Open in new window / Try shogun cloud
--- Log opened Sat Apr 14 00:00:17 2012
--- Day changed Sat Apr 14 2012
wikingi'm gonna test this one00:00
blackburnwiking: I still did not get how can that be you have no features for test data00:00
blackburnwhat then you have at all00:00
wikingwell i do have00:01
wikinglet's say i have 2 different features about an image00:01
wiking1 feature is based on the image00:01
wiking2 feature is based on the text around the image00:01
blackburntext around the image?00:01
wikingmost of the cases i have 1+200:01
wikingyes since the text is embedded into a text00:02
wikingand the text is for sure not completely unrelated to the image... just in very few cases it is00:02
wikingmost the time i have 1+2 features00:02
wikingbut sometimes00:02
wikingi only have feature 100:02
wikingas i don't have the 'context' of the iamge00:02
wikingjust the image itself00:03
blackburnI don't think it is DA however00:03
blackburnwiking: DA is all about different distributions of test and train data00:09
wikingwell according the slide u've sent00:11
wikingi've got that differently00:11
wiking suuuuugar00:14
blackburnwiking: which slide?00:16
blackburnwiking: yes, there is underlying assumption that target probability space is slightly different00:18
blackburnwiking: I have seen one fantastic paper00:19
blackburnthey learned svm recognizing pedestrians on *game*00:19
blackburnand used domain adaptation to construct real system00:19
blackburnthat is the case00:19
blackburnin your case I am unsure how can it help00:20
blackburnI mean if you have no features it is a latent model :)00:20
wikingwell my space is slighhtly different00:20
blackburnso latent DA? nice00:20
wikingi mean u didn't get it then00:20
wikingso you get the 2 different features right?00:20
wikingso i use that and train that SVM00:21
blackburnon train data?00:21
wikinga multiclass svm00:21
wikingtrain data has both features00:21
wikingfeatures from image and features from text00:21
blackburnand test has only text often?00:21
wikingi mix them in a way00:21
wikingtest has soometimes00:22
wikingonly image00:22
wikingso that's where i see DA could do the trick for me00:22
wikingwell i train a classifier with the both of the features00:23
wikingand then use that00:23
wikingin DA where i only have 1 type of feat00:24
wikingand yes i can make it both having train/test split for 1 type of features00:25
blackburnyou didn't convince me :D00:26
blackburnwiking: I still think that you should rather do some kernel mean matching or so00:29
wikingseems java as a language itself is going to be copyrighted :O00:32
blackburnwhat do you mean?00:33
wiking"With the Oracle v. Google trial date set for next Monday, the Judge has asked Google and Oracle to take a position on whether a programming language is copyrightable."00:33
blackburnoh i hate oracle00:34
blackburntheir databases are powerful crap heh00:35
blackburnand their intentions are totally evil00:35
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]01:12
-!- harshit_ [~harshit@] has quit [Ping timeout: 250 seconds]01:23
-!- harshit_ [~harshit@] has joined #shogun01:30
-!- harshit_ [~harshit@] has quit [Remote host closed the connection]02:05
-!- vikram360 [~vikram360@] has quit [Read error: Connection reset by peer]05:26
-!- vikram360 [~vikram360@] has joined #shogun05:26
-!- gsomix [~gsomix@] has joined #shogun08:56
-!- gsomix [~gsomix@] has quit [Client Quit]08:57
-!- gsomix [~gsomix@] has joined #shogun08:57
-!- n4nd0 [] has joined #shogun10:58
-!- wiking [~wiking@huwico/staff/wiking] has quit [Remote host closed the connection]11:26
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun11:26
