssi: Join to #shogun was synced in 6 secs11:42
HeikoSsonne|work: will do :)11:44
-!- hoijui [] has quit [Ping timeout: 246 seconds]11:45
-!- k_matsuda [] has joined #shogun11:51
@lisitsynHeikoS: is there any difference between berlin airports?12:21
HeikoSlisitsyn: better ask sonney2k12:21
@lisitsynsonne|work: ^12:21
HeikoSDont know12:21
HeikoSalways go there by car :)12:21
@lisitsynHeikoS: where are you going to stay btw?12:22
HeikoSlisitsyn: dont know, probably friends12:22
HeikoSor a hostel12:22
HeikoSI can let you know places, but again, sonney2k probably knows the place a bit better12:23
@lisitsynHeikoS: any estimate on cost of staying there?12:23
HeikoSlisitsyn: I never stayed in a hotel there12:23
HeikoSoh yes once12:23
HeikoSa very nice hostel12:23
HeikoSbut I forgot the price ;)12:23
HeikoSnot too much12:23
@lisitsynyay germany visa is >200$ for me12:23
@lisitsynHeikoS: I should also prove I have enough money to get back :D12:28
HeikoSlisitsyn: haha .)12:29
HeikoSdo you? :D12:30
@lisitsynHeikoS: I hope so!12:30
wikingsonne|work: yo12:31
wikingsonne|work: i have no idea why that python example fails12:31
wikinglisitsyn: you need sponsors? :D12:33
@lisitsynwiking: like sponsor that would take care of me?12:34
@lisitsynfor visa?12:34
wikingfor visa12:35
wikingor you are going with an agency?12:35
@lisitsynwiking: I am undecided - will take care of this in next few days12:35
wikinglisitsyn: if u need an invitation letter let me know12:36
@lisitsynwiking: but you can't invite me to germany right? :)12:36
wikingi can12:37
wikingi mean not to germany12:37
wikingbut to EU12:37
wikingi mean schengen12:37
wikingwhich is pretty much the same12:37
-!- sumit_ [73f91219@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]12:37
@lisitsynwiking: I just checked - I have to have some place to stay to get visa12:38
wikingi mean if u get the visa (schengen) you would be able to travel within schengen zone12:38
@lisitsynyes I know12:38
wikingwell i dont know if sonne|work can offer you a place12:38
wikingbut if that's too complicated12:38
@lisitsynhah I guess no12:39
wikingwe can tell that you are going to stay at my place in budapest12:39
wikingand that's all12:39
wikingyou get the visa for hungary, which is basically a visa for germany as well12:39
wikingand then it should be all good12:40
@lisitsynI see12:40
wikingif u want one of these days i can call the russian embassy in budapest12:40
wikingto ask them for most of the things...12:40
@lisitsynyou are helpful! thanks :)12:40
wikingbut maybe - if u decide to go this way - you should call the hungarian embassy in russia12:40
wikingjust to be sure about the papers we need to provide them12:41
@lisitsynI'll reach some agency and ask them12:41
wikinglisitsyn: +7-495-796-937012:41
@lisitsynwiking: where are you going to stay in berlin btw?12:41
HeikoSlisitsyn, wiking isn't it easier to do all this in Germany? I mean get a place to stay there?12:41
HeikoSlisitsyn: Once you booked a hostel you have a place to stay12:41
wikingHeikoS: hehe it's not that easy12:42
HeikoSsonney2k can give you a list12:42
wikingHeikoS: it's much better if he can tell to the embassy12:42
wikingthat he is staying at some persons private place12:42
HeikoSwell then, go ahead :)12:42
wikingbecause EU is an asshole12:42
wiking..? :)12:42
@lisitsynHeikoS: yes - but I just don't know if they provide proper papers for german visa12:42
wikingand they are even more assholes with russians...? :)12:42
@lisitsynwiking: haha12:42
wikingsorry but this is the truth12:43
wikingonce we had to apply for a schengen zone visa for a friend of mine12:43
wikingwho was from middle-east12:43
wikingit was hell12:43
wikingi was waiting when are they going to ask for a dna sample of my friend12:43
wikingand on the end by 'their mistake' they refused to give him the visa12:43
wikinglisitsyn: if the hungarian embassy in russia12:44
wikingcan provide schengen visa12:44
wikingthen you are afaik much better off than germans12:44
wikingjust because of the history between the 2 countries12:44
@lisitsynwiking: heh12:44
@lisitsynyeah may be12:44
wikingyou know old communist comrades ;)12:44
wikingbut then again12:45
wikingin some countries they cannot issue schengen visas12:45
wikingand it's done by another embassy12:45
wikinglisitsyn: so if u can please call these guys in moscow12:45
wikingand ask them12:45
@lisitsynwiking: HeikoS: I am asking where you are going to stay as we may book the same hotel?12:45
wikinglisitsyn: i have 0 ideas yet12:46
