IRC logs of #shogun for Thursday, 2012-05-17

« Wednesday, 2012-05-16 Index Friday, 2012-05-18 »
--- Log opened Thu May 17 00:00:40 2012
CIA-113shogun: Sergey Lisitsyn master * r6580ef2 / src/shogun/features/Labels.cpp : Fixed function introduced in the last commit - http://git.io/6vHSBQ00:18
blackburnhehe stupid coding00:18
-!- blackburn [~qdrgsm@188.122.250.167] has quit [Ping timeout: 252 seconds]01:34
-!- emrecelikten [~Anubis@217.131.4.22] has quit [Quit: Leaving.]01:54
-!- Francis_Chan [~Adium@58.194.224.108] has joined #shogun02:17
-!- av3ngr [av3ngr@nat/redhat/x-qkezysfryayimbjt] has joined #shogun02:53
-!- Francis_Chan [~Adium@58.194.224.108] has quit [Quit: Leaving.]03:47
-!- abn_ [av3ngr@nat/redhat/x-luuwxqrxbhzoojli] has joined #shogun03:57
-!- av3ngr [av3ngr@nat/redhat/x-qkezysfryayimbjt] has quit [Ping timeout: 272 seconds]04:02
-!- av3ngr [av3ngr@nat/redhat/x-jkdqgjkzryrnmdft] has joined #shogun05:10
-!- abn_ [av3ngr@nat/redhat/x-luuwxqrxbhzoojli] has quit [Ping timeout: 240 seconds]05:14
-!- av3ngr [av3ngr@nat/redhat/x-jkdqgjkzryrnmdft] has quit [Read error: Connection reset by peer]06:37
-!- av3ngr [av3ngr@nat/redhat/x-zzweqaftcracouxr] has joined #shogun06:38
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]06:49
-!- abn_ [av3ngr@nat/redhat/x-lsiqhomuwggtathm] has joined #shogun08:16
-!- av3ngr [av3ngr@nat/redhat/x-zzweqaftcracouxr] has quit [Ping timeout: 245 seconds]08:20
-!- uricamic [~uricamic@2001:718:2:1634:a8ca:b8c8:8e49:b04c] has joined #shogun08:56
-!- blackburn [~qdrgsm@188.122.250.167] has joined #shogun09:10
-!- Francis_Chan [~Adium@59.108.60.58] has joined #shogun09:41
-!- abn_ [av3ngr@nat/redhat/x-lsiqhomuwggtathm] has quit [Quit: That's all folks!]10:33
-!- Marty28 [~marty@cable-158-181-77-81.cust.telecolumbus.net] has joined #shogun11:00
-!- Marty28 [~marty@cable-158-181-77-81.cust.telecolumbus.net] has quit [Quit: Colloquy for iPad - http://colloquy.mobi]11:10
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun11:41
-!- blackburn [~qdrgsm@188.122.250.167] has quit [Quit: Leaving.]12:31
-!- wiking [~wiking@c-98-231-171-175.hsd1.md.comcast.net] has joined #shogun13:09
-!- wiking [~wiking@c-98-231-171-175.hsd1.md.comcast.net] has quit [Changing host]13:09
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun13:09
-!- Francis_Chan [~Adium@59.108.60.58] has quit [Quit: Leaving.]13:35
-!- vikram360 [~vikram360@117.192.186.253] has joined #shogun14:47
-!- pluskid [~pluskid@111.120.28.216] has joined #shogun14:51
-!- Francis_Chan [~Adium@58.194.224.108] has joined #shogun16:11
-!- vikram360 [~vikram360@117.192.186.253] has quit [Ping timeout: 260 seconds]17:16
-!- pluskid [~pluskid@111.120.28.216] has quit [Quit: Leaving]17:40
-!- blackburn [~qdrgsm@188.122.250.167] has joined #shogun17:46
n4nd0is there any reason why CList is not made with template?18:03
blackburnn4nd0: some serialization issue I am afraid18:04
blackburnbut I am not sure18:04
n4nd0aham18:04
n4nd0I think it would be more comfortable with templates18:05
n4nd0not so many casts when returning elements for example18:05
blackburnn4nd0: try to ask soeren once he is online18:11
n4nd0blackburn: I'll do that, thanks ;)18:11
-!- Francis_Chan [~Adium@58.194.224.108] has quit [Ping timeout: 252 seconds]18:29
-!- gsomix [~gsomix@37.61.181.74] has quit [Quit: Ex-Chat]18:37
-!- blackburn [~qdrgsm@188.122.250.167] has quit [Quit: Leaving.]18:48
-!- blackburn [~blackburn@188.122.250.167] has joined #shogun18:50
n4nd0hey wiking, how is it going?19:05
-!- uricamic [~uricamic@2001:718:2:1634:a8ca:b8c8:8e49:b04c] has quit [Quit: Leaving.]19:10
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Ping timeout: 272 seconds]19:17
-!- vikram360 [~vikram360@117.192.186.253] has joined #shogun19:27
blackburnstudents are getting crazy waiting for welcome packages20:00
-!- Netsplit *.net <-> *.split quits: @sonney2k20:10
-!- Netsplit over, joins: sonney2k20:11
-!- emrecelikten [~Anubis@82.222.148.100] has joined #shogun20:29
emreceliktenHello all20:30
