Open in new window / Try shogun cloud
--- Log opened Mon Jul 01 00:00:12 2013
-!- van51 [] has joined #shogun00:01
shogun-buildbotbuild #1164 of bsd1 - libshogun is complete: Failure [failed test_1]  Build details are at  blamelist: Sergey Lisitsyn <>00:04
shogun-buildbotbuild #1287 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Roman Votyakov <>00:39
shogun-buildbotbuild #1288 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Sergey Lisitsyn <>00:45
-!- van51 [] has quit [Quit: Leaving.]01:25
-!- pickle27 [] has quit [Quit: Leaving]01:57
-!- nube [~rho@] has joined #shogun02:28
-!- shogun-notifier- [] has quit [Quit: transmission timeout]02:38
-!- foulwall|mobile [~foulwall@] has joined #shogun02:57
-!- foulwall|mobile [~foulwall@] has quit [Read error: Connection reset by peer]03:33
-!- foulwall|mobile [~foulwall@] has joined #shogun03:33
-!- foulwall|mobile [~foulwall@] has quit [Read error: Connection reset by peer]03:40
-!- foulwall|mobile [~foulwall@] has joined #shogun03:41
-!- nube [~rho@] has quit [Quit: Leaving.]03:50
-!- foulwall|mobile [~foulwall@] has quit [Ping timeout: 268 seconds]03:53
-!- zxtx [] has quit [Ping timeout: 256 seconds]04:27
shogun-buildbotbuild #444 of nightly_default is complete: Failure [failed test]  Build details are at
-!- foulwall|mobile [~foulwall@] has joined #shogun04:52
-!- fw2k [~foulwall@] has joined #shogun05:37
-!- fw2k [~foulwall@] has quit [Ping timeout: 246 seconds]06:03
-!- foulwall|mobile [~foulwall@] has quit [Ping timeout: 246 seconds]06:03
-!- zxtx [] has joined #shogun06:29
-!- nube [~rho@] has joined #shogun06:35
-!- lisitsyn [] has left #shogun []06:46
-!- votjakovr [] has joined #shogun08:01
-!- sonne|work [] has joined #shogun08:46
-!- nube [~rho@] has quit [Ping timeout: 260 seconds]09:06
-!- iglesiasg [~iglesias@2001:6b0:1:1da0:c55f:fe6c:b69e:1325] has joined #shogun09:28
-!- mode/#shogun [+o iglesiasg] by ChanServ09:29
-!- shogun-notifier- [] has joined #shogun09:33
shogun-notifier-shogun: Thoralf Klein :develop * fad0b0d / src/shogun/machine/OnlineLinearMachine.h:
shogun-notifier-shogun: OnlineLinearMachine::get_w(): Copy weight before passing them to SGVector to prevent double-free on destruction.09:33
shogun-notifier-shogun: Fernando Iglesias :develop * c6e3416 / src/shogun/machine/OnlineLinearMachine.h:
shogun-notifier-shogun: Merge pull request #1196 from tklein23/double_free_hotfix09:33
shogun-notifier-shogun: Fixed double-free when using "SGVector<float32_t> OnlineLinearMachine::get_w()"09:33
-!- nube [~rho@] has joined #shogun09:35
-!- nube1 [~rho@] has joined #shogun09:36
-!- nube [~rho@] has quit [Ping timeout: 276 seconds]09:41
shogun-buildbotbuild #1165 of bsd1 - libshogun is complete: Failure [failed test_1]  Build details are at  blamelist: Thoralf Klein <>09:47
-!- travis-ci [] has joined #shogun09:54
travis-ci[travis-ci] it's Fernando Iglesias's turn to pay the next round of drinks for the massacre he caused in shogun-toolbox/shogun:
-!- travis-ci [] has left #shogun []09:54
-!- zxtx [] has quit [Ping timeout: 248 seconds]09:58
@iglesiasgwiking, ping10:00
-!- nube1 [~rho@] has quit [Ping timeout: 256 seconds]10:06
-!- nube [~rho@] has joined #shogun10:14
shogun-buildbotbuild #1166 of bsd1 - libshogun is complete: Failure [failed test_1]  Build details are at  blamelist: Fernando Iglesias <>10:23
shogun-buildbotbuild #1289 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Thoralf Klein <>10:27
@wikingiglesiasg: pong10:32
@iglesiasgwiking, how can we fix the annoying json serializable error in travis and bsd?10:36
@wikingiglesiasg: travis is quite easy10:36
@iglesiasgwiking, how then? :)10:37
@wikingiglesiasg: remove libjson0-dev from before_install10:37
@wikingand add a new line there after wget10:37
@iglesiasgwiking, an empty line?10:38
@wikingwget --no-check-certificate && unzip json-c-0.11-20130402.zip10:39
@wikingand then add the path to the ./configure10:39
@wikingiglesiasg: although you have to change the ./configure file10:39
@wikingas there's no way to set json path10:39
@wikingor better10:39
@wikingwget --no-check-certificate && unzip && cd <json dir> && sh && ./configure && make && sudo make install10:40
@wikingsomething like this10:41
@wikingand that way ./configure should be able to detect the json-c lib10:41
@wikingmmm no10:41
@iglesiasgwiking, this wget should be in the line that starts with script?10:42
@wikingyou don't need the 'sh' part10:42
@wikingthis should be right after the other wget10:42
@wikingthat handles gmock10:42
@iglesiasgat the end it seems it is not 'quite easy' :P10:42
@wikingwell it's this eays10:42
@wikingyou just need to add this line10:42
@wikingand remove libjson0-dev from the apt-get install line10:42
@wikingonly thing u need to figure out10:43
@iglesiasgadd it after before_install?10:43
@wikingis what's the dir name10:43
@wikingdo you see that line?10:43
@iglesiasgmy bad... I was reading the travis.yml in master. I just cloned the repo in the office computer10:44
@iglesiasgwiking, yeah, I see that line10:44
@wiking- wget --no-check-certificate && unzip && cd json-c-json-c-0.11-20130402 && ./configure && make && sudo make install10:44
@wikingthis is the line you have to add after line 4210:44
@wikingand before line 43 which starts with before_script10:45
@iglesiasgthank you10:45
@wikingand remove libjson0-dev from line 4110:45
@wikingthis way you'll have json 0.1110:45
@wikingand this should fix the travis build problems10:45
@iglesiasgwiking, ok I am going to try right now then, thank you!10:46
@wikingiglesiasg: just push it to develop10:46
@wikingand check the logs10:46
@wikingdid u push?10:47
@iglesiasgit seems there is no make file10:47
shogun-notifier-shogun: Viktor Gal :develop * b90b3e3 / .travis.yml:
shogun-notifier-shogun: Fix json version on travis10:48
@wikinggod bless webedit of github ;)10:48
@wikingiglesiasg: makefile is generated by ./configure10:49
@iglesiasgno space left on device10:49
@iglesiasgno Makefile generated10:49
@wikingwhere? :)10:49
@iglesiasgthe laptop I am working on, nothing to do with travis10:50
@wikingwell what does df -h say? :)10:50
@iglesiasg100% used10:50
@wikingheheh then you need to request for a bigger hdd ;)10:50
@wikingok i've made a mistake10:50
shogun-notifier-shogun: Viktor Gal :develop * 77e3c59 / .travis.yml:
shogun-notifier-shogun: Travis: fix cwd before building10:52
@wikingthis should do it finally10:52
@iglesiasgwiking, do you see some of the builds there gray?10:52
@wikingiglesiasg: yeah10:52
@wiking10:50 <@wiking> ok i've made a mistake10:52
@wikingdid you read this? :)10:52
@wiking10:52 < shogun-notifier-> shogun: Travis: fix cwd before building10:52
@iglesiasgI am used to see them red when a mistake is done rather than gray10:52
@wikingshould fix the grays10:53
@iglesiasgdon't really know what the gray is10:53
@wikingerror :)10:53
@iglesiasgred is error as well10:54
@wikingChecking for libjson (json-c) ... yes10:54
@wikingthis should do it10:54
-!- travis-ci [] has joined #shogun10:55
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 []10:55