wikingi think i'm going to fix this in may some time12:46
@lisitsynwiking: my plan is to check about direct -> germany thing and then compare12:46
@lisitsynwiking: so you have hungarian citizenship rite?12:47
wikinglisitsyn: as well, yes12:47
@lisitsynwiking: you have some more? :)12:48
wikinglisitsyn: yes seribna12:48
wikingso yeah for the hungarians we can provide all stuff like: i am a 'sponsor' of your stay during your visit and that you can stay at my flat during the time your visit12:49
wikinganyhow somebody has any ideas with the serialization issue on travis?12:50
@lisitsynthat's cool12:50
wikingi saw it once failed on buildbot as well12:50
@lisitsynwiking: thanks! let us consider that if germany deny me :D12:50
wikinglisitsyn: okok just be on time12:50
wikinglisitsyn: it does take them some fair amount of time to decide anything12:51
@lisitsynyeah I should take care of that next 2 weeks or I could get out of time12:51
wikinglisitsyn: yeah i really advise you to start it latest next week12:51
@lisitsynI wanted to start early but postponed postponed postponed12:51
wikingfor my friend it took them ages to do anything12:51
wikingand then on the end they denied12:51
wikingand 1.5 week before his trip12:51
@lisitsynwiking: some agencies here say they can make a german visa for 7500 roubles12:52
wikingthey told him that 'aaaahh we made a mistake'....12:52
wikingbut it was too late12:52
wikingand he couldnt do the trip anymore12:52
wikinglisitsyn: how much is that in usd?12:52
@lisitsyndivide by 3012:52
wikingohh shiiiiiit12:52
wikingthat's craaazy expensive12:52
@lisitsynwiking: but I wouldn't have to go to moscos12:52
wikingaaah otherwise you *have* to?12:52
@lisitsynwiking: yeah may be12:53
wikingi mean i know this friend of mine did it via mail12:53
@lisitsynwiking: german embassy says they want an interview12:53
wikingbtw: any of these places are nearer to you than moscow:
@lisitsynwiking: I am just like a cluster in k-means12:54
@lisitsynall of these are ~1000 km :D12:54
wikingah yeah you are in samara12:54
wikingcall that number i gave you12:54
@lisitsynwiking: yes, toliatty currently12:54
wikingjust to ask them12:54
wikingwhat are their requirements12:55
wikingmaybe they dont want an interview with you12:55
@lisitsynwiking: alright12:55
wikinglisitsyn: as far as i understood from some papers12:57
wikingyou as a russian citizen12:57
wikinghave some 'light' criterions for getting the schengen visa12:58
@lisitsynyeah it must be much easier than for your friend from middle east12:58
wikingHeikoS: ping13:01
wikingHeikoS: wtf is happening in that serialization script?13:02
HeikoSwiking:  which one?13:02
wikingcomplex serialization...py13:02
HeikoSwiking: dont know, did I write that?13:03
HeikoSthis one keeps on failing on my local  machine13:04
HeikoSsonney2k recently did some fixes for serialization13:04
HeikoShaven't looked into it for a while13:04
wikingbtw: anybody knows what's the hard limit for a repository size on github13:06
@lisitsynwiking: I have seen some info where they say 1gb is soft limit13:07
@lisitsynand when you reach 1gb they ask you to reduce the size13:07
-!- foulwall [] has joined #shogun13:11
wikingthat should be enough for me.. thnx lisitsyn13:13
foulwallsonney2k: Hey. maybe some more concrete suggests on my new pull request? I'm refactoring the code.13:15
@lisitsynwiking: okay other agency says 80EUR for visa13:22
-!- Yanglittle [deb20af9@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]13:34
wikinglisitsyn: heheh yeah i saw that yesterday or something13:41
wikingbtw we should have unit testing as well on travis i guess13:43
HeikoSwiking: yes, that would be very cool!13:46
wikingHeikoS: it's not complicated13:48
wikingthe only problem is that somehow it shoudl be done13:48
HeikoSyeah as always :)13:48
wikingonly once13:48
HeikoSI like this travis stuff13:49
wikingso only for the libshogun run13:49
wikingnot for all the modular cases13:49
HeikoSyes of course13:49
wikingwhich makes it of course a bit more complex :P13:49
HeikoSwe have the modular examples, that should be fine13:49
HeikoSunit tests with valgrind would also be nice :)13:49
wikingHeikoS: well that's only one more package :P13:50
HeikoSyeah :)13:50
wikingHeikoS: what's the command for running that? :)))13:51
wikingHeikoS: i know i've written it but still13:51
HeikoSno idea13:51
HeikoSI usually do cat Makefile | grep valgrind :D13:51
HeikoSI try hard *not* to remember those things in order to keep my brain tidy  ;)13:52