wikingyeeeey there's a domain transfer gsoc application!11:42
-!- xiangwang [~xiangwang@2001:cc0:2020:2013:f2de:f1ff:fe50:9012] has joined #shogun12:17
-!- n4nd0 [] has quit [Ping timeout: 264 seconds]12:25
-!- blackburn1 [~qdrgsm@] has joined #shogun12:25
CIA-64shogun: Sergey Lisitsyn master * r56eb7ca / (14 files in 6 dirs): Merge branch 'lars' of git:// (+16 more commits...) -
gsomixblackburn1, yo12:30
wikingblackburn1: hey... have u seen the gsoc email about domain transfer?12:33
blackburn1wiking: domain transfer? wtf ;)12:34
wikingok domain adaptation12:35
blackburn1some mohamed's mail you mean?12:35
blackburn1well yes should I worry? :)12:35
shogun-buildbotbuild #714 of libshogun is complete: Failure [failed compile]  Build details are at  blamelist: pluskid@gmail.com12:35
wikingworry why?12:35
wikingit's just great that somebody wants to do it as gsoc12:35
blackburn1huh! there are some more12:36
blackburn1worry because he kinda competes with me12:36
wikingwhy, i thought that this is a team work12:36
blackburn1for slot12:37
wikingah u want to do gsoc as well this year?12:37
wikingi didn't know12:37
blackburn1surprise :D12:37
wikingi thought you are mentoring this year12:37
blackburn1no, I am not an expert in anything :)12:38
blackburn1and undergrad student still12:38
wikinghahahahahah well in the code of shogun you are12:38
shogun-buildbotbuild #715 of libshogun is complete: Success [build successful]  Build details are at
wikingoooh damn i dont understand pgm ppm formaaat12:39
-!- blackburn1 [~qdrgsm@] has quit [Ping timeout: 246 seconds]12:42
shogun-buildbotbuild #652 of r_static is complete: Failure [failed test_1]  Build details are at  blamelist: pluskid@gmail.com12:43
-!- nickon [] has joined #shogun12:45
shogun-buildbotbuild #674 of cmdline_static is complete: Failure [failed test_1]  Build details are at  blamelist: pluskid@gmail.com12:47
wikingsomething got broken12:47
shogun-buildbotbuild #653 of octave_static is complete: Failure [failed compile]  Build details are at  blamelist: pluskid@gmail.com12:47
shogun-buildbotbuild #640 of python_static is complete: Failure [failed compile]  Build details are at  blamelist: pluskid@gmail.com12:48
shogun-buildbotbuild #653 of r_static is complete: Success [build successful]  Build details are at
-!- blackburn1 [~qdrgsm@] has joined #shogun12:53
blackburn1wiking: what is the problem with pgm?12:54
wikingblackburn1: nada... found out what was the problem... in the header i've got some comments but managed to remove them12:54
wikingbut we have a problem with the new merge12:54
blackburn1yes but it may be related to order of commits12:55
blackburn1let us wait a little12:55
blackburn1wiking: do you read pgm manually?12:56
wikingfscanf failed12:56
wikingfor reading the header12:56
blackburn1wiking: what do you use then?12:56
wikingi thought it's always 4 integers are coming in a given order12:56
wikingso i did an fscanf on the header of the pgm12:56
wikingbut then some images that i used to convert with imagemagic12:57
wikinghad some comments in them12:57
wikingfrom exif or i have no idea from12:57
shogun-buildbotbuild #675 of cmdline_static is complete: Success [build successful]  Build details are at
wikinganyhow imagemagick left those comments in the header12:57
blackburn1I mean python's imload have no problem with comments12:57
wikinghehehe i have my own utility ... it's a pretty lame feature extractor12:58
wikingbut still i want to see how it works12:58
wikingi saw new io classes12:58
wikingwell anyhow i have to get back somehow to be able to feed gmm with my custom created gmms... so i can start clustering13:00