@iglesiasguhm ok10:56
@wikingok let see how build 1185 ends10:57
sonne|workwiking: can you install the same thing on the bsd machine?11:01
@wikingsonne|work: noup sorry11:01
@wikingsonne|work: cannot do the same for bsd machine as i intend to keep packages there and not some hand compiled stuff :(11:02
@wikingas it makes things much easier to maintain11:02
sonne|workwiking: ok then let us disable python_modular on that machine11:02
@wikingsonne|work: by all means11:03
@wikingsonne|work: i'll let u know when ports gets libjson updated to 0.1111:04
sonne|workwiking: could   you get the ubu1 buildot back online?11:07
@wikingsonne|work: that one is out of my hands11:07
sonne|workwhat happened?11:07
@wikingsonne|work: it's a box at a uni11:07
sonne|workit is down...11:07
@wikingsonne|work: and none of us are actually at the uni11:07
@wikingand whenever it goes down11:07
@wikingwe have to ask for favours11:07
@wikingto bring it back11:07
sonne|workso now what?11:07
@wikingso it's not instant11:07
shogun-buildbotbuild #1290 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Fernando Iglesias <>11:08
sonne|workcan you ask or is it 'lost'?11:08
@wikingsometimes it takes 2 weeks to get it back11:08
@wikingi've asked already11:08
@wikingso it'll eventually come back11:08
@wikingbut i cannot tell exactly when11:08
sonne|workhmmhh where is foulwall...11:09
@wikingok i think travis is at least fix11:11
@wikingsince python_modular is green now11:11
sonne|workwiking: all error'd11:12
@wikingsonne|work: that's 118411:12
@wikingcheck ^ that one11:12
@wikingthat's 118511:12
@wiking3 are still building the rest is green11:12
@iglesiasgit seems to be fixed now, nice!11:13
sonne|workso current != current11:13
@wikingseems like11:13
shogun-buildbotbuild #1167 of bsd1 - libshogun is complete: Failure [failed test_1]  Build details are at  blamelist: Viktor Gal <>11:13
-!- travis-ci [] has joined #shogun11:18
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 []11:18
@wikingvoila' !11:19
@iglesiasgit's weird to get that message even if all the builds are right11:19
@wikingiglesiasg: it does it because of11:20
@wiking    on_success: change11:20
@wikingfrom .travis.yml11:20
@wikingi.e. when it goes from fail -> success it'll give a notification11:20
@wikingbut from now on until it's in success state it won't give any notification11:21
-!- iglesiasg [~iglesias@2001:6b0:1:1da0:c55f:fe6c:b69e:1325] has quit [Remote host closed the connection]11:25
shogun-buildbotbuild #1168 of bsd1 - libshogun is complete: Failure [failed test_1]  Build details are at  blamelist: Viktor Gal <>11:26
-!- HeikoS [] has joined #shogun11:27
-!- mode/#shogun [+o HeikoS] by ChanServ11:27
sonne|workwiking: current is still not green but the *85 is11:32
sonne|workdoes anyone have a clue about videorecording? I am thinking about renting gear for the workshop11:32
@wikingsonne|work: in best case u need just a decent camera11:33
@wikingbut do you know how's the light setup @ c-base?11:34
@wikingi think i can bring with me an hd camera11:34
@wikingsonne|work: does c-base have a good inet uplink? in that case we could even stream the conference?11:35
@wikingsonne|work: i can offer my server as a proxy11:36
@wikingsonne|work: so we basically stream the ogg to my server with icecast11:36
@wikingand then from there anybody can watch it11:36
@wikingso that the bandwidth is not sooo much of an issue11:36
@wikingand from the dv cam i can create an ogg/theora stream and put that into icecast11:37
sonne|workwiking: it is rather dark usually but I am totally clueless wrt all that. you would have to take care of the streaming solution. I was thinking of ustream or so11:38
@wikingas most of the current cams are only streaming mp4 or similar...11:38
@wikingsonne|work: ustream clinet is quite problematic if we have a custom recorder11:38
@wikingi.e. not an android/iphone phone11:39
@wikingicecast is pretty simple11:39
-!- nube [~rho@] has quit [Ping timeout: 248 seconds]11:39
@wikingand then after the live stream using the original stream we can convert them to webm etc and then put out to the site the html5 videos11:40
@wikingand that's all11:40
@wikingpertty simple thing11:40
@wikingdone that couple of times11:40
@wikingsonne|work: what would be a good thing to do is to mux the beamer and the camera stream into one video11:41
@wikingso that the people watching the videos can see what's on the presentation stream11:41
-!- nube [~rho@] has joined #shogun11:41
@wikingagain that shouldn't be that hard to do11:42
@wikingjust need a video stream from the beamer11:42
@wikingand then render that with the video stream of the camera11:42
sonne|workthe uplink might be okish. but IMHO it would be best if we have some low quality live stream and then HQ recordings (on youtube) later11:42
@wikingsonne|work: we can use our own machine for publishing the videos i think11:43
@wikinghtml5 is supported by all the major browsers11:43
@wikinganyhow we can play with icecast of the quality of the stream11:43
shogun-buildbotbuild #1291 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Viktor Gal <>11:45
-!- nube [~rho@] has quit [Quit: Leaving.]11:48
sonne|workwiking: get some setup working and we will use it! but please test it before :D11:51
@wikingsonne|work: i'm just trying to create a live stream from that camcoder11:55
@wikingsonne|work: we would need a good mic for sure11:57
@wikingsonne|work: possibly an RF mic that you can hook up to somebody's t-shirt and the receiver can be connected to the camera11:57
sonne|workhmmhh no idea where to get that though11:59
@wikingsonne|work: at that site that you sent12:00
@wikingsonne|work: u know these type of mic they use @ universities12:00
-!- lambday [67157f4c@gateway/web/freenode/ip.] has joined #shogun12:04
lambdayHeikoS:  hi12:04
@HeikoSlambday: hi!12:04
lambdayI implemented the get/set diagonal stuffs12:04
lambdayRationalApprroximationIndidivualJob is also implemented and unit-tested12:04
@HeikoSlambday: nice man12:05
@HeikoSso you agreed with the diagonals for matrices?12:05
lambdayas soon as I implement the LogRationalApproximation implementation, I'll be able to unit-test the shifts and all12:05
@HeikoSrather than shifts for operators?12:05
lambdayHeikoS: yes.. although we need to create a SGVector with the diagonal element separately, but that doesn't last long in the memory12:06
@HeikoSlambday: no thats fine and not much12:06
lambday15MB extra12:06
@HeikoSalso keep in mind that for sequential we only have one of them12:06
lambdaythat's affordable12:06
@HeikoSfor parallel, thats not the bottleneck12:06
@HeikoSand for distributed, it doesnt matter12:06
lambdayby the way, I just out of curiosity made the RationalApproximation class non-abstract (temporarily) to check whether shifts/weights/const-multiplier are being calculated correctly12:08
lambdayand it matches the results with krylstat12:08
-!- van51 [] has joined #shogun12:08
lambdayjust that I need a concrete implementation to unit-test that stuff12:08