sonne|workfoulwall: what do you mean?13:52
wikinganyhow first i debug something because i dont get some shit...13:52
foulwallsonne|work: I am trying to make the code as most as reusable as I can.13:54
sonne|workfoulwall: I guess the three parts: getting  input / shogun learning / applying & output should be enough?13:55
foulwallI wrote a base class called train&learn model13:56
foulwallclustering and svr is  derived  from the train&learn13:57
foulwallbut the split up is not easy to do. I'm trying to find a general method.13:58
-!- foulwall [] has quit [Ping timeout: 252 seconds]14:08
-!- foulwall [] has joined #shogun14:08
foulwallsonne|work: Can you see me?14:09
-!- vgorbati [d4029f22@gateway/web/freenode/ip.] has joined #shogun14:20
wikingi cannot unref an svm14:21
@lisitsynwiking: why?14:23
@lisitsynit frightens you?14:23
-!- gzhd79 [] has joined #shogun14:27
sonne|workfoulwall: what do you mean?14:31
sonne|worklisitsyn: btw do you send out confirmation emails or anything to the people registering for the workshop?14:33
@lisitsynsonne|work: will send a bunch today14:33
@lisitsynsonne|work: who are subscribers of that address?14:34
@lisitsyninternal? should I CC it?14:34
sonne|workyou and me so far14:34
foulwallsonne|work: do we need a model for all demo which contains an train and apply process or just split the function14:37
sonne|workfoulwall: just split it for now. However data input is very similar14:38
foulwallsonne|work: Ok.14:38
-!- vikram360 [~vikram360@] has joined #shogun14:49
wikinglisitsyn: for binary svm which has better performance (speed): libsvm or svmlight, or it all depends?14:51
@lisitsynwiking: kernel?14:51
wikingbtw i might need to test as well liblinear + HKM :P14:52
@lisitsynit depends of course but I'd expect svmlight could be faster14:52
-!- deerishi [73f88294@gateway/web/freenode/ip.] has joined #shogun14:53
@lisitsynwiking: HKM + liblinear rocks :)14:53
@lisitsynwiking: at least it makes you able to find better features / settings14:53
-!- foulwall [] has quit [Remote host closed the connection]15:30
-!- deerishi [73f88294@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]15:32
-!- deerishi [73f88294@gateway/web/freenode/ip.] has joined #shogun15:39
-!- iglesiasg [] has quit [Quit: leaving]15:40
sonne|workwiking: it really depends... in my experience if you have lots of data from one class and not so much from the other svmlight is faster15:47
wikingsonne|work: it's kind of like equaly distributed15:48
sonne|worksvm light better benefits multiple cores15:50
sonne|workbut otherwise libsvm might be a bit faster15:50
sonne|workmaybe more stable too15:50
deerishiHeikoS: u told me to implement a scik it example on KPCA,could you guide me a bit on how to go about it.i have been trying the shogun KPCA library,but so far did not get anything good15:52
HeikoSdeerishi: hi!15:52
HeikoSdeerishi: what I meant is just to use the data in the scikit learn and produce similar looking plots15:52
HeikoSusing shogun's KPCA implementation15:52
HeikoSyour code is fine15:52
HeikoSbut I would use a different example15:52
HeikoSthe scikit example is very nice since it illustrates the method very nicely15:53
deerishijust one quick question : shogun's does explicitly gives back the modified data KPCA ?the code i wrote gave back the densities of the data on a line15:54
deerishiohh.i think i have an idea,will try it out and let you know15:56
HeikoSdeerishi: no15:56
HeikoSit gives you a  matrix15:56
HeikoSwhich is DxN where D are the number of eigevectors you use15:56
HeikoSyou just have to plot carefully15:56
HeikoSpairs of points15:57
HeikoSdeerishi: you should be almost there! :)15:57
deerishii hv an idea,will work on it tonight.thanks!!15:58
HeikoSdeerishi: thanks for keeping on it, see you!15:58
-!- gsomix_ [~gsomix@] has joined #shogun15:58
-!- thoralf [] has joined #shogun16:00
sonne|workHo :)16:01
-!- gsomix [~gsomix@] has quit [Ping timeout: 245 seconds]16:01
-!- gzhd79 [] has quit [Quit: gzhd79]16:03
-!- gzhd79 [] has joined #shogun16:06
wikinglisitsyn: svmsgd seems to be cool ;)16:07
wikingi mean much faster for my dataset :P16:07
sonne|workwiking: well it just does x passes over the data16:11
sonne|workso no wonder16:11
wikingmmmm but unfortunately all the options16:12
wikingare way worse than random forest :(16:12
wikingi guess we dont have random forest, yet :P16:12
wikingpnorm preprocessor segfaults16:14
sonne|workthoralf: better chat in the channel :)16:15