wikingdamn my shell script is stupid should do something better about it13:01
CIA-64shogun: Victor Sadkov master * rda846bd / (2 files): Refactored to satisfy code formatting standards -
CIA-64shogun: Sergey Lisitsyn master * r416606c / (2 files): Merge branch 'refactoring' of git:// -
shogun-buildbotbuild #654 of octave_static is complete: Success [build successful]  Build details are at
wikinglol i've just read about new way of stealing a car: they put on your parking car of the rear window in the middle a paper, so when you sit in your car and you start the engine and start to get out from the parking slot you'll realize that there's a car at the rear window, you get out (of course leave everything in the car and the engine running) to remove the paper and at this moment somebody will jump into your car and run away with your car and all yo13:07
wikingbelongings in it ;)13:07
wikingsmart :)13:07
shogun-buildbotbuild #641 of python_static is complete: Success [build successful]  Build details are at
blackburn1wiking: heh13:07
wikingblackburn1: it's cool no? :)13:08
blackburn1wiking: I didn't get what exactly they put?13:09
wikinga piece of paper in the middle of the rear window13:09
wikingjust that you get distracted13:09
blackburn1I see13:09
wikingso that you'd get out from the car to remove it13:09
shogun-buildbotbuild #500 of octave_modular is complete: Failure [failed compile]  Build details are at  blamelist: pluskid@gmail.com13:11
shogun-buildbotbuild #488 of csharp_modular is complete: Failure [failed compile]  Build details are at  blamelist: pluskid@gmail.com13:12
shogun-buildbotbuild #477 of lua_modular is complete: Failure [failed test_1]  Build details are at  blamelist: pluskid@gmail.com13:18
-!- xiangwang [~xiangwang@2001:cc0:2020:2013:f2de:f1ff:fe50:9012] has quit [Quit: Leaving]13:20
-!- n4nd0 [] has joined #shogun13:24
shogun-buildbotbuild #475 of python_modular is complete: Failure [failed test_1]  Build details are at  blamelist:, blackburn91@gmail.com13:35
blackburn1uh that takes a while13:37
-!- karlnapf [] has joined #shogun13:37
blackburn1n4nd0: could you please remind me what I promised to do? :D13:38
n4nd0blackburn1: hmmmm13:40
n4nd0I think I have it on the tip of my tongue13:41
n4nd0blackburn1: no way, I don't remember :(13:42
gsomixn4nd0, hi. how are you?13:45
n4nd0gsomix: hey! I am fine, what about you?13:46
gsomixn4nd0, I'm lazy, hehe. what about covertree?13:46
n4nd0gsomix: it's going good :)13:47
n4nd0I hope I will have something ready to push during this weekend13:47
n4nd0gsomix: what about you, what are you working on?13:49
gsomixn4nd0, reference counting for SGVector. but I had some problems with time over the past days.13:51
n4nd0gsomix: aham, good luck with it anyway ;)13:52
blackburn1n4nd0: ah consider merging your ovo thing13:53
n4nd0blackburn1: yeah, but first I should test if it now it works fine13:54
n4nd0blackburn1: don't you think so?13:54
blackburn1n4nd0: why couldn't it work?13:54
blackburn1karlnapf: hey!13:54
karlnapfblackburn1, hi :)13:55
karlnapfheard there are 9 slots :D13:55
n4nd0blackburn1: who knows13:55
blackburn1karlnapf: yeap13:55
blackburn1karlnapf: could you please add subsetting to labels as well? ;)13:55
karlnapfpretty cool13:55
karlnapfblackburn1, yes its on my todo list13:55
karlnapfsorry for being away last week, too busy13:56
blackburn1ok nice thanks13:56
-!- xiangwang [~chatzilla@] has joined #shogun13:58
-!- xiangwang [~chatzilla@] has quit [Client Quit]13:59
blackburn1n4nd0: I guess that is I promised to do14:00
blackburn1karlnapf: what is your next weeks workload?14:01
