Open in new window / Try shogun cloud
--- Log opened Tue Aug 28 00:00:17 2012
-!- romi_ [] has joined #shogun03:06
-!- romi_ [] has quit [Remote host closed the connection]04:37
-!- n4nd0 [] has joined #shogun06:47
-!- n4nd0 [] has quit [Quit: leaving]07:32
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.] has joined #shogun12:01
-!- yoo [2eda6d52@gateway/web/freenode/ip.] has joined #shogun14:29
-!- yoo [2eda6d52@gateway/web/freenode/ip.] has quit [Client Quit]14:32
-!- yoo [2eda6d52@gateway/web/freenode/ip.] has joined #shogun14:43
yooI have a question about mkl classification14:44
yoousing p-norm mkl and the example below14:46
yoowith p=1 no problem14:46
yoowith p>1 I get different results according of orders the kernels are "append" to CombinedKernel14:47
yooregardless to the order the kernels are added, the first kernel always get a big weight ...14:48
n4nd0yoo: what classifier are you using?14:57
n4nd0it looks like there are some bugs with MKL and SVMLight14:57
yooI am using MKLMulticlass14:58
yooCould you confirm me the "bug" I get running the code below14:59
yoodespite those results, I dont know if its a real bug14:59
yooIs there another mkl multiclass machine than CMKLMulticlass ?15:01
n4nd0give me a sec15:03
yooMaybe CMKLMulticlass support only 1-norm ?15:05
n4nd0I am not sure about that, I have never used this part of the project but I am checking15:05
n4nd0about other MKL multiclass classes15:05
n4nd0I think there is an MKLMulticlassGradient and MKLMulticlassGLPK15:06
n4nd0did you see those?15:06
n4nd0yoo: in the doc of MKLMulticlass, it says that it is a class for L1-norm - so yes, I guess it is only for L1-norm15:07
yooyes but MKLMulticlassGradient and MKLMulticlassGLPK are helpers for MKLMulticlass15:07
n4nd0yoo: aham, I understand15:07
n4nd0however there are some things related about pnorm begin >= 115:09
n4nd0being, sorry15:09
yooin the MKLMulticlass norm15:10
yoovoid CMKLMulticlass::set_mkl_norm(float64_t norm) { pnorm=norm; if(pnorm<1 ) SG_ERROR("CMKLMulticlass::set_mkl_norm(float64_t norm) : parameter pnorm<1"); }15:10
-!- n4nd0_ [] has joined #shogun15:10
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.] has quit [Quit: Page closed]15:10
-!- n4nd0_ is now known as n4nd015:10
n4nd0yeah ok so the norm shouldn't be less than 115:11
n4nd0but your problem was with norm greater than 1 right?15:11
yooyes of course15:11
n4nd0I think it should work fine with pnorm>=115:12
yooallright could you just run the code I sent you, set p-norm with p=1.5 for example15:13
n4nd0let's see15:13
yooand then change the order of the kernels15:13
yoo"the order you call cker->append_kernel()"15:14
yooand tell me if you get correct results : meaning same weights with different corresponding orders15:14
yoo(thanks :p )15:15
n4nd0there was a compile error btw, line 21 missing a "15:15
yoocertainly when passing code to pastebin15:15
yooyes, because you need to set your path to shogun example data15:16
n4nd0yeah, the weights are different15:16
n4nd0firs time15:16
n4nd0and I just switched two calls to append15:17
yoothats strange15:17
n4nd0I certainly don't know15:17
n4nd0I don't know much about MKL apart from what the acronym means :S15:17
yoobecause linear kernel give poor results but when you place it in the first position, it takes 80% of weights ..15:18
yooI am not very used to it too15:18
yoomaybe I can come back later. Could you keep the attached code and ask to people who wrote the code ?15:19
n4nd0have you checked that the other parts of your example are correct?15:21
n4nd0for example, I am not very sure if I understand why before each call to append_kernel you call append_feature_obj?15:21
yooyes I think this part is correct. append_features is called for each kernel since the features could be different for each kernels.15:38
yoofrom doc "It keeps pointers to the added sub-features and is especially useful to  combine kernels working on different domains (c.f. CCombinedKernel) and to  combine kernels looking at independent features."15:40
n4nd0I see15:45
n4nd0but the thing is that for the kernels you are using, all of them use the same features15:47
n4nd0then I am not really sure why to append the same features several times15:48
yoobecause you need CCombinedFeatures object ?15:48
yooif not : segfault " Number of features/kernels does not match - bailing out "15:49
n4nd0aham ok, you are probably right15:49