sonne|workyeah I know I did write the hashing stuff in shogun16:15
sonne|workwiking: backtrace?16:15
sonne|workhard to believe since it is so simple16:16
wikingsomething is really weird16:24
wikingi have an SGMatrix16:24
wikingand with memcpy i fill it up16:24
wikingand right away after memcpy i can even print out the right values16:24
wikingbut then when i do on the whole matrix16:24
wikinga display_matrix()16:24
wikingit's a full 0 matrix16:24
wikingi mean the output16:24
wikingany ideas?16:26
-!- sids_aquarius [] has joined #shogun16:30
wikingok seem display_matrix is aint working16:30
sonne|workI used that and it worked16:33
wikingcoz it perfectly works if i put that matrix into a densefeatures class and then do a 1-by-1 feats->get_feature_vector(i); display_vector()16:33
wikingbut if i just do a display_matrix it's like as if the matrix is a big 016:34
sonne|workwiking: please put some test code somewhere in a github issue and mark it entrance...16:35
wikingsonne|work: ok will do16:35
vgorbatihi all, I am a student wisihing to participate in GSoC 2013. To get familiar with shogun I have forked, cloned the shogun-git workspace and compiled it. To continue, I want to try working on the issue Can anyone talk with me about it?16:37
wikingnow more the question is16:37
wikingthat why pnorm dies16:37
wikingsonne|work: it works fine with a small SGMatrix16:38
wikingand it dies with a big one16:38
sonne|workbig == ?16:38
wikingon a 19*1816:38
wikingit works fine16:38
thoralfsonne: Yes, I was wondering... since your name appears in the copyright notice of all the hashing kernels. ;)16:40
wikingsonne|work: hahaha i think i know why, although not 100%16:43
wikingsonne|work: u reckon this could happen when i apply PNorm (which works with fixed float64_t) on a DenseFeatures<float32_t> ?16:43
wikingi mean memory access problems16:43
wikingimo PNorm should be templated as well16:44
-!- gsomix_ [~gsomix@] has quit [Ping timeout: 272 seconds]16:55
thoralfI am trying to process text data with shogun, but the examples I found only deal with DNA data, producing one feature per CHAR.16:57
thoralfCan anyone give a pointer to a piece of code with WORD features? :)16:58
thoralfThe class StringWordFeatures pretends to be the right one, but I did not find out how to use it.16:59
thoralf(The class name within the python bindings.)17:01
-!- deerishi [73f88294@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]17:04
sonne|workthoralf: yeah that would be task too - we don't have features for text17:12
sonne|workthe word in there just denotes 2 bytes17:12
thoralfsonne: Thanks - at least I don't need to search any more.17:14
thoralfsonne: A preprocessing using NLTK wouldn't be hard, so I think this might be the way to go.17:15
wikingsonne|work: when i template a class why i dont have m_parameter  ? :)17:19
sonne|workwiking: what?17:26
-!- gzhd79 [] has quit [Quit: gzhd79]17:27
sonne|workthoralf: it would be nice to have some 'text document' -> hashed list of words / n-grams in shogun though17:27
wikingsonne|work: any comments on this diff?17:30
thoralfsonne: Parsing text in C/C++ is not the funniest thing to do, so actually I was trying to avoid this part. ;)17:33
-!- sumit [73f91219@gateway/web/freenode/ip.] has joined #shogun17:33
-!- gzhd79 [] has joined #shogun17:35
-!- foulwall [~foulwall@2001:da8:215:c252:193a:a95f:9fad:517a] has joined #shogun17:40
foulwallHello dudes.17:40
wikingsonne|work: the only thing is that in this case i cannot get SG_ADD working17:40
sonne|workfoulwall: hello17:41
sonne|workthoralf: yes indeed but I guess there is no way around this17:41
foulwallsonney2k: I've pulled a request. check it pls. I'm wondering the reliability of the code.17:42
sonne|workwiking: why do you need the templated thing?17:42
wikingsonne|work: well this is the thing17:42
wikingsonne|work: i've got segfault17:42
wikingsonne|work: when i applied the current PNorm on a DenseFeatures<float32_t>17:42
wikingsonne|work: since of course sizeof(float32_t) < sizeof(float64_t)17:43
wikingsonne|work: of course the best would be that the user doesn't have to give17:43
wikingthe PNorm<T>17:43
wikingas it should be known where it's applied17:43
wikingi mean the T type in the Features17:44
wikingso this should be somehow detected runtime :P17:44
sonne|workfoulwall: looks much more readable now17:44
@lisitsynwhy do we need templated features/preprocessors/etc exposed to the interfaces?17:44
wikinglisitsyn: any idea how to solve this in case of PNorm?17:44
