Open in new window / Try shogun cloud
--- Log opened Fri Aug 24 00:00:17 2012
-!- zxtx [~zv@cpe-75-83-151-252.socal.res.rr.com] has quit [Read error: Operation timed out]03:48
shogun-buildbot_build #75 of nightly_default is complete: Success [build successful]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_default/builds/7503:54
-!- av3ngr [av3ngr@nat/redhat/x-unomdjhlngrsqlva] has joined #shogun06:45
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun06:52
-!- zxtx [~zv@guest-wireless-nup-198-147-120-231.usc.edu] has joined #shogun07:16
-!- av3ngr [av3ngr@nat/redhat/x-unomdjhlngrsqlva] has quit [Ping timeout: 246 seconds]08:12
-!- av3ngr [av3ngr@nat/redhat/session] has joined #shogun08:24
-!- av3ngr [av3ngr@nat/redhat/session] has quit [Ping timeout: 248 seconds]08:55
-!- uricamic [~uricamic@2001:718:2:1634:e9e5:5ab4:359d:813d] has joined #shogun08:56
-!- av3ngr [av3ngr@nat/redhat/x-canvtavaulipugsg] has joined #shogun09:14
-!- zxtx [~zv@guest-wireless-nup-198-147-120-231.usc.edu] has quit [Ping timeout: 264 seconds]09:16
-!- av3ngr [av3ngr@nat/redhat/x-canvtavaulipugsg] has left #shogun ["Leaving"]09:35
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving]10:09
-!- zxtx [~zv@cpe-75-83-151-252.socal.res.rr.com] has joined #shogun10:33
-!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has joined #shogun10:37
-!- emrecelikten [~emre@trir-4d0d9a53.pool.mediaWays.net] has joined #shogun10:59
-!- yoo_ [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has joined #shogun11:17
-!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has quit [Quit: Page closed]11:29
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.83.179.44.135] has joined #shogun13:15
-!- yoo_ [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has quit [Quit: Page closed]13:19
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.83.179.44.135] has quit [Quit: Page closed]13:52
-!- nicococo [~nico@guest0.ml.tu-berlin.de] has joined #shogun13:57
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.83.179.44.135] has joined #shogun14:00
-!- blackburn [~blackburn@109.226.79.69] has joined #shogun14:21
blackburnHELLO14:24
-!- yoh_ [~yoh@washoe.dartmouth.edu] has joined #shogun14:37
blackburnn4nd0: around?14:40
n4nd0blackburn: yeah14:41
n4nd0blackburn: at skype with Nico :)14:41
blackburnoh any news on SO improvements?14:41
blackburndoes he have any cool ideas?14:41
-!- Netsplit *.net <-> *.split quits: yoh14:42
n4nd0blackburn: I have been asking here about the bundle methods14:46
n4nd0blackburn: I have the ideas more clear now :)14:46
blackburnwhat about subtracting <w,Psi(x,y_truth)> in argmax?14:47
n4nd0yeah we talked about that14:47
n4nd0he said that the score in the ResultSet (the return value of argmax) can be defined as we like14:48
n4nd0so yes, we can make that the score is14:48
n4nd0score = <w,Psi(xi,y)> + Delta(yi, y) - <w,Psi(xi,yi)>14:48
n4nd0that way in the generic risk we'll just do14:49
n4nd0call to argmax14:49
n4nd0and14:49
n4nd0risk += ret->score14:49
n4nd0blackburn: do you like it like that?14:49
blackburnn4nd0: yes that's ok14:49
n4nd0cool14:50
blackburnn4nd0: but we need to not forget about subtracting loss from multiclass model14:50
n4nd0blackburn: yeah, sure14:50
blackburnhowever this is kind of inefficient it is more general this way14:50
blackburnn4nd0: what about HM SVM?14:50
blackburnis it correct now with provided risk function?14:51
n4nd0yes14:51
n4nd0the only thing is that now (in the PR I submitted) the loss is taken into account twice14:51
n4nd0but I'll fix that within minutes now14:51
n4nd0nicococo: anything to add? :)14:53
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.83.179.44.135] has quit [Quit: Page closed]14:54
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun14:54
n4nd0blackburn: what do you think?14:55