karlnapfblackburn1, I just handed in my last coursework, so I will have a bit more time. However, there are still the exams in May, which are really important and therefore currently have priority14:03
blackburn1karlnapf: I stumbled upon maximum mean discrepancy embedding and will ask you for some advice later :)14:04
karlnapfbut I think I will be able to do some shogun stuff in between14:04
karlnapfreally? where?14:04
karlnapfThis is really cool stuff (-)14:04
blackburn1karlnapf: MMD or MMDE?14:04
blackburn1I assume you are MMD expert14:04
blackburn1MMDE is some kind of multitask stuff14:05
karlnapfmaximum mean discrepancy embedding?14:05
karlnapfnever heard it, what happens there14:05
blackburn1karlnapf: LLE(MVU) + MMD14:05
blackburn1Locally Linear Embedding14:06
blackburn1okay idea is to use MMD as distance if I get it right14:06
blackburn1and construct embedding14:06
karlnapfMMD is a distance between distributions14:06
blackburn1as latent space for united features space for domain14:06
blackburn1karlnapf: can't it be empirical?14:07
karlnapfyes, of course, you dont know the distributions14:07
karlnapfok, sounds cool, do you have a paper?14:07
karlnapflooks interesting14:09
karlnapfcool I am always looking for applciations of MMD14:09
karlnapfthe stuff I am doing will be kind of a large scale online two sample test, so a bit more statistics than ML14:10
blackburn1I see14:10
blackburn1so how can you apply it?14:10
karlnapfwhat do you mean?14:11
karlnapflike which applications?14:11
blackburn1karlnapf: yes14:12
karlnapfthe current MMD test ist in quadratic time. However, imagine you have infinite data and want the best test possible for a given amount of time14:12
karlnapfthen you do a kind of online-like kernel selection on a stream of data14:12
karlnapfeither you optimise kernel until test says distributions are different14:13
blackburn1but which distributions do you compare?14:13
karlnapfor until you can say, probably they are equal14:13
karlnapfdata comes from two distributions14:13
karlnapfgoal is a two sample test14:13
karlnapftwo sources of microarray data14:13
karlnapftest is cancer type is equal or not14:14
karlnapfIll tell you more about it later :)14:14
karlnapfcurrently in the early stages14:14
karlnapfoh and thats my master project14:14
karlnapfthe gsoc stuff is a bot different14:14
karlnapfbut anyway, I will study a bit now :) see you around!14:15
blackburn1ok looks clearer now14:15
blackburn1thanks for explanation :)14:15
blackburn1see you14:15
shogun-buildbotbuild #501 of octave_modular is complete: Success [build successful]  Build details are at
blackburn1wiking: ^ all ok :)14:16
shogun-buildbotbuild #489 of csharp_modular is complete: Success [build successful]  Build details are at
shogun-buildbotbuild #478 of lua_modular is complete: Success [build successful]  Build details are at
shogun-buildbotbuild #470 of ruby_modular is complete: Failure [failed test_1]  Build details are at  blamelist: sadkov993@gmail.com14:35
shogun-buildbotbuild #471 of ruby_modular is complete: Success [build successful]  Build details are at
-!- vikram360 [~vikram360@] has quit [Ping timeout: 265 seconds]15:31
-!- vikram360 [~vikram360@] has joined #shogun15:32
-!- blackburn1 [~qdrgsm@] has quit [Ping timeout: 276 seconds]15:59
-!- harshit_ [~harshit@] has joined #shogun16:27
-!- harshit_ [~harshit@] has quit [Ping timeout: 265 seconds]16:35
-!- pluskid [~pluskid@] has joined #shogun16:36
-!- harshit_ [~harshit@] has joined #shogun16:43
gsomixsonney2k, hey.16:57
-!- harshit_ [~harshit@] has quit [Ping timeout: 250 seconds]17:04
-!- nickon [] has quit [Quit: ( :: NoNameScript 4.22 :: )]17:17
-!- harshit_ [~harshit@] has joined #shogun17:17