foulwallsonne|work: I'm looking for a way to hide the related but not public method in the class.17:44
@lisitsynwiking: no - I am ok with that change but we are going bad17:45
sonne|workwiking: it is quite weird already that you were able to attach this processor to your float32_t data17:45
sonne|workthere should be a check in there17:45
wikinglisitsyn: indeed17:45
thoralfsonne: Mapping all text features (unigrams, n-grams) to numbers and storing them in sparse feature files would be a good start.17:45
@lisitsynwiking: have you seen my crazy ideas on API?17:45
sonne|workfoulwall: my suggestion would be to split up the file - I mean the svr and clustering class could be in some other class or so17:46
foulwallOh, I see.17:46
thoralfsonne: Since you probably want to implement all this text-processing-stuff (like stopword removal, stemming) in shogun, this could easily be outsourced to existing tools.17:46
foulwallI'm fixing that.17:46
sonne|workthoralf: but then what would you need shogun for?17:47
sonne|workI mean computing a hash is what you could outsource too17:47
thoralfsonne: MKL, Hashing?17:47
thoralfsonne: Yes and no: When outsourcing the hash, you don't hav influence to the hashing function any more.17:48
thoralfwhile learning.17:48
sonne|workcomputing hash on n-grams is easy to do in C++ at least17:48
sonne|workcomputing hashes on words (with some stop tokens) too17:48
sonne|workstopword removal / stemming is a different story17:49
sonne|workfoulwall: remove the old pull request then and create just one pull request with the split up code17:50
foulwallsonne|work: ok.17:50
sonne|workwiking: look at shogun/features/DenseFeatures.cpp - it uses SG_ADD17:50
sonne|worklisitsyn: we indeed mostly only have preprocessors for double/float17:50
sonne|workwiking: I would suggest to have this p-norm preproc just for double & float17:51
@lisitsynsonne|work: I tend to think we should get rid of templates here17:53
@lisitsynit is not possible to get rid of them at all but they should go deeper17:53
@lisitsynlike in policy-based design (scnr)17:54
-!- gzhd79 [] has quit [Quit: gzhd79]18:05
-!- foulwall [~foulwall@2001:da8:215:c252:193a:a95f:9fad:517a] has quit [Ping timeout: 245 seconds]18:08
wikingsonne|work: for some reason i get this18:15
wikingpreprocessor/PNorm.cpp:114:2: error: use of undeclared identifier 'm_parameters'18:16
wiking        SG_ADD(&m_p, "norm", "P-norm parameter", MS_NOT_AVAILABLE);18:16
wiking        ^18:16
wiking../shogun/base/SGObject.h:65:3: note: expanded from macro 'SG_ADD'18:16
wiking                m_parameters->add(param, name, description);\18:16
wiking                ^18:16
wikingpreprocessor/PNorm.cpp:114:2: error: use of undeclared identifier 'm_model_selection_parameters'18:16
wiking        SG_ADD(&m_p, "norm", "P-norm parameter", MS_NOT_AVAILABLE);18:16
wiking        ^18:16
wiking../shogun/base/SGObject.h:67:4: note: expanded from macro 'SG_ADD'18:16
wiking                        m_model_selection_parameters->add(param, name, description);\18:16
wikingusing SG_ADD18:16
wikingsonne|work: if i don't define template class CPNorm<bool>; etc.18:17
wikingsonne|work: linker starts to cry that it cannot find18:17
wikingthe reference on those18:17
wikingand for some reaseon they are referenced from class_list.cpp.o18:18
wikingany magic tricks to to avoid this error?18:18
sonne|workwiking: yeah you are deriving from a templated class18:18
sonne|workit is a massacre18:19
sonne|workso just create another PNorm preproc for floats18:19
sonne|workand relax18:19
wikingsonne|work: hahhahaha18:19
wikingsonne|work: or have a PNorm preproc now for all?18:20
sonne|workotherwise IIRC you have to specify which type of the base class you are referring to18:20
wikingbecause that's pretty fast18:20
wikingi mean implementation wise18:20
wikingas it's already ready on my machine18:20
wikinganyways i have this18:21
wikingin init18:21
wikingASSERT(features->get_feature_class()==C_DENSE) ASSERT(features->get_feature_type()==F_DREAL)18:21
wikingso it won't run on any other type of features18:21
wikingjust floats18:21
wikingfloat and double yeah18:21
wikingthe only question now is that why's that error for SG_ADD18:22
wikingok it's funny if i specify like this the function: template <class ST> void CPNorm<ST>::register_param ()18:25
wikingi mean template <> void CPNorm<float32_t>::register_param ()18:25
wikingthen it works but of course it'll cry that for the other possible template paramaters this function (register_param) does not exist18:25