-!- nicococo [~nico@guest0.ml.tu-berlin.de] has quit [Ping timeout: 265 seconds]14:55
n4nd0does it sound good?14:55
blackburnn4nd0: about what?14:56
blackburn:)14:56
n4nd0about having this working well finally14:57
n4nd0supported by Nico, it gives more confidence that it is ok14:57
blackburnwell yeah I don't mind to make it work :)14:58
blackburnn4nd0: can you please also document in argmax description14:59
blackburnthat in training mode it computes blablah14:59
blackburnwith blahblah assumption14:59
blackburn:D14:59
n4nd0ok14:59
blackburnI will add director SO model15:00
blackburnonce we have everything working15:00
blackburnthis would be really kind of killer feature I hope15:01
n4nd0blackburn: yeah the SO director is very interesting15:01
n4nd0I told gsomix about it15:01
n4nd0blackburn: SO model and StateModel is very interesting too!15:02
blackburnwell it is pretty easy to set up, not really needed any gsomix's expertise :)15:02
blackburndirector StateModel?15:02
n4nd0I would welcome StateModel very very much15:02
n4nd0yeah15:02
blackburnit is HM right?15:02
n4nd0you know when you use the HM-SVM you need another piece15:02
blackburnso i.e. if you have 3 states15:03
blackburnyou would need that15:03
n4nd0yes15:03
blackburnright?15:03
n4nd0apart from the number of states, the type of model too15:03
blackburnokay lets implement that too15:03
blackburnah llike semi markovian?15:03
n4nd0e.g. in speech recognition people used right Markov models15:03
n4nd0the transitions to the left (think of a Markov chain of states) are not allowed15:04
blackburnI think about starting PGM course *seriously* this time15:04
n4nd0yeah me too15:04
n4nd0end of September15:04
n4nd0:D15:04
blackburnn4nd0: well I don't care about that competitive stuff15:04
blackburnlections are here anyway15:04
blackburnerr lectures15:04
n4nd0true15:04
blackburnand you always can submit code15:04
n4nd0we should try to take to it and motivate each other not to drop it15:05
blackburnmotivation should come from research interest15:05
n4nd0that one seems so far away...15:06
blackburnwhat do you mean?15:06
n4nd0I see research something sooo far15:08
blackburnno it is already here in front of you15:09
blackburnstaring into your eyes :)15:09
n4nd0haha we shall see15:09
n4nd0the compilation time in my computer kills me15:11
n4nd0it makes me lose time for real15:11
blackburn--disable-optimization, multiple threads?15:11
blackburnccache?15:11
blackburnare you using that15:12
blackburnor are you talking about interfaces?15:12
n4nd0I use ccache make -j2 normally15:12
n4nd0disable-optimization on15:12
n4nd0normally just with python-modular support15:12
blackburnwell what can I suggest is to use libshogun only15:13
blackburnor lets try to find a way to split interface file15:13
blackburnn4nd0: hah were you here already then we disabled optimizations in interfaces?15:13
blackburnbefore it took 2-3 times more15:14
n4nd0I don't think so15:14
n4nd0really?15:14
blackburnwe are now using -O0 in interfaces15:14
blackburnbut before it was global15:14
blackburn-O3 or -O015:14
blackburnfor everything15:14
n4nd0can imagine15:15
blackburnn4nd0: but in general I agree we really need to improve dev experience15:16
blackburnn4nd0: what mostly slows down compilation is template classes15:20
n4nd0I see15:21
blackburnn4nd0: one of reasons I felt a little pain in my bottom when you introduced matrix features :D15:21
n4nd0it could be so cool if it is just re-compiled what is changed15:21
n4nd0when one changes a header file, all over again15:21
n4nd0haha15:22
n4nd0they were required I think, didn't find anything that could do the job15:22
n4nd0though I looked for it!15:22
blackburnyes, dense wouldn't work15:22
blackburnokay dense subset is not really needed we could merge it to dense15:23
