IRC logs of #shogun for Tuesday, 2011-11-29

« Monday, 2011-11-28 Index Wednesday, 2011-11-30 »
--- Log opened Tue Nov 29 00:00:59 2011
-!- blackburn [~blackburn@188.168.5.8] has quit [Quit: Leaving.]06:28
-!- blackburn [~blackburn@188.168.2.74] has joined #shogun11:22
sonne|workblackburn: what happened?12:11
blackburnsonne|work: nothing, I did git bisect but went to wrong commit I guess12:12
blackburnit takes too long to do it again now12:12
sonne|workyou mean you marked one commit bad that wasn't ?12:13
sonne|work(or good)12:13
blackburnmay be..12:14
blackburnmay be I did not clean everything well when recompiling12:14
blackburnI could do it again later, but not today12:15
sonne|workhmmhh too bad - I guess one should do git clean -dfx or similar12:15
sonne|workand use ccache to speed up the procedure12:15
blackburnsonne|work: btw I ran GMNPSVM this night12:15
blackburnbefore going sleep hah12:15
blackburnand it was one more percent accurate12:16
sonne|workI made some progress on the typemap stuff but I would rather want to adjust all typemaps now12:16
blackburn88.8% or so12:16
sonne|workas I expected12:16
sonne|workI guess you could only gain more when doing these distortions to make the machine learn about invariances12:16
blackburnI forgot to add time measure12:16
sonne|workanyway back to work12:16
sonne|workcu12:16
blackburnok see you12:16
blackburnsonne|work: there is a way I can gain still - RoI, custom color/shape kernels, etc12:19
blackburnbattle is not lost :D12:20
sonne|workis this number accuracy?12:39
sonne|workhow many classes?12:39
blackburnsonne|work: 4312:40
sonne|workdamn good then already12:40
sonne|workwill be hard to improve12:40
blackburnsonne|work: CNN got 99% :D12:40
sonne|workhow many examples / class?12:41
sonne|worksounds like cheating...12:41
blackburnhttp://benchmark.ini.rub.de/?section=gtsrb&subsection=news12:42
blackburnsonne|work: http://benchmark.ini.rub.de/?section=gtsrb&subsection=results12:43
blackburnas you can see Commitee of CNNs performed better than human :D12:43
sonne|workas usual12:43
sonne|work;-O12:43
sonne|workLDA on HOG seems to be doing pretty well12:44
blackburnI hope SVM will do the same12:44
blackburnthese guys went crazy12:45
blackburnLDA + HOG + ... + ...12:45
blackburnthey combined everything :D12:45
blackburnsonne|work: is LDA success surprising to you? I do not know much except how it works12:46
blackburnhahah12:47
blackburnSHOGUN KLTSA applied, took 7.867928s12:47
blackburnSKLEARN LTSA applied, took 483.886713s12:48
blackburnsonne|work: that's ^ not fair to publish :D12:48
sonne|worknice work12:48
blackburnwhy is it so slow..12:49
blackburnsonne|work: the most strangest commit msg I've ever seen: http://code.google.com/p/locally-linear-embedding/source/detail?r=3012:53
sonne|work:)12:54
blackburnI guess I won't compare with such strange project :D12:54
sonne|workI am impressed that they have 96 % with random forests13:06
sonne|worksame for LDA + HOG13:07
sonne|worknot bad13:07
CIA-33shogun: Sergey Lisitsyn master * r6aebf7d / src/shogun/converter/LocalTangentSpaceAlignment.h : Warn that LTSA is slower than KLTSA with linear - http://git.io/rd7uPA13:07
blackburnsonne|work: I heard random forests are good for similar tasks13:07
blackburnsonne|work: for vehicle (numbers? signs? bad english) recognition13:08
blackburn:D13:08
blackburnsonne|work: madeleine's issue solved completely, all went well13:49
blackburnbut we know that java has some leaks13:49
sonne|workhow did she solve it?13:49
sonne|workwhich leaks?13:49
sonne|worktoo bad that this is not public on the mailinglist13:50
blackburnsonne|work: you did, and I13:50
blackburnsonne|work: same script in java was eating too much memory13:50
blackburnoh I have ten minutes to get myself to job haha13:50
sonne|workbut where13:50
sonne|workhmmhh13:50
blackburnsonne|work: I will try to reproduce it later13:51
blackburnbe right back in twenty minutes haha13:51
sonne|workyeah lets get a releasy out13:51
-!- blackburn [~blackburn@188.168.2.74] has quit [Quit: Leaving.]13:51
sonne|workrelease out before that13:51
sonne|workcu13:51
sonne|workI will be away then13:51
-!- in3xes [~in3xes@180.149.49.230] has joined #shogun14:19
-!- blackburn [5bdfb203@gateway/web/freenode/ip.91.223.178.3] has joined #shogun14:23
blackburnoh wow14:25
blackburnblocker bug on canada project :D14:28
blackburncanadian*14:28
-!- blackburn [5bdfb203@gateway/web/freenode/ip.91.223.178.3] has quit [Ping timeout: 265 seconds]15:45
-!- in3xes_ [~in3xes@180.149.49.230] has joined #shogun16:04
-!- in3xes_ [~in3xes@180.149.49.230] has quit [Client Quit]16:06
-!- in3xes [~in3xes@180.149.49.230] has quit [Ping timeout: 240 seconds]16:08
-!- blackburn [~blackburn@188.168.2.74] has joined #shogun17:59
@sonney2kcanadian?20:05
blackburnsonney2k: yeah nevermind :)20:06
blackburnrecall I work in telecom software company20:06
-!- supernaught [~hal@host3-196.wifi.ubc.ca] has left #shogun []20:57
@sonney2kblackburn, how big is wc -l interfaces/python_modular/modshogun_wrap.cxx for you?22:22
blackburn41891222:23
@sonney2k42671222:23
blackburnsonney2k: are you trying to reduce its size again?22:24
@sonney2kI am fixing python typemaps22:30
blackburnI see22:33
@sonney2kwe wrap a lot of stuff that isn't supported22:51
@sonney2ke.g. sparsevector22:52
@sonney2kor functions that return sgvector of gaussian kernels etc22:52
blackburnsonney2k: yeah a lot of useless stuff22:54
@sonney2kand I ran into some weird crashes when we disable certain data types22:57
CIA-33shogun: Soeren Sonnenburg master * r4bce5ac / (3 files in 3 dirs):23:16
CIA-33shogun: Overhaul python typemaps23:16
CIA-33shogun: - check that we don't have a list as input23:16
CIA-33shogun: - refactor code to call static functions to only have minimal23:16
CIA-33shogun: functionality inside the actuall typemap23:16
CIA-33shogun: - re-enable boolean type by default to fix crasher - http://git.io/SybNbA23:16
--- Log closed Wed Nov 30 00:00:59 2011
« Monday, 2011-11-28 Index Wednesday, 2011-11-30 »

Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!