-!- pluskid [~pluskid@] has quit [Ping timeout: 252 seconds]17:32
-!- pluskid [~pluskid@] has joined #shogun17:33
-!- wiking_ [] has joined #shogun17:44
-!- wiking_ [] has quit [Changing host]17:44
-!- wiking_ [~wiking@huwico/staff/wiking] has joined #shogun17:44
-!- wiking [~wiking@huwico/staff/wiking] has quit [Ping timeout: 244 seconds]17:48
-!- wiking_ is now known as wiking17:48
-!- pluskid [~pluskid@] has quit [Quit: Leaving]17:58
-!- karlnapf [] has left #shogun []18:06
-!- harshit_ [~harshit@] has quit [Quit: Leaving]18:33
-!- vikram360 [~vikram360@] has quit [Ping timeout: 276 seconds]19:28
-!- vikram360 [~vikram360@] has joined #shogun19:28
-!- wiking [~wiking@huwico/staff/wiking] has quit [Remote host closed the connection]19:29
-!- wiking [] has joined #shogun19:29
-!- wiking [] has quit [Changing host]19:29
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun19:29
-!- n4nd0 [] has quit [Quit: leaving]19:39
-!- blackburn [~qdrgsm@] has joined #shogun19:49
-!- harshit_ [~harshit@] has joined #shogun19:49
@sonney2kgsomix, hey - how is it going?19:50
gsomixsonney2k, moin. I added code to ref count in the SGVector. and I did SGVector -> SGVector& for arguments to some functions and methods.19:56
gsomixbut I have a question. can you look at code preprocessors?19:57
gsomixmethods `apply_to_feature_vector'19:58
gsomixsonney2k, this method should return a reference to new SGVector on heap with SG_REF, isn't it?20:03
gsomixor should leave as is?20:03
@sonney2kgsomix, uh you are right20:07
@sonney2kI am not sure if this reference (on the heap) is still valid after the function returns20:07
blackburnhey there20:08
gsomixblackburn, yo20:09
blackburnsonney2k: I have to ask you to review feature learning code20:09
blackburnwould be pretty unfair :)20:09
blackburnsonney2k: and btw I merged lars20:10
wikingi have only a mean vector and the variance vector of a gaussian... how do i create a CGaussian from that? :)20:11
-!- Marty28 [~Marty@] has joined #shogun20:11
blackburnwiking: what is the problem?20:11
wikingwell the covariance is a matrix...20:11
wikingCGaussian(SGVector<float64_t> mean, SGMatrix<float64_t> cov, ECovType cov_type=FULL);20:12
blackburnuse DIAG or so20:12
blackburnor patch it :D20:12
wikingyou mean that in case of a DIAG matrix where the diag elements are actually the variance that should be fine for creating a gaussian?20:12
blackburnwiking: yes some kind of20:13
blackburn3 0 020:13
blackburn0 1 020:13
blackburn0 0 320:13
blackburnwould work :)20:13
wikingand 3,1,3 would be the variance?20:13
wikingmmm is there a function to create a diag matrix where with given diagonal elements, i.e. from a vector?20:13
gsomixsonney2k, why not? or do you mean that problem is that someone should to free returned SGVector?20:14
blackburnwiking: where?20:14
wikingor simply create an eye matrix20:14
wikingmmm in shogun20:14
blackburnwhat is the inteface?20:14
@sonney2kgsomix, because the on the heap allocated object might be freed already...20:14
wikingi'm asking whether there is one :)20:15
blackburnwiking: there is one in cmath (create_diagonal_matrix) however if you use python it would be easier to use numpy for that20:15
wikingso like diag (v) in matlab20:15
gsomixsonney2k, so what is strategy to convert method's returns (SGVector)? references only in the right places?20:22
gsomixwe can not return references to local variables or references to allocated SGVector in methods. should we return copy?20:24
gsomixsonney2k, ok. I figured out how it should work. :)20:49
shogun-buildbotbuild #209 of nightly_none is complete: Failure [failed compile]  Build details are at
blackburnuh how bad bad bad21:04