blackburnargh no pluskid around15:24
blackburnokay I have to add a unit test then15:25
n4nd0blackburn: PR updated15:45
blackburnokay let me check15:47
CIA-52shogun: Sergey Lisitsyn master * r093b502 / (7 files): Merge pull request #756 from iglesias/master (+6 more commits...) - http://git.io/_rSCmA16:02
CIA-52shogun: Chiyuan Zhang master * rba02284 / src/shogun/multiclass/MulticlassOneVsOneStrategy.cpp : tie-breaking by decision values in One-vs-One. - http://git.io/uEjFmg16:10
CIA-52shogun: Sergey Lisitsyn master * r2f3688a / src/shogun/multiclass/MulticlassOneVsOneStrategy.cpp : Merge pull request #753 from pluskid/multiclass - http://git.io/kxI4lg16:10
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving]16:20
-!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has joined #shogun16:28
yoohi all16:32
blackburnhi16:33
-!- alexlovesdata [~binder@goldenezahl.ml.tu-berlin.de] has joined #shogun17:14
-!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has quit [Quit: Page closed]17:41
-!- uricamic [~uricamic@2001:718:2:1634:e9e5:5ab4:359d:813d] has quit [Quit: Leaving.]17:44
CIA-52shogun: Sergey Lisitsyn master * rd50e297 / src/interfaces/modular/modshogun_ignores.i : Ignored SGVector operator+ and operator+= - http://git.io/t-TRzA17:45
@sonney2kyoh_, I think you can safely assume that debian wheezy will have shogun from squeeze18:03
-!- gsomix [~gsomix@178.45.65.87] has joined #shogun18:05
yoh_sonney2k: 0.9.3? how ? atm there is no single version in wheezy and only viable one to enter is in unstable/sid which is 1.1.0...   or do you mean that wheezy will not have any version? ;)18:11
yoh_btw -- probably for my own sake I will just build backports of 1.1.0 and distribute from neuro.debian.net18:12
blackburnyoh_: why not 2.0?18:12
yoh_blackburn: is it released? well tested? is there a debian package? ;)18:12
blackburnyoh_: is that done for 1.1.0? :D18:12
blackburn2.0 will be released on 1st of sept18:13
yoh_blackburn: 1.1.0 was out there for a while so hopefully it is a bit more tested ;)18:13
blackburnyoh_: no it is not18:13
blackburnI am sure 2.0 have less bugs than 1.1.018:14
yoh_blackburn: it would be great! ;)  so, ok -- is 1.1.0 worth hassle over 0.9.3?18:15
blackburnyes, sure18:15
blackburnbut 2.0 is worth even more :)18:15
blackburnnot feature-wise but bug-wise18:16
blackburnthere were a few bugs that has to be merged 1.1.0 if it was supported18:16
blackburnto 1.1.0*18:16
blackburnyoh_: 0.9.3 has no java, no ruby.. etc18:19
blackburnyoh_: two major (I do mean major) new features are director swig stuff and structured output framework18:20
yoh_good ;-) FWIW: myself I care about python  ;)18:23
blackburnyoh_: then directors are for you probably18:24
blackburnand zero-copy features18:24
blackburnyoh_: directors allow you to take care of e.g. kernels computing using python code18:25
blackburnor to define structured model via python code18:25
blackburn(in case you didn't know what it is) :)18:26
yoh_so, 1st of sept there also would be debian packages uploaded? ;)18:31
blackburnyoh_: I am just a stupid developer, let you rather ask soeren about that :)18:32
yoh_sonney2k: btw -- from the debian build log it is not clear -- are you running unit/regresion tests at pkg build time?  it might be cool if you did ;-)18:34
blackburnto be serious I am unsure about that - sonney2k takes care about that18:34
blackburnyoh_: can I see these logs?18:34
yoh_blackburn: no way -- they are a property of the communist part of C.!18:35
yoh_nah... of D.!18:35
yoh_;) https://buildd.debian.org/status/fetch.php?pkg=shogun&arch=armhf&ver=1.1.0-6&stamp=134553985318:35
yoh_more from here: https://buildd.debian.org/status/package.php?p=shogun18:35
blackburnoh my it is pretty big18:36
blackburnsuper lu is absent18:37
blackburnstrange, okay not a problem18:37