lambdayoh and yes, using arprec for computing sn, cn, dn will surely increase accuracy12:09
lambdaycn is disturbing using std:: lib12:09
@HeikoSlambday: wow thats awesome12:10
@HeikoSreally getting there! :)12:10
lambdayyeah! getting excited to see this work! but unfortunately my thesis adviser NEEDS to see me everyday in his lab :'(12:11
lambdayso, sitting in his lab, chatting with you :D12:11
lambdaymy thesis is also interesting.. they are trying to build online classifier for malware12:11
@HeikoSlambday: cool12:12
@HeikoSlambday: please be careful about uni stuff :)12:12
@HeikoSso what methods are you using?12:12
lambdayHeikoS: yes trying not to get my ass kicked! I'm really irregular :'(12:12
@HeikoSand classifying what exactly?12:13
lambdayHeikoS: umm, till now they were using logistic regressions12:13
lambdayumm... its a project from quickheal (antivirus)12:13
@HeikoScool stuff12:13
@HeikoSsend me a copy when you are done12:13
lambdaythey regularly update their database with new info...12:13
@HeikoSI see12:13
lambdayHeikoS:  sure :) this will go on for next 1 year12:13
lambdayyes,  so they need some online classifier sort of thing... now they are moving towards transducer based kernels...12:14
lambdaysequence kernels :-/ (I am yet to read)12:14
@HeikoSI dont know that stuff :)12:15
@HeikoSonline learning is interesting though12:15
@HeikoSmaybe some potential for shogun stuff there :) but thats later12:16
lambdayI so wish that I'd use shogun for experiments12:16
lambdaybut my prof sticks to weka and java12:16
@HeikoSI am currently working on MCMC for shogun, and we will use your gsoc project once things work12:16
lambdayHeikoS: awesome!!! :D12:16
@HeikoSlambday: thats fine, shogun is sometimes a bit unstable ;)12:16
shogun-buildbotbuild #1292 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Viktor Gal <>12:17
lambdaybtw why SGMatrix::clone doesn't use memcpy but rather copies elementwise? just wondering12:18
@HeikoSlambday: there might be reasons for that12:18
@HeikoSbut I dont see any now :)12:18
shogun-notifier-shogun: Roman Votyakov :develop * 7103b5c / src/shogun/machine/gp/ (8 files):
shogun-notifier-shogun: refactor likelihood models12:21
shogun-notifier-shogun: Heiko Strathmann :develop * 667f2f3 / src/shogun/machine/gp/ (8 files):
shogun-notifier-shogun: Merge pull request #1200 from votjakovr/feature/gp_refactoring12:21
shogun-notifier-shogun: refactor likelihood models12:21
@HeikoSvotjakovr: thanks for the patch :)12:21
votjakovrHeikoS: hey, yep :)12:22
@HeikoSvotjakovr: hi! :)12:22
@HeikoSvotjakovr: whats the state of the logit stuff? what is missing for it to work?12:22
votjakovrHeikoS: unit tests + that integral evaluation12:23
votjakovrHeikoS: btw i'll also implement probit likelihood12:23
@HeikoSvotjakovr: cool!12:25
@HeikoSvotjakovr: that should be straight, right?12:25
votjakovrHeikoS: yep12:26
-!- lambday [67157f4c@gateway/web/freenode/ip.] has quit []12:27
@HeikoSvotjakovr: cool then, I will include an example on the workshop12:27
votjakovrHeikoS: btw GPstuff toolbox uses Gauss-Kronrod quadrature formula for integral aprroximation12:27
@HeikoSvotjakovr: really? I thought they use this mixture stuff12:28
@HeikoSvotjakovr: I think it doesnt matter too much, maybe choose the one that you can do fastest12:28
@HeikoSbut I think its good to stay in synch with GPML box12:29
votjakovrHeikoS: ok, i'll use mixture12:29
@HeikoSvotjakovr: is there a comparison of methods?12:30
@HeikoSI mean if the Kronrod quadrature is better, and there is code, why not use it12:31
@HeikoSor maybe add both (but later, just choose one for now)12:32
votjakovrHeikoS: i choose approximation with mixture of cdf, because it's simpler to implement it :)12:34
@HeikoSvotjakovr: ok then! :)12:34
votjakovrHeikoS: which example do you whant to include on the workshop?12:35
@HeikoSvotjakovr: just a toy binary classification12:36
@HeikoSbut with learning hyperparameters if possible12:37
shogun-buildbotbuild #1169 of bsd1 - libshogun is complete: Failure [failed test_1]  Build details are at  blamelist: Heiko Strathmann <>12:37
votjakovrHeikoS: Ok :) But will include an example for regression also?12:40
@HeikoSvotjakovr: yes of course, best thing was to have one big example where one can click together points (like the existing svr/svm ones) and then have buttons to learn hyperparameters, plot predictive distributions, and use different inference methods (sparse, exact)12:41
@HeikoSvotjakovr: so if you enjoy coding such a thing, feel free! otherwise I will probably do some hacking on the weekend on this12:42
votjakovrHeikoS: Ok, i'll try to do it :)12:44
@HeikoSvotjakovr: nice :) but please focus on the binary stuff first since thats more important for GSoC for now12:44
votjakovrHeikoS: yeah, i agree12:48
shogun-buildbotbuild #1170 of bsd1 - libshogun is complete: Failure [failed test_1]  Build details are at  blamelist: Roman Votyakov <>12:48
-!- lisitsyn [] has joined #shogun13:02
-!- lambday [67157c4f@gateway/web/freenode/ip.] has joined #shogun13:06
shogun-buildbotbuild #1293 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Heiko Strathmann <>13:18
-!- van51 [] has quit [Quit: Leaving.]13:22
-!- foulwall [~user@2001:da8:215:6100:5063:d7c2:d2e0:7584] has joined #shogun13:24
-!- iglesiasg [] has joined #shogun13:26
-!- mode/#shogun [+o iglesiasg] by ChanServ13:26
sonne|workfoulwall: around?13:30
foulwallHi sonne|work , I would like to turn my working hour to night... sorry for your waiting. This week I have to go for college fieldwork from Monday to Friday, and after that I'll turn my working hour to normal13:33
sonne|workfoulwall: hmmh, please tell such things before you actually do them.13:35
sonne|workfoulwall: when do you have time to discuss? The latest commit state is rather b0rken13:35
foulwallsonne|work: Maybe 11 utc this week? and after this week we can make it back to 9 utc13:36
sonne|worknot sure if I make it 11 utc maybe a little later (lunch break...)13:37
sonne|workfoulwall: currently the clustering example is broken13:38
sonne|workwhen you click clear13:38
shogun-buildbotbuild #1294 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Roman Votyakov <>13:38
sonne|workfoulwall: you still have the old points13:38
foulwallsonne|work: ok, I found that bug today, and is the GaussianProcessRegression class updated? now gp-demo seems not accept the argument.13:40
sonne|workfoulwall: yes you have to fix the example to work with the new interface votjakovr implemented13:41
foulwalland I'll rewrite kernel_matrix visualization today. and implement a simple data importer.13:41
foulwallsonne|work: was there any docs? I don't know how to fix that?13:42
sonne|workfoulwall: well you took it from the examples/ dir and that one got updated so just do the same thing13:43
foulwallsonne|work: gotcha!13:43
sonne|workfoulwall: and each example should have 'interactive data', 'load toy data', 'generate data'13:44
foulwallsonne|work: ok13:45
foulwallhey sonne|work, I found a little bug on shogun's zh-cn manual.