blackburnhi20:31
emreceliktenStumbled upon this: http://groups.csail.mit.edu/netmit/sFFT/code.html20:34
blackburnhah yeah that cat is cute20:34
-!- blackburn [~blackburn@188.122.250.167] has quit [Quit: Leaving.]20:52
-!- blackburn [~blackburn@188.122.250.167] has joined #shogun20:54
sonney2kemrecelikten, heh :)21:06
emreceliktenMight be useful in future I guess21:07
emreceliktenNot the cat :P21:07
sonney2knot until you did a fft^-121:07
emreceliktensonney2k: :D21:08
blackburnis anybody aware of fourier features usage in CV?21:08
sonney2kn4nd0, YMMV. it used to be template based but IMHO it is much more clean to use SGObject as the required base, since you know you can do SG_REF/UNREF21:09
sonney2kbut I agree lots of casting...21:10
blackburnsonney2k: when did you plan next meeting?21:13
sonney2kcheck the logs 2x.05.21:19
-!- mode/#shogun [+o sonney2k] by ChanServ21:21
@sonney2kblackburn, I've changed all the label stuff - but one question pops up. what shall classifiers output?22:02
@sonney2kI mean BinaryLabels, MulticlassLabels, Regressionlabels22:03
@sonney2kare all specific22:03
blackburnsonney2k: hehe depends whether you did add some base class there22:03
@sonney2kno22:03
blackburniirc it is ok to specialize class in descendant22:03
blackburnthen you will have to make apply class specific..22:04
@sonney2kwell we have CLabels as base class22:04
@sonney2kbut it won't help us22:04
blackburnwhy not?22:04
@sonney2kbecause from python labels will be an abstract class22:04
@sonney2klike features22:04
blackburnsonney2k: as I can see the only way is to move apply to specific machines22:05
@sonney2kso I only see two solutions: RealvaluedLabels or OutputLabels22:05
@sonney2kblackburn, you mean binarymachine?22:06
@sonney2kmulticlassmachine?22:06
@sonney2kregressionmachine?22:06
blackburnsonney2k: yes binary machine, multiclass machine all has they own apply22:06
@sonney2kmakes sense22:06
@sonney2kso Machine would loose its 'apply' interface function22:06
@sonney2konly problem I see is that we have kernel machine22:06
@sonney2kand that can be binary.mc, regression22:07
@sonney2kany of it22:07
blackburnKernelBinaryMachine..22:07
@sonney2ksame w/ linear22:07
blackburnyeah..22:07
@sonney2kblackburn, not so easy because e.g. SVRLight derives from SVMLight22:08
blackburnheh22:09
@sonney2kso I would rather suggest to remove apply() from CMachine22:09
@sonney2kand then introduce some type specific apply to the machines we have22:10
blackburnsonney2k: there is the diff?22:10
blackburnwhere*22:10
blackburnsonney2k: SVR/SVM can be solved I think - we can merge regression and binary labels22:11
@sonney2kblackburn, which name do you prefer: DenseOutputLabels, RealvaluedLabels, ...22:15
blackburnsonney2k: dense?22:16
blackburnno not dense22:16
blackburnRealLabels is ok I think22:16
blackburnsonney2k: we should think about multiple outputs here..22:16
@sonney2kblackburn, I currently have regression labels - which would be the same as RealLabels22:17
@sonney2kI think I would keep them though22:17
blackburnsonney2k: yes, why?22:17
@sonney2kbecause it is obvious that you want regression labels for regression, mc labels of mc, binary for binary classifiers...22:18
blackburntrue22:18
blackburnbut I am ok with real labels and multiclass labels22:19
blackburnnot binary, regression and mc I mean22:19
@sonney2kblackburn, we need binary... I mean otherwise there is no way to check if labels are just +1/-122:19
blackburnsonney2k: there is an advantage of real as binary - we can train on outputs of machine22:20
@sonney2kblackburn, I am talking about inputs22:20
@sonney2koutputs will only be RealLabels22:21
blackburnoh different classes?22:21
blackburninput/output?22:21
@sonney2kyes of course22:23
@sonney2kfor binary we really prefer real valued outputs22:24
@sonney2kwe can add a constructor to binary labels that converts from real valued ones though22:24
@sonney2ksplitting at default threshold 022:25
@sonney2kbtw for multiclass I would also want to have some function that returns the outputs of all the sub-machines22:26