CIA-52shogun: Sergey Lisitsyn master * rcbc0150 / src/shogun/structure/StructuredModel.cpp : Changed type of loop variable in StructuredModel - http://git.io/JZa_1g20:07
@sonney2kyoh_, of course we don't run any tests... otherwise the build will take >1d on arm etc20:22
blackburnsonney2k: are you willing to have modsel typemaps in 2.0?20:22
blackburnwe have to freeze at some point20:22
blackburnsonney2k: has TU bb gone?20:27
blackburncygwin one20:27
@sonney2kblackburn, yeah20:28
@sonney2kblackburn, extra typemaps for these kind would be nice but if they dont' make it we put them in 2.1 right?20:29
@sonney2kor 2.0.120:29
blackburnexactly20:29
blackburnbut 2.0 should be more or less stable20:29
@sonney2kblackburn, however these are totally non-intrusive and won't hurt if not yet ready20:29
@sonney2ksure20:29
@sonney2kas stable as possible20:29
blackburnI have time now but a little lost what to do20:29
blackburnnow I am finishing multiclass output requested by yoo a few times20:30
@sonney2kblackburn, update NEWS? try to get tests to work perfectly (even old ones?)20:30
blackburnoh NEWS20:30
blackburnI love to update news :D20:30
blackburnsonney2k: however warnings are at soft zero20:31
blackburnthat's actually nice20:31
blackburnahh I need to implement director structured model20:32
blackburnvery important thing to have20:32
@sonney2kyeah but could as well be in 2.0.120:32
@sonney2ktests / doc is probably most important now20:32
blackburnno, disagree here20:32
blackburnsounds very important20:32
blackburnhowever20:32
@sonney2kblackburn, very few people will use directors!20:32
blackburnmay be I am wrong20:32
blackburnsonney2k: BMRM + python code sounds vvverrry interesting20:33
blackburnpure prototyping20:33
@sonney2kahh and we have to write some blog entry with nice figures ...20:33
@sonney2kthough I didn't receive any further ones...20:34
blackburnlol20:34
blackburnwell I really thought about that20:34
blackburnbut still so hard to came with any idea for me20:34
blackburn:(20:34
@sonney2kblackburn, well multiclass is easy ... just illustrate e.g. 3-5 gaussians and decision surfaces for all the MC methods20:34
@sonney2kmultitask is more difficult20:34
blackburnhmm multitask could be20:35
@sonney2kbut one could create one toy example showing the solution for the 'vanilla' svm and one for a dual task one20:35
blackburnlike20:35
@sonney2kso show how solutions changes...20:35
blackburnpictures coming from 3 sources20:35
blackburnI'll try something20:35
blackburndimension reduction is the easiest here20:36
@sonney2kfor SO  / latent variable SO we really need some fancy example like license plate recognition, speech recognition, background detection (in images)...20:36
@sonney2kbut that is lots of work20:36
blackburnsonney2k: good example, sound like Phd20:36
blackburn:D20:36
@sonney2kblackburn, no20:36
@sonney2knot when you have all the methods available already20:37
@sonney2kblackburn, it is of same effort as your traffic sign detector20:37
blackburnsonney2k: I still didn't get how could that even work (background detection with SO)20:37
blackburnsonney2k: I totally failed with L1/Lq btw :(20:37
blackburnI thought it is worth to regularize cells with L220:37
@sonney2kblackburn, it is a segmentation problem you have to decide for each pixel if it is background or foreground20:38
blackburnsonney2k: yes, I understand20:38
blackburnbut how technically not yet20:38
@sonney2khowever doing so independently (naiive way) misses correlations... so SO can help here20:38
@sonney2kblackburn, I am not an image guy so better ask wiking - he did things like this IIRC20:39
blackburnsonney2k: can that be done with HM having #pixels states?20:39
@sonney2kblackburn, ahh for GPs we have nice figures right?20:39
@sonney2kblackburn, between 2 ... #pixels states :D20:40