foulwallinterface section, interfaces=python_modular not python-modular14:33
foulwalland python_static not python, etc... this manual looks old.14:38
sonne|workfoulwall: yes it is. I cannot update it - cannot parse kanjis :D14:38
shogun-notifier-shogun: Roman Votyakov :develop * 550de70 / src/shogun/machine/gp/ (3 files):
shogun-notifier-shogun: change default value of the derivative wrt hypers14:58
shogun-notifier-shogun: Heiko Strathmann :develop * f19478d / src/shogun/machine/gp/ (3 files):
shogun-notifier-shogun: Merge pull request #1201 from votjakovr/feature/gp_refactoring14:58
shogun-notifier-shogun: change default value of the derivative wrt hypers14:58
@HeikoSvotjakovr: what are the cases when the derivatives cannot be computed, what happens then?14:58
votjakovrHeikoS: for example we can't comute derivative wrt parameter, which is not used by likelihood model15:01
@HeikoSthen the method should throw an error15:01
@HeikoSnot quietly return an empty vector15:01
@HeikoSvotjakovr: are there other cases (where no error should be thrown?)15:02
votjakovrHeikoS: yep, please look at CLaplacianInferenceMethod::get_marginal_likelihood_derivatives15:03
@HeikoSvotjakovr: I see15:04
@HeikoSif (deriv.num_cols*deriv.num_rows > 0)15:04
@HeikoSthis one you mean right?15:04
@HeikoSmakes sense15:04
votjakovrHeikoS: yep, it's not a perfect solution, but it works15:05
@HeikoSvotjakovr: yeah, something like try-catch would be nicer, but ok lets leave it as it is15:05
shogun-buildbotbuild #1171 of bsd1 - libshogun is complete: Failure [failed test_1]  Build details are at  blamelist: Heiko Strathmann <>15:11
-!- foulwall [~user@2001:da8:215:6100:5063:d7c2:d2e0:7584] has quit [Remote host closed the connection]15:16
-!- FSCV [~FSCV@] has joined #shogun15:17
-!- lisitsyn [] has quit [Quit: Leaving.]15:20
shogun-buildbotbuild #1172 of bsd1 - libshogun is complete: Failure [failed test_1]  Build details are at  blamelist: Roman Votyakov <>15:23
-!- iglesiasg [] has quit [Quit: Leaving]15:35
shogun-buildbotbuild #1295 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Heiko Strathmann <>15:40
-!- lambday [67157c4f@gateway/web/freenode/ip.] has quit [Ping timeout: 250 seconds]15:45
shogun-notifier-shogun: foulwall :develop * 27be102 / examples/undocumented/python_modular/graphical/
shogun-notifier-shogun: fixed interactive gp demo, under votjakovr's new GaussianProcessRegression class16:13
shogun-notifier-shogun: Soeren Sonnenburg :develop * 38a4003 / examples/undocumented/python_modular/graphical/
shogun-notifier-shogun: Merge pull request #1202 from foulwall/develop16:13
shogun-notifier-shogun: fix graphical interactive gp demo, under votjakovr's new GaussianProcessRegression class16:13
shogun-buildbotbuild #1296 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Roman Votyakov <>16:15
shogun-buildbotbuild #1173 of bsd1 - libshogun is complete: Failure [failed test_1]  Build details are at  blamelist: Soeren Sonnenburg <>16:26
-!- van51 [] has joined #shogun16:33
shogun-buildbotbuild #1174 of bsd1 - libshogun is complete: Failure [failed test_1]  Build details are at  blamelist: foulwall <>16:38
-!- foulwall [~user@2001:da8:215:c252:6564:69ee:7ca1:6375] has joined #shogun16:43
votjakovrfoulwall: hi, thanks for the fixes!16:45
foulwallvotjakovr: you're welcome16:46
shogun-buildbotbuild #1297 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Soeren Sonnenburg <>16:54
sonne|workfoulwall: now only the demo needs your fix :)17:01
-!- lisitsyn [] has joined #shogun17:12
foulwallsonne|work: bugs are already fixed, after data_importer I'll send a full pr.17:19
foulwallsonne|work: and I added a menu on gp-demo, make user available to choose their custom kernel function17:21
@wikingsonne|work: ok tomorrow hopefully i'll manage to get a firewire 800-to-400 converter and then i'll be able to test the live streaming... in my little town there's no such thing (i've tried 6 different shops :P)17:25
shogun-buildbotbuild #1298 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: foulwall <>17:29
lisitsynoh guys I didn't resist and bought iphone17:29
-!- pickle27 [] has joined #shogun17:33
@HeikoSlisitsyn: hahahha :)17:43
-!- lambday [67157e4f@gateway/web/freenode/ip.] has joined #shogun17:44
-!- nube [~rho@] has joined #shogun17:44
@HeikoSlisitsyn: I hope you like it at least ;)17:47
@wikinglisitsyn: why not nexus?17:47
lisitsynHeikoS: enjoy it a lot comparing to philips x513 I had before17:48
lisitsynwiking: don't know if it is even available here17:48
-!- lisitsyn [] has quit [Quit: Leaving.]17:54
-!- nube [~rho@] has quit [Ping timeout: 264 seconds]18:03
-!- nube [~rho@] has joined #shogun18:10
pickle27hey everyone18:19
pickle27today is Canada Day!18:19
van51pickle27: so, what aboot it? :P18:34
pickle27van51: just a fun fact!18:34
pickle27van51: its Canada's Birthday so its a national holiday with fireworks etc18:35
van51pickle27: aah sounds nice! happy birthday then :)18:35
-!- foulwall [~user@2001:da8:215:c252:6564:69ee:7ca1:6375] has quit [Ping timeout: 264 seconds]18:36
-!- nube1 [~rho@] has joined #shogun18:40
-!- nube [~rho@] has quit [Read error: Connection reset by peer]18:40
-!- foulwall [~user@2001:da8:215:c252:1c25:3b2:82f0:6f94] has joined #shogun18:41
@wikingsonne|work: ping?18:45
@wikingHeikoS: here?18:48
@HeikoSwiking: hi!18:48
@wikingHeikoS: do u have mplayer at hand?18:48
@HeikoSwiking: nope, why?18:49
@wikingmmm can u play this stream:
@HeikoSwiking: btw whats with this unit test stuff, it feels like ages ago18:49
@HeikoSyep can play18:49
@HeikoSsome books18:49
@wikingok cool18:49
@wikingdid u hear me?18:49
@HeikoSnoting moves though18:49
@HeikoShave no sound here18:50
@HeikoSput your hand in18:50
@wikingu seen my fingers?18:50
@HeikoSimage is stuck18:50
@HeikoStime also doesnt move18:50
@wikingmmm i think there's a lag18:50
@HeikoSah there is a finger18:50
@wikingok cool18:50
@HeikoSbut framerate is about 1 fps18:51
@HeikoSor even less18:51
@HeikoSno the finger is gone18:51
@wikingthe problem is that i'm generating the input with a raspberry pi18:52
@wikingso the rendering time is not optimal ;)18:53
@wikinganyhow the thing works18:53