@sonney2kreal-valued...22:26
blackburnyeap22:27
blackburnsonney2k: I have used some OvR ROC/PRC/F1/Recall/Precision measures - do you think it should be added to evals too?22:27
@sonney2kblackburn, for multi-output labels we should have a separate classes ...22:28
@sonney2ksure22:28
@sonney2kblackburn, btw what was the issue with auROC?22:28
blackburnsonney2k: hmm?22:29
@sonney2kblackburn, you had auROC=1.0 but accuracy <122:29
blackburnsonney2k: it is possible22:29
@sonney2khow?22:30
@sonney2kahh22:30
@sonney2kbia22:30
@sonney2ks22:30
@sonney2kso no bug?22:30
blackburnsonney2k: hmm no idea but it is22:30
blackburnno no bug I compared with other impl22:30
blackburnof you I guess?22:30
blackburnin mldata utils22:30
@sonney2kblackburn, yes22:30
blackburnsonney2k: what is bias here?22:31
@sonney2ksimple example is add + 1000 to all classifications22:31
@sonney2kauROC will stay correct22:31
@sonney2k100%22:31
@sonney2kaccuracy will drop22:31
blackburnhmm why?22:31
@sonney2kto 0 I guess22:31
blackburnI admit I do not fully understand role of ROC22:31
@sonney2kbecause sign(output+100000) -> +1 for everything22:31
blackburnokay so ROC is TPR/FPR22:32
@sonney2kfor all possible bias terms22:32
@sonney2kso independent of bias22:32
@sonney2kauROC22:32
blackburnsonney2k: ah so neither TPR nor FPR is not changed, right?22:33
blackburnsonney2k: will the graph change at all?22:34
@sonney2kno22:35
@sonney2ksame thing22:35
blackburnsonney2k: sounds pretty useless22:35
@sonney2kwhat sounds useless?22:37
blackburnsonney2k: roc22:37
@sonney2kwhy?22:37
@sonney2kimho accuracy is useless22:37
@sonney2kbecause it depends on bias22:37
blackburnsonney2k: quality is pretty bad but auROC is ok..22:37
@sonney2kquality is actually superb22:38
@sonney2kyou should never trust the bias a machine outputs22:38
@sonney2kbut adjust it to what you want later22:39
@sonney2k(wanted fpr or tpr)22:39
blackburnsonney2k: so only b term is to be adjusted then?22:39
blackburnsonney2k: cool wisdom hehe22:40
@sonney2kyes22:40
blackburnsonney2k: my ROC/PRC analysis led to some interesting results in my multiclass task22:42
blackburnI found that one class really sucks hehe22:42
blackburnhttp://dl.dropbox.com/u/10139213/share/00000_00001.png22:44
blackburnsonney2k: what is the sign? :D :D22:44
@sonney2kno cows :D22:45
blackburnsonney2k: traffic light ahead22:45
blackburnI have absolutely no idea how to recognize something what I can't recognize w/o prior :D22:46
@sonney2kyeah no...22:48
@sonney2ktoo tough22:48
@sonney2kblackburn, CSupervisedKernelLocalTangentSpaceAlignment - does it need binary or MC features?23:09
@sonney2kerr labels23:09
blackburnsonney2k: it needs to be removed23:09
@sonney2k?23:09
blackburnsonney2k: *really* useless23:09
blackburnand makes no sense at all23:10
@sonney2kthat doesn't answer the question23:10
@sonney2kbinary or mc?23:10
blackburnsonney2k: mc but I will remove it anyway23:10
blackburnsonney2k: let me better remove it now23:10
blackburnsonney2k: I mean dim reducition is generally crappy23:11
blackburnbut it is really really useless shit23:11
blackburnI just wanted to remove it once I am done with libedrt things23:12
blackburndone23:13
CIA-113shogun: Sergey Lisitsyn master * rf1d5d22 / (4 files in 2 dirs): Removed SKLTSA algorithm - http://git.io/k40L9g23:13
blackburnin the end I believe it was mistake to add it :D23:14
blackburnsame effect can be obtained with custom distance - but it is still totally useless to do that in that tangent mean23:15
blackburnwhere are students?23:18
wikingwhere23:34
blackburnheh23:34
shogun-buildbotbuild #836 of python_static is complete: Failure [failed test test_1]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/python_static/builds/836  blamelist: blackburn91@gmail.com23:39
blackburnno space left? cool23:41
--- Log closed Fri May 18 00:00:40 2012
« Wednesday, 2012-05-16 Index Friday, 2012-05-18 »

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