@sonney2k#pixels sounds a bit stupid20:40
blackburnwhy?20:40
blackburnargh20:40
blackburnso hidden states are20:40
@sonney2kwould then be like SVM20:40
blackburnback and fore?20:40
@sonney2kyes20:40
blackburnwhat are observations?20:41
blackburnpixels?20:41
@sonney2kbut maybe some more20:41
blackburnintensities?20:41
blackburnor some fancy stuff like hogs would work too?20:41
@sonney2ksome features computed based on pixels / neighborhood20:41
@sonney2kI would guess so20:41
blackburnlbp etc?20:41
blackburnI see20:41
@sonney2kyeah20:41
blackburnwell20:41
blackburnI could develop an example then20:41
blackburnif I understand correctly20:42
blackburnwe have everything20:42
blackburnto do that20:42
@sonney2kwell one thing maybe not20:42
blackburnwhich one?20:42
@sonney2kI only understand how things work for 1d images20:42
blackburn:D20:42
@sonney2kso you would go through an image from left to right20:42
blackburn1d images?20:42
blackburnlol20:42
@sonney2k(one line only :)20:42
@sonney2kand get a state sequence F-B-B-B-B-F-F ... etc20:43
blackburnwhy not to flatten it row by row then?20:43
@sonney2kblackburn, the dependencies matter20:43
blackburnahh wait wait20:44
blackburnso we observe only neighborhood20:44
blackburnof each pixel20:44
blackburnnot the whole image20:44
blackburnright?20:44
@sonney2kyes20:44
blackburnI start to understand these pictures I have seen with CRFs segmentation20:44
blackburnthere they had 9 observations and some network of it20:44
@sonney2kbut in SO you have the state model (simplest 2 states F & B)20:44
@sonney2kand you learn some weights going from F -> B and back20:44
@sonney2kso this somehow needs to be taken into account20:45
@sonney2kyeah it is more like a network20:45
@sonney2k(e.g. markov network)20:45
blackburnokay I see20:46
blackburnyou made me interested I should try20:46
@sonney2kblackburn, in this sense speech recognition is much simpler20:46
blackburnthis is the application I do not understand even more20:47
blackburnin means of features20:47
blackburnhow do they construct features?20:47
@sonney2kyou just have a number of states (phonems) and combine these phonems into most likely words20:47
blackburnhow do they detect phonems?20:47
@sonney2kblackburn, well a sample of a phonem20:47
blackburnis it a solved problem?20:47
@sonney2kblackburn, yes20:48
blackburnokay I see20:48
@sonney2kwell basically you have some known phonems in training20:48
@sonney2kand then you can fit gaussians or so on them20:48
@sonney2k(HMMs with guassians are the standard approach for that task)20:48
blackburnsonney2k: can I fit segmentation HM with one image?20:49
blackburnwould that work at all?20:49
blackburnnot in generalization means but at least would it segment the image I used?20:49
@sonney2kwell if you used say 30% of the image for training then maybe20:50
@sonney2kblackburn, the methods here are all supervised20:50
blackburnin other words20:51
@sonney2kso not like a HMM that could (in theory) also segment the image in one go20:51
blackburnwould it reproduce the same pattern?20:51
blackburnat least approximately20:51
@sonney2kblackburn, yes20:51
blackburnsounds like pretty interesting thing20:52
@sonney2kdepending on #states it will reproduce the exact image :D20:52
blackburnI have to do that because I hardly believe :D20:52
@sonney2klet me better work on the website...20:54
blackburninstead of?20:54
blackburnsonney2k: do you have any (just any) idea what could make our HMM class fail?20:56
blackburnI assume it was working before?20:57
@sonney2kblackburn, does it fail now?20:58
blackburntest - yes20:58
@sonney2kahh test... ohh well it is difficult. randomly initialized parameters NP hard problem...20:59
@sonney2kI would need to check...20:59
@sonney2kif the 3 dice problem still works then it is good though20:59