-!- zxtx [] has joined #shogun18:53
@wikingHeikoS: u dont have audio on your machine?18:53
@wikingor you dont get the audio strema?18:53
@wikingsonney2k: ok the video/audio streaming works.18:56
@wikingsonney2k: we just need a good (strong with cpu time) linux machine dedicated for rendering the video...18:57
-!- van51 [] has quit [Quit: Leaving.]18:57
@HeikoSwiking: sorry no not really currently19:01
@wikingHeikoS: ok no worries19:01
@wikingHeikoS: i see with mplayer that the audio is there as well19:01
@wikingand i've heard it as well19:01
@wikingthe only thing is that we need a dedicated machine for this19:01
@wikingand i still need to figure out how to mux it in a way19:02
@wikingah i think i know19:02
@wikingtr will do it for me19:02
lambdayHeikoS: hi19:03
@HeikoSlambday: hi!19:03
lambdayshifts, weights, etc tested.. works fine *but*19:03
@wikingmmm what is the mux command in linux?19:04
lambdayHeikoS: I tried to add a log-det test as well.. something is wrong.. its coming out to be near the actual log-det value.. not even close19:04
lambdaytill shifts I tested..19:04
@HeikoSlambday: what did you try exactly?19:05
lambdaysimilar like the one before.. estimate log-det using rational approximation this time19:05
lambdayHeikoS: well, could you please tell me if this is what we are supposed to do..19:06
lambdayQ is the matrix19:06
lambdayso, I move shifted ops like this --> (zI-Q)19:06
@HeikoSlambday: try the matrix log itself first maybe19:06
@HeikoSso multiply with a fixed vector rather than random19:07
lambdayHeikoS: alright...19:07
@HeikoSbut that wont solve it I guess19:07
lambdayso, I solve system with (zI-Q)x=s, s=sample19:07
@HeikoSso you are solving system individually currently?19:07
@HeikoSso check the solutions for all these systems19:08
@HeikoSand before check whether the shifts are correct ( I think you did that already right?)19:08
lambdayyes shifts are correct as per krylstat19:08
@HeikoSlambday: okay19:08
-!- lisitsyn [~lisitsyn@] has joined #shogun19:08
@HeikoSthen it could be the systems19:08
@HeikoSmaybe wrong solution19:08
lambdaybut can't check whether individual solutions are correct19:08
@HeikoSso check them19:08
@HeikoSwhy not?19:09
lambdaykrylstat uses cocg_m19:09
@HeikoScompare against matlab/octave19:09
@HeikoSand if those are correct, there is probably something wrong with the formulae19:09
lambdayI think something is wrong with the formula I am using here..19:10
lambdayafter I solve the systems, I multiplied weight w*x, then sum it all up over all the shifts19:11
lambdaythen took the imag, applied operator to that imag vector once again, then multiplied with the constant multiplier...19:11
lambdayI mean, pretty much what krylstat does19:11
@HeikoSlambday: dont worry19:11
@HeikoSjust check step by step19:11
@HeikoSmaybe a little toy example in octave reveals things19:12
@HeikoSand if its also wrong in octave, you know its the math19:12
-!- shogun-notifier- [] has quit [Quit: transmission timeout]19:13
lambdayalright, I am checking19:13
@HeikoSlambday: also check the individual poarts of the imag computation19:15
@HeikoSmaybe we missed something there in the associative rules19:16
-!- shogun-notifier- [] has joined #shogun19:20
shogun-notifier-shogun-demo: foulwall :master * d5fc7b8 / / (11 files):
shogun-notifier-shogun-demo: 1. fix bug in clustering. 2.adapt gp demo to new19:20
shogun-notifier-shogun-demo: GaussianProcessRegression class. 3. user now can choose kernel_function19:20
shogun-notifier-shogun-demo: in gp demo. 4. rewrite the kernel_matrix visualization demo in new19:20
shogun-notifier-shogun-demo: framework, but got a defective result. I'll check if there's any better19:20
shogun-notifier-shogun-demo: way than conrec.js for heatmap drawing, or else I'll fix the defect.19:20
shogun-notifier-shogun-demo: Soeren Sonnenburg :master * 2a074ed / / (11 files):
shogun-notifier-shogun-demo: Merge pull request #11 from foulwall/perceptron19:20
shogun-notifier-shogun-demo: 1. fix bug in clustering. 2.adapt gp demo to new19:20
lambdayHeikoS: okay I am checking.. but as per Hale paper, it should be alright :(19:20
lambdayHeikoS: I'm checking log(Q) first... may be that will reveal the issue19:21
@HeikoSlambday: yep good idea!19:21
-!- foulwall [~user@2001:da8:215:c252:1c25:3b2:82f0:6f94] has quit [Ping timeout: 264 seconds]19:21
-!- foulwall [~user@2001:da8:215:c252:b8f1:eb8a:e017:c0a1] has joined #shogun19:32
@sonney2kfoulwall, ok clustering works again!19:35
@wikingsonney2k: ok so this camera is 720x576, ie dvd quality19:36
@sonney2kfoulwall, GPs do mostly (but creating a new train data set should clear)19:36
@wikingsonney2k: if we want better quality we need to rent19:36
@sonney2kwiking, IMHO good enough19:36
@wikingsonney2k: other than that the icecast worked nicely19:36
@wikingsonney2k: i'm just investigating how we can render the slides as well into the video stream nicely19:36
@wikingsonney2k: i'll prepare the scripts19:36
@sonney2kwiking, can we store to the recordings too for putting them on youtube later?19:37
@wikingsonney2k: but can u bring a good laptop/pc which can handle the rendering19:37
-!- van51 [] has joined #shogun19:37
@sonney2kwhat is needed? I only have a core2duo19:37
@wikingsonney2k: hehehe well the streaming is totally command line: dvgrab -format raw -V -input /dev/video0 -| ffmpeg2theora -a 0 -v 5 -f dv -x 320 -y 240 -o /dev/stdout -|oggfwd -p -d "Shogun test stream" -n "Shogun workshop" 8000 ...19:37
@sonney2kvan51, evening gentlement19:37
foulwallsonney2k: ok, I'll fix them tomorrow!. too late now, cu19:38
@wikingsonney2k: so i can insert there a | tee dvstream.dv |19:38
@sonney2kwiking, from usb? then maybe someone with sth faster would work!19:38
@sonney2kfoulwall, don't forget your weekly report!19:38
@wikingsonney2k: i have firewire output as well19:38
van51sonney2k: hello19:38
@sonney2kwiking, a fine19:38
@wikingsonney2k: just make sure that machine has a firewire input19:38
@wikingsonney2k: but since it's a firewire 40019:38
@wikingsonney2k: it's as good as usb 2.0 high speed ;P19:38
@sonney2kwiking, I guess we could later do the conversion19:38