wikingbut at least it doesn't cry for not having m_parameters and m_model_selection_parameters variables18:26
wikingi have to leave if you guys have a sane soluation for this problem please write it here i'll read the logs and then commit this fix18:28
sonne|workwiking: yes as I said it needs the actual type18:33
wikingsonne|work: so i should put there by hand all the type? :)18:33
wikingok i see18:33
sonne|workso it will work in the specialization18:33
wikingwill do when i'm online again18:33
sonne|workthat is why I dont' like deriving templated classes from templated classes18:33
-!- gsomix [~Miranda@] has joined #shogun18:34
-!- gzhd79 [] has joined #shogun18:35
-!- sumit [73f91219@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]18:39
gsomixsonne|work: :) how are you?18:40
gsomixsonne|work: hang in there18:43
-!- gzhd79 [] has quit [Quit: gzhd79]18:43
@lisitsynsonne|work: try to not hang yourself!18:49
-!- utsavjain [~tiripS@] has joined #shogun18:53
-!- dvalcarce [] has joined #shogun18:54
dvalcarcehi everyone!18:54
-!- dvalcarce [] has quit [Remote host closed the connection]18:56
-!- dvalcarce [] has joined #shogun18:57
utsavjainHi, I like to participate in GSOC'13 with your organization. I am interested in Machine learning and its tool. What is the first steps, I should follow18:57
dvalcarcesorry for not connecting before, i have taken a lot of subjects this semester18:58
dvalcarcei am working on integrating my demos with django right now18:58
@lisitsynutsavjain: subscribe to the mailing list and check archives of the mailing list - you may find some information there18:59
utsavjainThanks! :)19:00
sonne|workdvalcarce: you are david?19:01
dvalcarcei am daniel valcarce19:01
dvalcarcei think i sent you an email two days ago19:01
sonne|workdaniel yes19:02
sonne|workdvalcarce: you did this cool binary svm with gaussian kernel heatmap demo right?19:02
dvalcarcehaha, yes :)19:02
sonne|workdvalcarce: so just hang in here if you can19:03
dvalcarcesonne|work: ok, no problem19:04
-!- vgorbati [d4029f22@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]19:08
dvalcarcesonne|work: there is a regression demo at <server>/demo/svr/entrance right now, should i add my demos at <server>/demo/binary/entrance and <server>/demo/multi/entrance urls?19:19
sonne|workdvalcarce: demo/classification/{binary,multiclass} ?19:21
sonne|worksounds better IMHO19:22
-!- vgorbati [d4029f22@gateway/web/freenode/ip.] has joined #shogun19:27
-!- HeikoS [] has quit [Quit: Leaving.]19:29
-!- hoijui [] has joined #shogun19:30
@lisitsynsonne|work: just a quick thing - welcome letter should contain (we would be happy to see you, ask us if have questions)19:36
@lisitsynsonne|work: anything else?19:36
sonne|worksome instructions I guess too19:37
@lisitsynsonne|work: what kind of instructions?19:37
sonne|workand any user input19:37
@lisitsynsonne|work: say we don't send invitations, how to find c-base, etc?19:38
sonne|worklisitsyn: I will send you something I received when registering for similar stuff19:38
@lisitsynsonne|work: okay thanks19:38
@lisitsynsonne|work: so you are in france this weekend?19:39
@lisitsynerr early next week19:40
sonne|worktaking vacations :D19:40
@lisitsynsonne|work: we should say we do not provide any accomodation and any invitation letters19:41
sonne|worksurely not but I dont' think it is necessary to say19:41
@lisitsynsonne|work: I am afraid people might think19:42
@lisitsynabout that19:42
sonne|workmaybe they need some info about hotels in berlin?19:42
@lisitsynsonne|work: I need19:42
sonne|workbut I don't think it is so difficult to find one19:42
sonne|workgoogle hotel berlin19:42
sonne|workor hostel berlin19:42
@lisitsynsonne|work: hah19:42
@lisitsynsonne|work: I called an agency today I hope they will find out something for me19:43
sonne|workohh cool sunsetting here19:43
@lisitsynsonne|work:  Rungestrasse 20, 10179 Berlin, ????????19:44
@lisitsynsonne|work: right?19:44
sonne|workanything around train station jannowitzbruecke / ostbahnhof is close enough19:46
sonne|workalright gtg19:46
@lisitsynsee you19:46
wikingweka's smo gives like 10% better accuracy19:48
wikingthan anything with shogun till now19:48
@lisitsynwiking: ehm!19:48
@lisitsynwiking: any guess?19:49
wikingsad news19:49
wikingreally have no fucking clue why19:49
wikingi'm now already trying kernel normalizers et19:49
@lisitsynwiking: same C?19:49
@lisitsynwiking: you must be sure the C, the epsilon, etc are the same you know :)19:51
