Open in new window / Try shogun cloud
--- Log opened Fri Feb 24 00:00:19 2012
-!- n4nd0 [] has quit [Read error: Connection reset by peer]00:03
-!- n4nd0 [] has joined #shogun00:03
-!- n4nd0 [] has quit [Read error: Connection reset by peer]01:02
-!- n4nd0 [] has joined #shogun01:02
-!- n4nd0 [] has quit [Ping timeout: 244 seconds]01:55
-!- axitkhurana [~Akshit@] has joined #shogun02:26
-!- axitkhurana [~Akshit@] has left #shogun []03:37
-!- axitkhurana [~akshit@] has joined #shogun04:13
-!- abhilb [75ce3638@gateway/web/freenode/ip.] has joined #shogun04:34
-!- axitkhurana [~akshit@] has quit [Remote host closed the connection]05:16
-!- blackburn [~qdrgsm@] has joined #shogun07:21
-!- n4nd0 [] has joined #shogun07:34
-!- blackburn [~qdrgsm@] has quit [Ping timeout: 260 seconds]07:58
-!- abhilb [75ce3638@gateway/web/freenode/ip.] has quit [Quit: Page closed]09:15
-!- n4nd0 [] has quit [Ping timeout: 240 seconds]14:12
-!- n4nd0 [] has joined #shogun15:11
-!- tjn4 [458618b7@gateway/web/freenode/ip.] has joined #shogun15:42
-!- tjn4 [458618b7@gateway/web/freenode/ip.] has quit [Client Quit]15:42
-!- n4nd0 [] has quit [Quit: leaving]16:31
-!- blackburn [~qdrgsm@] has joined #shogun18:14
-!- karlnapf [] has joined #shogun19:35
karlnapfsonney2k, around?19:39
karlnapfblackburn, around?19:39
karlnapfhow is it going?19:40
blackburnfine, what about you?19:40
karlnapfsame here :)19:41
karlnapfvery busy19:41
blackburnyeah, feel the same sometimes19:42
blackburnkarlnapf: yesterday we have discussed ideas with soeren19:42
karlnapfah nice19:42
blackburnyou probably better to check logs19:42
karlnapfwas about asking about that19:43
karlnapfwill do19:43
karlnapfmany new things?19:43
blackburnno, no new things at all I'd say19:43
blackburnI am going to do multitask learning, you know?19:43
karlnapfwow, nice one :)19:44
karlnapfI had some lectures by massi pontil about that19:44
blackburnI don't know much about it yet19:45
karlnapfme neither, just the lecture, havent studied it so well yet, but its probably a nice thing and very recent19:45
blackburnI checked your part and it is pretty complex as well19:46
karlnapfthe mmd/HSIC?19:46
blackburnyeah no idea what it means but probably ys19:47
karlnapfthese are really nice things in my eyes19:47
karlnapfI hope I will do my Master project also about this19:47
karlnapfAbout kernel selection for these things19:48
blackburnmaster? I though you are phd..19:48
blackburnI'm going to do my bachelor project about road sign recognition19:48
karlnapfno, I am a Master student, hopefully phd in autumn :)19:49
blackburnah last year masters?19:49
karlnapfnice, so lots of image recognition stuff19:49
blackburnI see19:49
karlnapfits a one year program here19:49
blackburnI see19:49
blackburnhmm something is wrong with svmocas19:50
blackburnjust got 16% accuracy19:50
karlnapfthats not good19:50
blackburnwhile liblinear was 96%19:50
karlnapfI have beed fiddeling around with multiclass svm recently19:50
karlnapfmulticlassMKL also19:50
blackburnyeah I'm doing multiclass as well19:50
karlnapfand I think shogun desparetely needs sigmoid fitting (also for multiclass)19:51
karlnapfoh ok.19:51
blackburnwhy not to use genuine multiclass svms?19:51
karlnapfI have a lot of changes in my local repo, will commit when I find some time19:51
karlnapfwhats that?19:51
blackburnlike gmnpsvm19:51
blackburnno need to do sigmoid fitting19:51
blackburntakes a lot of time to train though19:51
karlnapfwell, but users may want to have probs for standard svm19:52
karlnapfreally? Just wrote some code for it19:52
blackburnI thought you mean different thing19:52
karlnapfworks pretty fine19:52
karlnapfah ok19:52
karlnapfI also implemented it for one-against-one svms19:52
karlnapfResults are nice19:52
blackburnone against one?19:52
karlnapfWhen we discussed about how to integrate it into shogun I will do19:52
karlnapfyes for multiclass svm19:52
blackburnI don't like OvO :)19:53
blackburnhad 903 svms here19:53
karlnapfone against one currently only has this vote stuff19:53
karlnapfoh :)19:53
karlnapfwell thats probably not feasible19:53
karlnapfbut if its used, its nice to have prob as output19:53
blackburnlet me describe what I meant19:53
karlnapfgo ahead :)19:54
blackburnwhen you do some OvR or OvO19:54
blackburnscaling can vary much19:54
blackburnand it is not fitted very well19:54
blackburnso both vojtech and soeren told me it depends on such scaling19:54
blackburnthat's why gmnpsvm and larank works better19:54
karlnapfah ok19:55