foulwallok, I'll write on my cellphone19:39
@wikingsonney2k: well yeah we shoudl do this small resolution streaming on the fly19:39
@wikingsonney2k: with a core2duo it should be fine19:39
@wikingsonney2k: i was testing with a raspberry pi :P19:39
@wikingsonney2k:  i mean the grabbing + encoding + streaming19:39
@wikingsonney2k: and it was ok :P19:39
@wikingbut obviously we need something more decent19:40
@wikinga c2d is going to be just fine19:40
@wikingimho we should save the dv raw input19:40
@wikingso we can do some postprocessing and encode the video in higher resolution19:40
@wikingand put it up later...19:41
@wikingsonney2k: only thing we need is a good microphone19:41
@sonney2kwiking, don't you have a 'faster' notebook19:41
@sonney2kyeah that remains a problem19:41
@wikingsonney2k: i do... but it's currently osx19:42
@wikingand dvgrab and shit won't work19:42
@sonney2kwiking, usb mic?19:42
@wikingsonney2k: naaah19:42
@sonney2kwiking, what then?19:42
@wikingsonney2k: i need something that i can plug in to the camcoder19:42
@wikingwhich is an analog input19:42
@sonney2kvan51, so what is the status?19:42
@wikingyou know the usual headphone jack19:42
@sonney2kwiking, yeah but then it wouldjust work right?19:42
@sonney2kI mean how is this osx?19:43
van51sonney2k: well, I think it's ok now in general19:43
@wikingsonney2k: yeah because then the camcoder reads the audio from that input instead of the built in mic19:43
@wikingsonney2k: in osx i have the problem with libusb/libfirewire missing ....19:43
@wikingsonney2k: and dvgrab is totally dependent on that19:43
@wikingsonney2k: and i don't have an equivalent utility on osx like dvgrab19:44
van51sonney2k: I' m just thinking now that it may be redundant to handle a CStringFeatures object inside and that it could be switched to a plain SGStringList19:44
@wikingthat simply reads the dv stream from usb/firewire19:44
@wikingsonney2k: but in worst case if we cannot find a spare machine19:44
@wikingwhat i can do is simply19:44
@wikingboot from a usb stick a linux19:44
@wikingand that should do it19:44
@wikingi'll bring an external storage19:45
@wikingto save there the raw dv stream19:45
@wikingi'll still have to figure out how to grab the video stream from the beamer19:45
@wikingand mux it into the video stream19:45
@wikingbut i'll solve this in the next couple of days19:45
@wikingsonney2k: i hope the machine that is hooked up to the beamer is some linux distro ;P19:46
@wikingsonney2k: can we do a pilot test on 11th ?19:46
@wikingsonney2k: with the beamer etc....?19:46
@sonney2kwiking, in the morning of the 12th only19:46
@wikingsonney2k: ok19:46
@wikingsonney2k: then i'll come early and that's it19:46
@wikingshoudln't take more than 20 mins max19:46
@wikingsonney2k: can u take care of the mic?19:47
@sonney2kwiking, but you could probably ask the c-base guys on the 11th if you just go there for a drink with sergey lisitsyn19:47
@wikingsonney2k: rent it from somewhere in berlin?19:47
@wikingi mean the microphone19:47
@sonney2kwiking, no idea I will try but no promises :/19:47
@wikingsonney2k: or at least can u give me some urls of multimedia rent places in berlin19:48
@wikingand i'll pintpoint u the mic we would need19:48
@wikingsonney2k: if we can setup the live stream nicely19:48
@sonney2kvan51, I only ``know'' the one I posted19:48
@sonney2kwiking, ^19:48
@sonney2kvan51, it is ok with using StringFeatures19:49
@wikingsonney2k: we should be advertise that people who are watching the stream can ask questions here on the irc channel19:49
@sonney2kwiking, during discussion19:49
@sonney2kmakes sense :D19:49
@wikingsonney2k: i:P19:49
van51sonney2k: It's just a bit messy when you initialize it, but that's all19:49
van51sonney2k: I beliebe it's good to be merged then :)19:49
@sonney2kvan51, ok then I will have another look19:50
@wikingsonney2k: what was the link?19:50
@wikingok found it19:51
@wikingsonney2k: c-base should have some basic a/v equipment no?19:51
@sonney2kyeah they have everything. which doesn't mean they lend everything to us19:52
@wikingsonney2k: can u ask them if they have an rf mic?19:52
@wikingsonney2k: that's the only thing we would need19:52
@wikingand inet uplink :P19:53
@wikingsonney2k: something similar19:54
@wikingsonney2k: as u can see there's one part with the jack that would go for the camcoder and the other part is the mic and the rf transmitter19:54
@wikingnothing fancy19:54
@wikingsonney2k: i'll even bring a tripod so they dont even have to lend us that :P19:55
@wikingyeah that's the stuff19:57
@sonney2kwiking, alright then let me buy it20:02
@sonney2kvan51, could you do some benchmarks then?20:04
@sonney2kvan51, do you have a decent machine?20:04
van51sonney2k: yeah I think so20:04
@sonney2kvan51, a long time ago I organized this challenge
@sonney2kso why not take this data set20:05
@sonney2kvan51, data is here
@sonney2kyou need train data and labels20:06
@sonney2kthat is the description20:07
@sonney2kvan51, take say a subset of a few thousand examples first and try to predict on some held out set of say 50k examples20:08
@sonney2kvan51, ask if you need more details :D20:08
@sonney2kwiking, ok ordered20:08
van51sonney2k: ok!20:09
@sonney2kvan51, I still see some issues20:10
van51sonney2k: ok, tell me20:11
@sonney2kvan51, one is that dim is not used - maybe we drop dim altogether and just use num_bits20:11
@sonney2kthen get_nnz_features_for_vector should return the string length not dim20:11
van51sonney2k: dim is used in the comparison of vector lengths20:11
van51and I think it's more intuitive to specify the dimension than the number of bits20:11
@sonney2kvan51, yeah but not when reducing the hash20:11
@sonney2kit should be hash % dim then20:12
@sonney2kbut it is a speed penalty20:12
@sonney2knot much but it is20:12
van51sonney2k: also for the number of nnz features isn't dim the smallest upper bound?20:13
@sonney2kvan51, you should always assume that string lenght << dim20:13
@wikingsonney2k: lol ok cool20:13
@wikingsonney2k: the server has 100mbit/sec uplink from where we'd stream20:13
@wikingsonney2k: so it can handle quite some connections20:14