wikinglisitsyn: c = 1.019:55
wikinge = 1^-1219:55
@lisitsynwiking: what is the kernel?19:55
-!- dvalcarce [] has quit [Quit: dvalcarce]19:58
-!- dvalcarce [] has joined #shogun20:02
wikinglisitsyn: any ideas?20:03
@lisitsynwiking: I'd compare kernel values to be sure20:04
wikingthis is crazy sad20:10
wikingwhat shogun does here :)20:10
@lisitsynwiking: you've got to understand why :)20:13
wikinglisitsyn: 93.053  % vs 0.82015220:18
wikingthis is weka.smo vs shogun.libsvm20:19
wikingand speed wise as well crazy faster in weka20:20
@lisitsynwiking: why do you compare accuracy btw?20:24
wikinglisitsyn: because i'd like to have the same accuracy at least20:25
wikingin shogun20:25
wikingas i hate using weka20:25
@lisitsynwiking: I mean accuracy is not a good measure to compare20:26
wikingshould i print f1?20:26
@lisitsynwiking: no - but I'd have to ask you to find out if there is a bug :)20:26
-!- dvalcarce [] has quit [Quit: dvalcarce]20:27
wikinglisitsyn: where?20:27
@lisitsynwiking: in libsvm probably20:28
wikinglisitsyn: i've tried svmlight20:28
wikingas well20:28
@lisitsynwiking: but I'd check if C has the same meaning20:28
wikingsame shit20:28
@lisitsynwiking: then we either have a bug in apply (not train) or your parameters differ20:28
wikingwelll parameters are now the same20:29
wikingeven the cache size of the kernel20:29
@lisitsynwiking: normalization? standartization?20:31
wikinglisitsyn: normalizing20:31
@lisitsynwiking: did you normalize it in shogun?20:31
wikinghence i've 'fixed' in shogun PNorm20:31
wikingthis is why i had to template it20:32
wikingto be able to normalize the input20:32
@lisitsynwiking: why do you use pnorm?20:32
wikinglisitsyn: what is better for normalizing?20:32
@lisitsynwe have sumone and normone iirc?20:32
@lisitsynwiking: not better but what is P?20:33
wikinglisitsyn: just checking now20:33
wikingprecision is 0.77002220:33
wikingon shogun20:33
@lisitsynwiking: I see a lot of flags here20:34
@lisitsynI'd expect weka does some good stuff in normalizing something20:34
wikingbtw pnorm(2.0) = normone20:35
@lisitsynwiking: yes sure20:38
@lisitsynand pnorm(1.0) is sumone :)20:38
@lisitsynwiking: did you enable bias in libsvm?20:38
-!- dvalcarce [] has joined #shogun20:40
@lisitsynwiking: should be enabled by default though20:40
wikingi did this now20:40
wikingjust in case20:40
wikinglet's see if precision changes20:40
-!- RushabhM [] has joined #shogun20:41
wikinglisitsyn: afaik it was enabled because precision is the exact same20:44
@lisitsynwiking: yes20:44
wikingrandomforest just nails it20:45
wikingthere's nothing in shoung we can get near to these numbers on this data set20:45
wikingthere should be something done about it20:45
@lisitsynwiking: that's why you are here - no? :)20:45
-!- RushabhM [] has quit [Ping timeout: 245 seconds]20:46
-!- RushabhM [] has joined #shogun20:48
@lisitsynwiking: oops :D20:50
-!- RushabhM [] has quit [Read error: Connection reset by peer]20:53
-!- RushabhM [] has joined #shogun20:54
wikingbut really20:54
wikingi will now implement randomforest into shogun20:54
@lisitsynwiking: there is no one that minds it :)20:55
wikinghahaha yeah20:55
-!- RushabhM [] has quit [Read error: Connection reset by peer]20:55
wikingi mind to do it20:56
-!- RushabhM [] has joined #shogun20:56
wikingbut i cannot fucking believe that we cannot do as good as weka20:56
@lisitsynwiking: there is no silver bullet - I guess RF doesn't work in kernel spaces :)20:56
@lisitsynso sonney2k hadn't to use it for his stuff20:57
wikingbut come on20:57
wikingit's polykernel + smo20:58
wikingwe ought to have something that is near to what weka can offer20:58
@lisitsynwiking: I am somewhat sure it is a matter of preprocessing the data20:58
wikinglisitsyn: yeah i thought so as well21:01
-!- iglesiasg [] has joined #shogun21:02
-!- mode/#shogun [+o iglesiasg] by ChanServ21:02
-!- RushabhM [] has quit [Remote host closed the connection]21:04
dvalcarcesonne|work: I've just finish demos integration in django, tomorrow i will check there are no bugs and send a pull request21:13
dvalcarcegood bye everyone :)21:13
@iglesiasgbye bye21:14
-!- dvalcarce [] has left #shogun []21:14
@lisitsyniglesiasg: how is it going?21:14
@iglesiasglisitsyn: hey, all good here, what about you?21:14
@lisitsyniglesiasg: wiking is struggling with our svms now :)21:15
@iglesiasghehe with what exactly?21:15
@iglesiasgin order to get something to work in travis?21:15
@lisitsyniglesiasg: no, weka works better than our svms21:16