blackburnI see21:00
@sonney2k(3 dice with different distributions are drawn lots of times - a 3 state HMM can learn the distribution of the dice figures and when they were drawn)21:00
yoh_sonney2k: well -- it would be neat if tests were run... at least selectively for some architectures... and yeah -- it does take a while to build some of my packages on ARM because of that... ARM people should make faster chips/ports ;)21:06
yoh_imho it is better to make machines sweat than later on troubleshoot bug reports (although it is indeed unlikely to see shogun users on eg s390 ;) ) ;)21:08
blackburnWe have processed the evaluation for your project named Implement multitask and domain adaptation algorithms with Shogun Machine Learning Toolbox.21:09
@sonney2kblackburn, and failed as usual ;-)21:11
@sonney2kI have 8 emails of this kind21:12
blackburnsonney2k: how do I manage to fail it year by year?21:20
blackburnI am tired of it already :(21:20
@sonney2kat least now the Rubel rolls right :D21:20
@sonney2kblackburn, http://94.23.237.10:5000/21:47
* sonney2k Zzzz...22:00
blackburnsonney2k: doesn't work for me22:06
-!- AndChat49209 [~AndChat49@host86-178-156-19.range86-178.btcentralplus.com] has joined #shogun22:33
AndChat49209.22:33
AndChat49209.22:34
-!- AndChat49209 [~AndChat49@host86-178-156-19.range86-178.btcentralplus.com] has quit [Client Quit]22:34
-!- AndChat49209 [~AndChat49@host86-178-156-19.range86-178.btcentralplus.com] has joined #shogun22:35
blackburn. :)22:35
AndChat49209Its heiko with mobile but dont know how to change name22:36
blackburnhello mobile heiko22:36
blackburnhow are you and your thesis?22:36
-!- heiko1 [~heiko@host86-178-156-19.range86-178.btcentralplus.com] has joined #shogun22:37
heiko1hey22:37
blackburnhey22:37
heiko1dude this is horros on the mobile keyboard :)22:37
heiko1thesis is coming along, has been a prett yintense week22:37
blackburnI imagine22:38
heiko1as always: its more work than I thought :)22:38
blackburnhaha22:38
blackburnyeah22:38
heiko1but still 2 weeks left22:38
heiko1so this will be ok22:38
blackburnI thought I will do my bachelor thesis in a week LOL22:38
heiko1thats optimistic :)22:39
blackburnvery22:39
heiko1but I have talked to people who havent start writing things down yet22:39
-!- AndChat49209 [~AndChat49@host86-178-156-19.range86-178.btcentralplus.com] has quit [Client Quit]22:39
blackburnI knew quite a few guys who acted this way22:39
blackburnthey are in army now22:39
blackburn:D22:39
heiko1lol :)22:39
blackburnwon't work for you though22:39
heiko1no too old already22:40
heiko1(and did civil service)22:40
blackburnI have 6 more years of hiding22:40
blackburn:D22:40
heiko1oh man22:40
heiko1come to Germany :)22:40
blackburnheiko1: well it won't work :)22:40
blackburneverybody would come to germany if it allows to avoid army :D22:41
heiko1ok22:41
heiko1is your thesis processed already?22:41
blackburnthere is a traditional russian way22:41
blackburnprocessed in which mean?22:41
heiko1like feedback22:41
blackburnI am unsure what do you mean :)22:42
blackburnfeedback from?22:42
heiko1supervisors22:42
heiko1like marks22:42
blackburnah22:42
blackburnyes I defended it already22:42
heiko1oh really, nice22:42
heiko1congrats then :)22:42
blackburnhah thanks22:42
blackburnbut I am a bachelor for two months already22:43
heiko1cool22:43
heiko1whats next?22:43
blackburnmasters here22:43
heiko1computer science?22:43
heiko1what courses and stuff?22:44
blackburnoh well22:44
blackburnyou know we are crazy here22:44
blackburnI will study just what they give me22:44
blackburn:D22:44
heiko1and that is?22:44
blackburnI have nothing to choose but it should be related to supercomputing22:44
blackburnI don't care much about that - I am coming along with shogun and other stuff22:45