-!- vikram360 [~vikram360@] has quit [Read error: Connection reset by peer]21:06
-!- vikram360 [~vikram360@] has joined #shogun21:06
@sonney2kgsomix, I actually don't know - so how do you do it?21:15
CIA-64shogun: Sergey Lisitsyn master * r277ba68 / (src/shogun/regression/LARS.cpp src/shogun/regression/LARS.h): Proper constructor for #438 -
gsomixsonney2k, we return copy if we have new instance. simple way.21:17
* gsomix afk21:17
blackburndamn how I missed that :)21:18
blackburnsonney2k: are you ok with transfer/domain_adaptation & transfer/multitask structure?21:22
@sonney2kgsomix, ok21:27
@sonney2kblackburn, too tired now21:27
blackburnsonney2k: see ya21:27
-!- harshit_ [~harshit@] has quit [Quit: Leaving]21:31
-!- vikram360 [~vikram360@] has quit [Read error: Connection reset by peer]21:32
wikingmmm guys21:46
wikingever since i've upgraded to the latest shogun21:47
wikingi'm having problems with combined kernel21:47
blackburnwiking: and what is the problem?21:47
wiking[ERROR] Index out of Range: idx_a=0/0 idx_b=0/021:47
-!- Marty28 [~Marty@] has quit [Quit: ChatZilla [Firefox 11.0/20120310010446]]21:48
blackburnwiking: ok then just fix it as I assume you want to be a shogun developer ;)21:48
wikinghahahahah lol21:49
wikingi wonder where has been the change21:49
blackburnso something changed very recently?21:49
-!- vikram360 [~vikram360@] has joined #shogun21:49
blackburnthen my suggestion is still here - just try to bisect it ;)21:50
blackburnhow long has it been outdated?21:50
wikingso this is my code21:50
wikingthe getKernel function is just giving me the kernel i want21:51
blackburnI bet it is empty21:51
wikingi have 10 bucks that it's my fault21:51
wikingi've commented out the kernel.init ... part21:52
wikinghehe now it works of course21:52
wikinglets see if i'm getting the same error21:52
wikingi mean21:53
wikingerror rate :)21:53
wikingas the homo kernel mapper21:53
wikingeven better ;)21:54
-!- genix [~gsomix@] has joined #shogun21:54
wikingblackburn: got a sec?21:56
wikingi've just realized that u could actually do a combined kernel like thingy21:56
wikingwith homo kernel mapping :)21:56
blackburnsome kind of CHI2 + intersection?21:57
wikingjust do the mapping for the given features21:57
wikingso you take the features21:57
wikingand then do the homo mapping21:57
wikingthe desired kernel for the given feature21:57
-!- gsomix [~gsomix@] has quit [Ping timeout: 245 seconds]21:57
wikingand then just concatenate those features21:57
blackburnokay actually I think this should be done with some features union helper21:58
wikinghehe yeah21:58
wikingwell i have my own function for it at the moment21:58
blackburnI'll implement one in a hr21:58
wikingbut i guess what we would like to have is in shogun21:58
wikingbut there's something like this no?21:58
wikingsonney2k: was talking about21:59
wikingthe CombinedDotFeatures21:59
blackburnhmm actually yes21:59
blackburnit makes more sense even21:59
blackburnno need to do that21:59
-!- genix is now known as gsomix21:59
wikingi'll give it a shot now21:59
wikingsee how it'll react22:00
-!- karlnapf [] has joined #shogun22:11
CIA-64shogun: Heiko Strathmann master * r4711307 / src/shogun/classifier/mkl/MKL.cpp :22:22
CIA-64shogun: -added initialisation of mkl_block_norm to std value22:22
CIA-64shogun: -added comment about strange set_sub_kernel_weights22:22
CIA-64shogun: -fixed a mem leak regarding this comment -
CIA-64shogun: Heiko Strathmann master * r2e21eb9 / src/shogun/kernel/CustomKernel.cpp : added debug messages -
CIA-64shogun: Heiko Strathmann master * rbf5e9ea / src/shogun/machine/KernelMachine.cpp : added debug messages -