karlnapfhow does it work, whats the output/input?19:55
blackburnwhat is 'it'?19:55
blackburnboth gmnp and larank is OvR19:55
blackburnbut it trains like genuine multiclass svm19:55
karlnapfmmh ok19:56
karlnapfI have got to do some reading on this stuff19:56
blackburn\sum_k ||w_k|| + C \sum_{i \ne k} \sum_j \xi_i^j19:56
karlnapfI am currently facing a problem with 4 classes19:56
karlnapfI would love to have a multi class MKL machine which output probabilities for this19:56
blackburnI see19:56
blackburndo you find MKL useful?19:57
karlnapfkind of like the idea19:57
karlnapfmy problem is from bioinformatics19:57
karlnapfI have two kernels: one based on biochemical data, and one based only on sequence oinformation19:57
blackburnI was curious if it is applicable to not bio data19:57
karlnapfAnd the kernel weights would give me a kind of measure how important the features are19:57
karlnapfMMhh, dont really know19:57
karlnapfbut I like combining kernels19:57
blackburnthere was one nice guy here in irc,  he wrote HomogeneousKernelMap preprocessor19:58
blackburninteresting stuff19:58
karlnapfcool ,will check it out :)19:58
blackburnmaps features to linear space19:59
blackburnwhere dot product is a kernel19:59
blackburnexplicitly I mean19:59
karlnapfwhich dimension?19:59
blackburnmore -> better19:59
blackburnbut x3 is sufficient20:00
blackburnI lose only 0.5% of accuracy20:00
blackburnbut linear20:00
blackburnmaking me possible to use liblinear with 3m train time20:00
blackburninstead of libsvm with 6h :)20:00
karlnapfliblinear is so fucking fast20:00
blackburnsvmocas is faster20:01
blackburnbut I am afraid of its results now20:01
karlnapfbtw I currently can reproduce a segfault using it (with modelselection) will try to fix that these days20:01
blackburnyou were pretty active last month20:01
karlnapfyes, want to integrate everythign I am using/doing here into shogun20:02
blackburnin fact I think you should have write rights..20:02
karlnapfespecially for model, there is lots of stuff to do20:02
karlnapfwell, I dont care much about this, also, the time I invest varies a lot20:02
blackburnsame here20:02
blackburnI have been doing nothing about development since january I guess20:03
karlnapfThink its ok, its opensource, youre not paid for this :)20:03
karlnapfI think youre doing great stuff here anyway :)20:03
blackburnbut gsoc is paid so going to be active this summer20:03
karlnapfyes, ineed :)20:04
karlnapfblackburn, I have some work to do so will leave now20:04
karlnapfsee you later20:04
blackburnsee yah20:04
@sonney2kkarlnapf, hi - yes20:15
karlnapfsonney2k, hi, I was about leaving now, just wanted to hear about the gsoc stuff, but talked a bit with blackburn and read yesterdays channel log20:16
blackburnsonney2k: gmnpsvm (JS kernel) 97.14%, liblinear (OvR,JS) 96.00%20:17
@sonney2kkarlnapf, I just updated the webpage20:18
karlnapfsonney2k, ok cool, will have a look, btw did you talk to arthur?20:18
blackburnhmm a little different features20:18
@sonney2kblackburn, already model selected?20:18
blackburnwell same C20:19
@sonney2kkarlnapf, no20:19
blackburnsonney2k: the problem - I had 16% with svmocas20:19
blackburnI am afraid it is broken20:19
@sonney2kblackburn, hmmhh I guess we introduced a bug at some point :(20:19
-!- axitkhurana [~akshit@] has joined #shogun20:19
karlnapfsonney2k, ok, have to go now, see you20:19
blackburnI fixed one before20:20
@sonney2kkarlnapf, regarding arthur20:20
@sonney2kis it OK if I put him as mentor?20:20
@sonney2kI think it makes sense...20:20
karlnapfsonney2k, I have asked him, but he is quite busy currently. I think he will agree and it makes sense. Will see him next week and then ask20:20
@sonney2kkarlnapf, yes ICMl deadline today20:21
karlnapfexactly ;)20:21
karlnapfok, bye then20:22
@sonney2kkarlnapf, cu!20:22
blackburnsonney2k: anyway, with gmnp I had 97.32% with HIK and 97.14% with HIK approximation with that homogay kernel map20:23
blackburnso I'd say it is pretty nice thing20:23
@sonney2kbasically all the same20:24
blackburnyes, so it is pretty good approximation20:24
-!- karlnapf [] has quit [Quit: Leaving.]20:24
blackburnsonney2k: I did rejects in very simple way20:24
blackburnjust with thresholding I get 99% accuracy among non-rejected images20:24
blackburnpretty nice for me :)20:25
-!- axitkhurana [~akshit@] has quit [Ping timeout: 260 seconds]21:36
-!- n4nd0 [] has joined #shogun21:38
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]23:08
--- Log closed Sat Feb 25 00:00:19 2012