blackburnI already had to learn everything by myself22:45
heiko1ok22:45
blackburnso no reason to change that22:45
blackburn:)22:45
heiko1well then :)22:45
blackburnheiko1: what about you then? :)22:46
blackburnare you going to do phd there?22:46
heiko1well, thesis first22:46
heiko1I got a small RA position here for about half a year afterwards22:46
heiko1and then I need another one :)22:46
blackburnresearch assistant?22:46
blackburnI have no clue what it is22:46
heiko1but i will apply for a phd that starts in sept22:46
heiko1yeah assistant22:46
blackburnso it is just like a normal job?22:47
heiko1no its uni and research stuff22:47
heiko1but applied22:47
heiko1kernel methods to detect xy22:47
blackburnyeah I mean normal research job22:47
heiko1yes22:47
blackburnI see22:47
blackburnis it well-paid in england btw?22:47
heiko1yes, especially in london22:48
blackburnI heard londongrad is not cheap to live22:48
blackburn:)22:48
heiko1no, its not cheap at all22:48
heiko1but my girlfriend has a well paid job, and I will join in that fashion, so its fine -- finally last year was so anoyying money wise22:49
blackburnwhat does she do?22:49
heiko1she works in the citi, for a digital media data mining company22:49
blackburnwow so a data mining?22:49
heiko1well, old fashioned data mining22:49
heiko1nothing advanced22:50
heiko1but collecting data22:50
blackburnstill geeky22:50
blackburn:)22:50
heiko1shes more in the organisation and planning22:50
blackburnI see22:50
blackburnis she german too?22:50
blackburnI am currently debugging MC output stuff22:51
blackburnwhich collects ROCs22:51
heiko1yeah22:51
heiko1monte carlo ?22:51
heiko1multiclass :)22:51
blackburnmulticlass :)22:51
heiko1sorry22:51
blackburnbut I don't mind to have some monte carlo out there22:52
blackburnmay be some casino royale22:52
heiko1Oh, I want to work on the labels a bit on the weekend22:52
blackburnwould be interested22:52
blackburn:D22:52
heiko1lol :)22:52
blackburnhave you seen Ernst&Young guy?22:52
blackburnin the ML22:52
heiko1yeah22:52
heiko1pretty cool :)22:52
heiko1machine learning arrives in the consulting22:53
blackburnwe are all perplexed how can it be used22:53
heiko1dont know22:53
blackburnyeah22:54
heiko1all the bankers are using (bayesian) linear regression22:54
blackburnwe was puzzled with the fashion he names himself22:54
heiko1one of our profs here always tells us whats going on in the citi22:54
heiko1linear regresseion :)22:54
blackburnheh22:54
heiko1I did not get your last scentence22:55
blackburnhe names himself22:55
blackburnarthur (sean)22:55
blackburnat first he wrote sth like22:55
blackburncheers, sean22:55
blackburnbut email is arthur22:55
blackburnthat's kind of puzzle :D22:55
heiko1ah, mmh no idea22:56
heiko1blackburn, what about the labels22:56
heiko1with the confidences /scores22:56
blackburnah22:56
heiko1have you seen s?rens email?22:56
blackburnoh bad bad22:56
blackburnwhich one?22:56
heiko1I liked the idea of registering multiple vectors with a name22:56
blackburnI remember one22:56
heiko1and then you set the "current" one22:56
heiko1and call a method22:56
heiko1this way you can switchj22:56
blackburnthe most confusing thing for me is that we all wanted to release22:57
heiko1yeah maybe leave that then22:57
heiko1the method is there22:58
heiko1though its only for binary22:58
blackburnI have checked multiclass stuff a little22:58
heiko1argh there is still the migration stuff22:59
blackburnoops22:59
heiko1that will be horrible22:59
blackburnso do we need to set up all the maps now?22:59
heiko1all these changes to the vectors cause loads of memory bugs22:59
heiko1dont know yet22:59
heiko1if we can migrate, maybe22:59
heiko1but I think this wont work22:59
heiko1so we will break22:59
heiko1and start from scratch22:59
heiko1that stuff is a mess, especially in c++ without any libraries23:00