CIA-64shogun: Heiko Strathmann master * r24bab87 / src/shogun/classifier/mkl/MKL.cpp : SGVector instead of plain array -
CIA-64shogun: Heiko Strathmann master * r2f00592 / (3 files in 3 dirs):22:22
CIA-64shogun: Merge pull request #437 from karlnapf/master22:22
CIA-64shogun: memory bugfix, initialization, comments and debug messages -
CIA-64shogun: Heiko Strathmann master * rd716ae0 / examples/undocumented/libshogun/modelselection_grid_search_mkl.cpp :22:27
CIA-64shogun: added example for mkl modelselection (currently not working, but should22:27
CIA-64shogun: in the future, once bugs are fixed) -
CIA-64shogun: Heiko Strathmann master * rc12e96c / examples/undocumented/libshogun/modelselection_grid_search_mkl.cpp :22:27
CIA-64shogun: Merge pull request #453 from karlnapf/master22:27
CIA-64shogun: added mkl modelselection example -
wikingkarlnapf: the model selection example works?22:37
karlnapfwiking, no, not yet22:37
karlnapfsome strange stuff22:37
blackburnlong way :)22:38
karlnapfLuckily I could tell this guy from cambridge how to use an alternative22:38
karlnapf(not using SVMLight)22:39
blackburnwhich guy?22:39
wikingkarlnapf: i'm just trying to experiment with mkl22:39
karlnapfhe wrote on the mailing list a couple of times about mkl stuff22:39
wikingbut i'm always getting worse results than simply doing a combinedfeatures/kernel22:39
karlnapfwiking, what are you trying to do?22:39
wiking+ gmnpsvm22:40
karlnapfI mean for simple examples, MKL should be a bit worse than standard stuff since when optimising, there are a lot of approximations happening.22:40
karlnapfbut I am not an MKL expert, soeren is22:40
blackburnsoeren hates it22:40
wikingyeah soeren dislikes it ;)22:41
karlnapfwhy did he write so many papers then ;)22:41
wikinggood question D:22:41
blackburnjust for fun may be :D22:41
karlnapfwiking, is the result much worse?22:41
wikingkarlnapf: well we are talking about 7-8% accuracy22:41
wikingso 80 vs 87-8822:42
karlnapffor binary classification?22:42
karlnapfmmh thats a lot22:42
wikingmulti class sorry22:42
wikingyeah i thought it should be actually better22:42
wikingbut it's not :(22:43
karlnapfmmh, dont know, there are so many possible sources for that22:43
karlnapfMKL is usually not better than just trying combinations by hand22:43
karlnapfonly faster22:43
wikingmmm not faster either ;)22:43
karlnapfhehe ;)22:44
karlnapfWell to be honest, I dont know, havent digged into that stuff.22:44
blackburnkarlnapf: ah one more issue22:44
karlnapfI am currently more interested to get rid of memory bugs in MKL with SVMLight :)22:44
karlnapfthis is ugly22:44
blackburnI'd like to enable cross validation for dim reduction things :)22:44
karlnapfpeople want to use model-selection and it doesnt work22:44
blackburnI'll ask you later22:45
karlnapfblackburn, yes nice22:45
karlnapfI guess you would want to modify/add stuff on the ModelSelection class22:45
karlnapfCurrently, a CMachine interface is needed22:45
-!- vikram360 [~vikram360@] has quit [Read error: Connection reset by peer]22:47
blackburnkarlnapf: yeah however that would be really painful change22:50
karlnapfblackburn, I think it would be better to extend the modelselection class22:50
karlnapfhowever, you would need some shared interface between all classes that do want to do some parameter selection22:51
blackburnkarlnapf: oh I feel too much pain thinking about it :D22:53
wikingwhat's with svm/ScatterSVM.h ?22:53
blackburnwiking: LOLZ22:53
karlnapflol ;)22:53
wikingwhat's that thing? :)))22:53
wikingwith being experimental :D22:53
blackburnyou're really going to try EACH feature of shogun22:53
wikingi'm just checking the svm machines22:54
wikingand i've just stumbled upon that stuff :))22:54