@wikingsonney2k: ok then the only thing is that lets have enough spare batteries for that device....20:14
van51sonney2k: why is that? a document can have an arbitrary number of tokens, but you hash it in a specific dimension size, namely dim20:14
-!- HeikoS [] has quit [Quit: Leaving.]20:15
@sonney2kvan51, yes but it is usually sparse!20:16
van51sonney2k: ok, I get what you 're saying20:17
@sonney2kvan51, so if you string is say 10k long and you dim is 256**8 (for 8-grams) you would get still get about only 10k as nnz features == upper bound of number of n-grams in the string20:17
@sonney2kvan51, but IMHO you can just drop dim20:18
@sonney2kit is totally sufficient to do this bit based20:18
van51sonney2k: ok, but I would like to leave in in the constructor20:18
van51sonney2k: so that it can be specified in this way as well20:18
@sonney2kvan51, but then add a check that it is a dim=2**n20:25
van51sonney2k: I just explain it in the documentation after all20:27
-!- nube [~rho@] has joined #shogun20:27
van51sonney2k: I'll test it and repush20:28
-!- nube1 [~rho@] has quit [Ping timeout: 276 seconds]20:29
@sonney2kvan51, hmmhh maybe nbits is better and less misleading than a parameter that can be used wrongly20:30
@sonney2kvan51, let me check again20:30
van51sonney2k: yeah I meant I changed it to num_bits and explain in the documentation what it means for the dimensionality of the feature space20:30
@sonney2kvan51, well you could do return hash % dim;20:31
@sonney2kI don't know about nowadays processors maybe it even is as fast as hash & mask;20:31
van51sonney2k: I have no idea :p20:32
@sonney2kvan51, well benchmark it :D20:32
van51sonney2k: hehe true!20:33
@sonney2kjust use the webspam data set for that20:33
van51sonney2k: ok. then when I update it again you can merge and if the benchmark suggests something else, we update it?20:34
@sonney2kvan51, yes20:35
@sonney2kwiking, excellent20:36
@sonney2kwiking, problem is only that we might have people showing demos20:36
@sonney2kso it is not clear if they run on each and every machine20:36
@sonney2kwiking, it looks perfect though20:37
@wikingsonney2k: yeah that's why is this shit there: TwinPact100 vga capture20:37
-!- van511 [] has joined #shogun20:37
@sonney2kwiking, wait you have one that does vga capture too?20:37
-!- van51 [] has quit [Quit: Leaving.]20:37
@wikingsonney2k: noup :(20:38
@sonney2kwiking, too bad... that would have been excellent20:38
@wikingsonney2k: i'll try to find it... maybe one of my friends has something similar20:38
@wikingdoh it's pretty expensive shit20:38
-!- nube [~rho@] has quit [Ping timeout: 256 seconds]20:39
@sonney2kno idea what the c-base beamer has20:40
@wikingsonney2k: i'll try to look around during this week20:42
@wikingsonney2k: since this way of course it doesn't matter who and what beams to the beamer :P20:42
@sonney2kvan511, ok so please drop the num_bits stuff for now and do instead of return hash & ((1 << num_bits) - 1) return hash % dim; then we will see20:44
van511sonney2k: baah I just finished doing it the other way :p20:44
@sonney2kvan511, well also fine!20:44
@sonney2kvan511, when you do benchmarking jsut make sure you don't use some const dim=128 when you do hash % dim later20:45
@sonney2kvan511, the compiler would do a hash & (mask) thing then20:45
van511sonney2k: you mean that a const in a mod operation is transformed to a mask?20:46
@sonney2kvan511, alright looking forward to the benchmark - please do it systematically as in use some script to run it on a few strings up to 'lots' and then store timings20:46
@sonney2kvan511, exactly - compilers are clever these days20:47
@sonney2kvan511, for a few bit sizes like 8,16,24,32 bit20:48
@sonney2kvan511, and remember to compile with optimizations enabled!20:48
van511sonney2k: ok, got it!20:48
@sonney2kvan511, we could then even use these tables in a blog post / create a figure out of it comparing % vs mask20:49
@wikingsonney2k: can u ask the exact type of the beamer they have @ c-base plz?20:49
@sonney2kvan511, and later you can just train liblinear with a) converted features b) with dotfeatures on that data set20:49
@sonney2kand see how it goes20:49
@wikingsonney2k: maybe i can just bring a simple vga2usb adapter20:49
@sonney2kwiking, I have no idea20:49
van511sonney2k: the benchmark will be with the dotfeatures though20:50
@wikingsonney2k: well let's just ask the type of beamer they have and i'll try to figure out something20:50
van511sonney2k: that last you said is just for comparison between the two20:50
-!- iglesiasg [] has joined #shogun20:54
-!- mode/#shogun [+o iglesiasg] by ChanServ20:54
@wikingsonney2k: or we get another camera that is zoomed and directed on the projectors screen ;P20:56
-!- nube [~rho@] has joined #shogun21:00
lisitsynvotjakovr: weekly report please ;)21:01
lisitsynfoulwall: weekly report please21:01
lisitsynhushell: weekly report please21:01
@wikinglisitsyn: get us a cheap russian solution for vga/dvi to usb :)21:01
lisitsynwiking: vga to usb?21:01
@wikinglisitsyn: yep21:02
@sonney2kwiking, yeah I asked. I asked a bit too much maybe hope we get a response :)21:02
@wikingsonney2k: cool21:02
lisitsynwiking: is that possible?21:02
@wikinglisitsyn: yeas for 300 USD  ;)21:02
@wikingit's basically a video4linux device21:03
@wikingit's just a dv stream...21:04
@sonney2kvan511, well sure makes sense. I suspect you would win the competition if you do this with shogun then ;)21:05
-!- van511 is now known as van5121:05
@sonney2klisitsyn, anything from gsomix?21:06
van51sonney2k: that would be cool :D21:06
lisitsynsonney2k: he is out of touch for me :)21:07
van51sonney2k: btw I have pushed and I am waiting for travis21:07
@sonney2kvan51, I am serious here yahoo & chap are the ones that won :)21:08
van51sonney2k: yea I saw the results21:08
van51sonney2k: wasn't chap at yahoo at the time? :P21:08
@sonney2kvan51, and we have liblinear sgdqn etc in shogun but the feature representation helps a *lot* wrt performance21:08
van51what's wrt?21:09
@sonney2kand that was just sparse doubles IIRC21:09
@sonney2kvan51, with respect oo21:10
* sonney2k thunderstorm!21:10