blackburnargh mkl storage fails with mkl multiclass23:00
heiko1and whenever somethings changed, it fails23:00
heiko1oh reall, what happens?23:00
blackburnsome error message I can't see23:01
blackburnhmm23:01
blackburnahhhhh23:01
blackburnheiko1: it handles MKLMulticlass as multiclass23:02
heiko1and then?23:03
blackburnhmmm23:03
blackburnsomething strange happens then23:03
heiko1mmh, if errors should be easy to track23:04
blackburnheiko1: we should mark down all these *outputs as experimental stuff23:04
blackburnto aavoid some psycho trying to find us and eat23:04
heiko1mmh23:05
heiko1I mean the idea is that only classes that are supported work with it23:05
heiko1so every class gets another implementation of the output class23:05
heiko1the output class can ignore things23:06
blackburnheiko1: are you able to check code?23:06
heiko1yeah, I want to go to bed, but send it :)23:07
blackburnah nevermind I'd rather paste23:07
blackburn 84 ????????????SG_PRINT("%smulti-class machine %d:\n", i);23:07
blackburnsomething is wrong out there23:07
blackburnwhat was the intention?23:07
blackburnthere are two formatters but one var23:07
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun23:08
blackburnn4nd0: !23:08
n4nd0blackburn: hey!23:08
n4nd0are we all happy with the final evaluations? did everyone pass?23:08
blackburnyes I think so23:08
heiko1should :)23:08
heiko1at least I am happy23:08
blackburnI wanted to develop a simple application that segments an image23:09
n4nd0blackburn: ok23:09
blackburnso it is a two state model I guess23:09
blackburnI wanted to ask if we are ready for that23:09
blackburncan that be done using current framework?23:09
n4nd0I have not much idea about image segmentation23:09
n4nd0can you tell me more about the application?23:09
blackburnn4nd0: okay let me try to describe23:10
heiko1blackburn, talk to you later, I need to sleep23:10
blackburnheiko1: okay sure23:10
n4nd0heiko1: good night!23:10
blackburnn4nd0: simplest case is image having foreground and background23:10
heiko1night guys! :)23:10
blackburnnite heiko123:10
blackburnn4nd0: so two hidden states23:10
blackburnand each pixel has its own state23:11
blackburnobservations are not so clear to me though23:11
n4nd0blackburn: the actual values of the pixels I guess23:12
blackburnyes but in neighborhood of each pixel23:12
blackburnI guess23:12
n4nd0and probably something that gives infor  about the pixels around23:12
n4nd0exatly23:12
n4nd0exactly*23:12
blackburnI still have weak belief how can that work :D23:13
n4nd0I'd say that the framework should work for that23:13
n4nd0but we need to find some information about the features that people use for this23:13
blackburnit is not a problem lets try to come along with pixels23:13
n4nd0blackburn: have you read about it in some paper?23:13
blackburnno, not really we discussed it with soeren23:13
blackburnso it comes from air23:14
blackburnand can be wrong :D23:14
n4nd0I think so far sounds reasonable23:14
blackburnI have to draw23:14
-!- heiko1 [~heiko@host86-178-156-19.range86-178.btcentralplus.com] has left #shogun []23:14
n4nd0but more research on the topic is probably necessary though23:14
blackburnI had a link somewhere23:14
blackburnto web service allowing to draw23:15
blackburn:D23:15
blackburnn4nd0: do you know any service like that?23:16
n4nd0mmm no23:16
blackburnhttp://www.twiddla.com/49623423:16
blackburnn4nd0: here you  go23:16
blackburnn4nd0: waiting for you ^ :)23:18
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Ping timeout: 240 seconds]23:38
CIA-52shogun: Sergey Lisitsyn master * rcfc8322 / (8 files in 3 dirs): Introduced CV MC storage - http://git.io/a_MgjA23:43
-!- gsomix [~gsomix@178.45.65.87] has quit [Ping timeout: 256 seconds]23:56
--- Log closed Sat Aug 25 00:00:17 2012