@iglesiasglisitsyn: mmm we have space for improvement then21:16
@iglesiasgin what sense to they work better btw?21:16
@lisitsyniglesiasg: the learned model appears to be better21:19
@iglesiasglisitsyn, wiking : have you tested with any particular dataset?21:19
-!- vgorbati_ [d4029f22@gateway/web/freenode/ip.] has joined #shogun21:20
-!- vgorbati [d4029f22@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]21:22
-!- Asix3 [452f9490@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]21:22
wikingiglesiasg: yes21:55
-!- vgorbati_ [d4029f22@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]21:57
@iglesiasgwiking: so what did you try?22:14
@iglesiasgI read something in logs about liblinear22:14
wikingiglesiasg: i'm trying to do binary classification on a quite simple dataset22:20
wikingbut for some reason weka beats by a big margin any shogun classifier until now22:21
@iglesiasgwiking: mmm not good22:23
wikingmmm we don't have a possibility to standardize a dataset right?22:26
@lisitsynwiking: PruneVarSubMean22:27
@sonney2kwiking, what do I hear weka works better?22:37
@sonney2kwhat does that mean?22:37
@iglesiasgthis is a sad day for all of us :(22:38
@sonney2kiglesiasg, it is user error22:39
@sonney2kno idea what wiking does but I am pretty sure that this is not a shogun bug22:39
@sonney2ktraining with poly kernel needs a couple of tweaks to get the data behave well22:40
@sonney2kthe svm I mean22:40
@lisitsynsonney2k: what are the tricks?22:40
@sonney2kthere a lots of factors like kernel normalizer, or features normalization or inhomogeneous poly kernel or homogeneous...22:41
@sonney2keverything can make a big difference22:41
@iglesiasglisitsyn: use weka *lol*22:41
@sonney2kdepending on the data22:41
@sonney2kthat weka trains a lot faster indicates that the data is better normalized22:41
wikingsonney2k: yeah i'm trying with different normalization22:42
wikingkernel and simple norm of the vectors22:42
@sonney2kwiking, why do you use poly kernel?22:42
@sonney2kwhy not gaussian?22:42
wikingsonney2k: well i've tried different22:42
wikinguntil now this is the best22:43
@sonney2kgaussian is much easier to get good results with22:43
@sonney2kwiking, just to understand - which parameters did you use for poly kernel?22:43
wikingsonney2k: for polykernel itself i used the same cache size and degree as in weka22:44
wikingdegree only 122:44
@sonney2kwiking, linear kernel?22:45
@lisitsyndegree =1?22:45
@sonney2kyou mean with linear kernel you get good results?22:45
@sonney2kwhy don't you use liblinear / svmocas then?22:46
@sonney2kwiking, which kernel normalizer?22:46
wikingsonney2k: not even near22:46
wikingsonney2k: sqrtdiag22:46
@lisitsynwiking: what is the difference between linear and poly d=1?22:46
@sonney2kfor linear it doesn't make much sense22:46
@sonney2kto put things on a hypersphere22:47
@sonney2klisitsyn, normalization22:47
@lisitsynsonney2k: ahh22:47
@sonney2kotherwise nothing22:47
@lisitsynbut I wouldn't go for that just because of normalization22:47
@sonney2kwiking, do not compare accuracy - compare with something bias independent22:48
@sonney2klike area under ROC curve22:48
wikingsonney2k: lemme check22:48
@sonney2kbias is very very sensitive22:49
@sonney2k(might be...)22:49
@sonney2kbetter plot both ROC curves even22:49
wikingwhat is the the macro?22:49
* sonney2k leaving train22:49
wikingin contingencytable?22:49
@sonney2kback in a few mins22:49
@lisitsynwiking: ROCEvaluation22:49
@lisitsynuse it22:49
wikingah lol22:50
wikinglol [ERROR] assertion m_current_values.vector && idx<get_num_labels() failed in file labels/Labels.cpp line 5722:57
wikingi wonder what's happening22:58
@lisitsynwiking: what did you do?22:58
wikingjust added CROCEvaluation roc; roc.evaluate(pred, gt);22:59
@lisitsynwiking: debug please :D23:00
@sonney2kwiking, seems like your labels & predictions don't have the same size23:07
wikingi wonder why dont we have such assert in contingencytable :S23:08
@lisitsynwiking: I am not sure it is the reason - did you check call stack?23:08
wikingnot yet23:08
-!- hoijui [] has quit [Ping timeout: 252 seconds]23:11
@sonney2kwiking, so did you check now?23:11
@lisitsynsonney2k: different lights in west/east berlin - really?23:14
@sonney2klisitsyn, you know the east :)23:22
@sonney2kwe can be lucky to even have electricity here23:23
@lisitsynsonney2k: hah why?23:23
-!- vikram360 [~vikram360@] has quit [Ping timeout: 255 seconds]23:41
--- Log closed Fri Apr 19 00:00:28 2013