@sonney2kvan51, and dont' forget your weekly report!21:11
van51sonney2k: it's sent!21:11
@sonney2kvan51, one more thing with our PR - you always have to use the *SAME* seed21:13
van51sonney2k: I got that from Olivier's paper21:13
@sonney2kvan51, otherwise the order of n-grams will matter21:13
van51sonney2k: they were saying that for each feature you use a different hash function21:14
@sonney2khash = calculate_token_hash(&sv.vector[start], end-start, num_bits, hash);21:14
van51sonney2k: and that you can achieve that by having the seed set to the previous hash value21:14
@sonney2kyeah but it depends on the previous feature here21:14
-!- nube [~rho@] has quit [Ping timeout: 248 seconds]21:15
@sonney2kwhen you have 'aaa' 'bbb' it might make a big difference if you compare it to 'bbb' 'aaa' then21:16
van51sonney2k: no21:16
van51sonney2k: sorry wait21:16
van51sonney2k: I;m reading it again21:16
@sonney2kit would make sense in my eyes when you have say certain feature classes - say several texts describing one image or so21:17
@sonney2kthen each class could have its own seed21:17
van51sonney2k: you are right21:17
van51sonney2k: it just says to use a different seed for each feature21:18
van51like the index of the feature21:18
@sonney2kvan51, it can make sense btw21:18
@sonney2kvan51, on the DNA data set it would probably21:19
van51sonney2k: yeah but this is for general purposes21:19
@sonney2kvan51, there is a tradeoff sometimes order of strings matters a lot sometimes not. for now just use some constant like 0xdeadbeaf21:20
@sonney2kas seed21:21
van51sonney2k: ok!21:21
van51sonney2k: good catch btw21:21
van51sonney2k: I will also change it in the converter now then21:25
van51so that the tests won't fail21:26
-!- nube [~rho@] has joined #shogun21:27
* sonney2k Zzzzz21:31
lisitsynokay sonney2k falls asleep I wake up21:32
-!- nube [~rho@] has quit [Ping timeout: 276 seconds]21:34
-!- lisitsyn [~lisitsyn@] has quit [Quit: Leaving.]21:36
-!- nube [~rho@] has joined #shogun21:47
-!- lisitsyn [~lisitsyn@] has joined #shogun21:54
-!- lisitsyn [~lisitsyn@] has quit [Quit: Leaving.]21:55
-!- lisitsyn [~lisitsyn@] has joined #shogun21:57
-!- nube [~rho@] has quit [Ping timeout: 264 seconds]21:58
pickle27lisitsyn: hey!22:06
lisitsynpickle27: hey22:07
pickle27I was thinking a bit about the shogun interface for my GSoC project22:07
pickle27and I was looking over some older emails with Andreas22:07
pickle27what I'm thinking now is that the AJD algorithms should be preproccessors22:08
pickle27and then the BSS algorithms (which use the AJD algs) should be converters22:08
lisitsynpickle27: ahh so like init / transform?22:08
pickle27sort of I am still looking through the pre processors22:09
pickle27ADJ is kind of like the best PCA for several matrices22:09
pickle27so the result would be applicable to new matrices that are expected to be part of the same group22:10
pickle27however unlike PCA AJD takes a SGNDArray instead of a single Matrix22:10
pickle27so we could make an object very similar to PCA22:11
pickle27anyways I messaged Andreas back to see if he could elaborate on the use case for that22:11
lisitsynwhat is n-dim input here?22:12
pickle27a series of matrics22:12
pickle27could also be just an array/list of matrices22:12
-!- nube [~rho@] has joined #shogun22:13
pickle27I need to move out of the matrix realm and into the feature realm soon. Hopefully Andreas will have an example use case22:15
lisitsynpickle27: could you please give me some understanding of these series?22:15
pickle27although internally it will look similar to the pca object22:15
lisitsynwhat is one matrix here?22:15
pickle27lisitsyn: I'll try but I only really understand what they are at the numerical level22:15
pickle27Its basically a set of matrices and the goal is to find the single matrix V that best diagonalizes all the matrices in the set22:16
pickle27where PCA diagonlizes only a single matrix AJD diagonalizes a set22:16
lisitsynpickle27: where these matrices come from?22:17
pickle27lisitsyn: right now they are just synthetic data22:17
pickle27I was hoping Andreas could provide a machine learning example22:17
lisitsynpickle27: I mean we all get used to feature vectors22:18
lisitsynbut it is like feature matrix?22:18
lisitsynsee what I mean?22:19
pickle27when you're using AJD for BSS its a series of covariance matrices22:19
pickle27Im not sure what it would be in the pre processing application22:20
-!- shogun-notifier- [] has quit [Quit: transmission timeout]22:20
lisitsynpickle27: what is the data these cov. matrices are computed from?22:20
pickle27the mixed input signal you are trying to separate22:20
pickle27mixed input signals*22:21
pickle27lisitsyn: I'm looking at the PCA preprocessor now and it works on a cov matrix too22:24
lisitsynpickle27: yes indeed22:24
pickle27lisitsyn: I think that is probably what you would apply AJD too as well22:24
lisitsynbut *one* covariance matrix22:24
lisitsynwhere multiple come from?22:25
pickle27yeah I've got to look into that22:25
pickle27Andreas did mention using AJD for pre processing though22:25
pickle27so he has some idea22:25
-!- iglesiasg [] has quit [Quit: Leaving]22:28
lisitsynpickle27: alright lets see what he answers then22:29
lisitsynI don't get the concept yet22:29
pickle27yeah Im not quite seeing the connection to applications other than BSS22:30
-!- nube [~rho@] has quit [Ping timeout: 256 seconds]22:30
-!- lisitsyn [~lisitsyn@] has quit [Quit: Leaving.]22:42
-!- lisitsyn1 [~lisitsyn@] has joined #shogun22:42
-!- lambday [67157e4f@gateway/web/freenode/ip.] has quit []22:42
-!- nube [~rho@] has joined #shogun22:45
-!- zxtx [] has quit [Ping timeout: 256 seconds]22:52
-!- pickle27 [] has quit [Quit: Leaving]22:54
-!- zxtx [] has joined #shogun22:55
-!- zxtx [] has quit [Ping timeout: 248 seconds]23:00
-!- lisitsyn1 [~lisitsyn@] has left #shogun []23:00
-!- votjakovr [] has left #shogun ["Fallen asleep!"]23:02
-!- gsomix [~gsomix@] has joined #shogun23:11
-!- zxtx [] has joined #shogun23:31
-!- gsomix [~gsomix@] has quit [Ping timeout: 256 seconds]23:51
-!- gsomix [~gsomix@] has joined #shogun23:55
--- Log closed Tue Jul 02 00:00:06 2013