blackburnwiking: okay actually it is experimental and definitely not better than larank/gmnp22:54
blackburnwiking: I really believe you should address your issues not with trying different machines ;)22:55
blackburnwiking: I mean your accuracy problem could be solved by analysis of results22:56
wikingonce again22:56
karlnapfblackburn, I am thinking of moving the subset logic (stack etc) from Features/Labels to Subset class. However, that would break all objects that have a subset from older shogun version (migration can do nothing here)22:57
blackburnkarlnapf: do you understand my words? :)22:57
karlnapfwhat do you think?22:57
blackburnkarlnapf: why?22:57
karlnapfblackburn, because then the logic would be in one class22:57
karlnapfinstead of having it multiple times in different classes22:58
karlnapfBefore these stacks, everythig was easy22:58
-!- gsomix [~gsomix@] has quit [Ping timeout: 265 seconds]22:58
-!- harshit_ [~harshit@] has joined #shogun22:58
karlnapfbut now code gets more complicated and you dont want to have a similar copy of the recomputation of the subset stack in every class which uses it22:58
blackburnI can hardly imagine that logic in stacks though22:58
karlnapfits a pain anyway because normally this should be done via multiple inheritance, but we cannot use this, so each class has to implement wrapper classes anyway22:59
-!- vikram360 [~vikram360@] has joined #shogun22:59
karlnapfjust the methods add/remove22:59
blackburnI think Soeren would not like breaking migrations22:59
karlnapfthe current active subset has to be recomputed in it22:59
karlnapfthis stuff is all not migratable anyway23:00
karlnapfi dont know how to translate the current changes .....23:00
blackburnactually my opinion is that users can do migrations manually :)23:00
karlnapfin this case, i agree23:01
karlnapfin other cases, we should help them23:01
karlnapflike name changes /type changes/new variables etc23:01
karlnapfsonney2k, are you around? whats your opinion?23:03
blackburnhe is sleeping I guess23:03
blackburnor I don't know23:03
blackburnactually he tends to appear randomly23:03
karlnapfsame here :)23:04
karlnapf(not the sleeping thing)23:04
blackburnkarlnapf: what is your opinion on moving lib* -> shogun/lib23:05
karlnapffrom machine to lib23:06
karlnapfdont know23:06
blackburnfrom classifier23:06
karlnapfwhy not?23:06
karlnapfbecause its an external library?23:06
karlnapfI was a bit confused about that fact actually when I tried using shogun for the first time23:06
karlnapfon the other hand23:06
karlnapfall classifier are at one place now23:07
karlnapfyou would loose that23:07
blackburnkarlnapf: actually not23:07
blackburnkarlnapf: multiclass is already forked23:07
blackburnfrom classifier23:07
blackburnand both are using libocas and liblinear23:07
blackburnok I'll do pull request and ask you and soeren for comments23:08
blackburnone more thing I don't like is liblinear is called SVM_linear23:08
karlnapf(both things)23:10
karlnapfI will rewrite the subset stuff now23:10
karlnapfwill be much easier23:10
karlnapfand I just thought of a way for the migrations23:10
blackburnsunday surprise for soeren23:10
blackburnkarlnapf: do you like shogun/lib/external23:11
blackburnfor that?23:11
karlnapfyes, that would actually be nice23:11
karlnapfblackburn, do we got a list type for SGVectors?23:19
blackburnlist of sgvectors?23:19
blackburnno idea23:19
karlnapfmmh I will just use DynArray then23:19
-!- n4nd0 [] has joined #shogun23:21
-!- vikram360 [~vikram360@] has quit [Ping timeout: 265 seconds]23:27
-!- vikram360 [~vikram360@] has joined #shogun23:27
--- Log closed Sun Apr 15 00:00:19 2012