yoobut one more time, I am not an expert ..15:49
n4nd0me neither15:49
n4nd0take a look at python_modular example though15:50
yoook i'll do that right now15:50
yoogive me a minute15:50
n4nd0as you can see there the features are not given in the constructor to the individual kernels15:50
n4nd0maybe that doesn't make a difference but let's better be sure15:50
yooand they test p-norm with p =1.5 or 2.015:54
yoobut I can see : feats_train.append_feature_obj(subkfeats_train)15:55
yoo3 time fr each kernel15:55
yoowhat d you mean "he features are not given in the constructor to the individual kernels" ?15:55
yoooh I see15:56
n4nd0I have run the test like that15:56
n4nd0and it seems that the weights still vary15:56
n4nd0if one changes the order of them15:57
n4nd0of the kernels I mean15:57
yooyes me too15:59
yoodid you run it in python or c++ ?15:59
n4nd0have you tried changing them in python?16:00
yoonot yet16:01
n4nd0let me know when you have16:01
yoois there a get_kernel_weights method in python ?16:05
n4nd0yeah sure16:08
n4nd0if you are using it from libshogun interface, it should be available from python modular16:09
n4nd0yoo: did you find any problem trying it?16:09
yoohere is the code if you want to test it16:13
yoochange the path to inputs16:13
n4nd0so what did you find?16:14
yoothe same pb ..16:15
yoowork well only with 1-norm16:15
n4nd0I have just tested it in python16:16
n4nd0changing the other of the kernels introduced16:17
yooand ?16:17
n4nd0using 1.5 norm16:17
n4nd0the weights do not change16:17
n4nd0sometimes there's a small variation in the last weight16:17
n4nd0but it is about e-13 so I guess that is normal16:17
yoobut they should right ? if you change the order16:17
yoovalues should stay the same in different ordrers16:18
n4nd0well I guess they are given in that ascending order16:18
yoono way16:18
yoothey are given in the order you ve called append16:18
yoojust try with 1-norm, you will see16:18
n4nd0lol this is a funny one16:20
n4nd0when I change the norm from 1.5 to 116:20
n4nd0SystemError: [ERROR] glpk.h from GNU glpk not included at compile time necessary here16:20
yoodont get that16:21
yoooutputs are16:21
yoo[ 0.97250202  0.02749798  0.        ]16:21
yoo[ 0.97250202  0. 0.02749798     ]16:21
yooworks fine16:21
n4nd0for which example?16:24
n4nd0norm 1 you mean right?16:24
n4nd0yeah what I said about the ascending order is probably stupid16:27
n4nd0sonney2k: ping16:27
n4nd0sonney2k: do you know something about a problem using MKL with norm>=1?16:27
n4nd0it looks like yoo gets different weights depending on the order he includes the subkernels16:28
yoo and for reproducing in c++/python16:32
CIA-52shogun: Sergey Lisitsyn master * rb630803 / (8 files in 3 dirs): Renamed L1L2 to L12 -
yoon4nd0: could you keep me inform if you have news on this issue ? (
n4nd0yoo: ah so you are the HM-SVM guy too! :)17:25
n4nd0yeah sure17:25
n4nd0but probably they will say something about this here at the chat17:25
n4nd0it will be in the logs17:25
n4nd0but I'll tell you if I see any answer about it posted17:26
-!- blackburn [~blackburn@] has joined #shogun17:41
-!- heiko1 [] has joined #shogun17:41
yooblackburn: hi17:48
yooblackburn: I have not finished the compilation procedure of shogun using cmake, it turns out there is severals macro to handle17:49
blackburnyoo: macro like?17:50
yooblackburn: I have encountered pbs testing mklmulticlass as you can see in the chat logs, have you already experienced smthg similar ?17:50
heiko1blackburn, how to use this python script you sent around for extracting code?17:51
heiko1could you give me your call?17:51
blackburnyoo: yes I've seen17:51
blackburnheiko1: hmm let me remember17:51
blackburnheiko1: python % -s [since_date] -b [end of gsoc date]17:53
n4nd0heiko1: where is that script?17:53
heiko1blackburn, which format is the date?17:53
heiko1and do I have to put my name somewhere?17:53
blackburnheiko1: oops17:53
blackburnn4nd0: heiko1:
blackburn?     ??????????????                    ?????17:59
blackburnoops I was cleaning my kboard17:59
-!- blackburn1 [~blackburn@] has joined #shogun18:01
-!- blackburn [~blackburn@] has quit [Ping timeout: 244 seconds]18:03
heiko1damn it doesnt work18:04
heiko1strange error messages18:04
n4nd0too bad18:04
n4nd0shouldn't there be something built in git to extract a patch?18:04
heiko1n4nd0 does it work for you?18:05
n4nd0didn't try18:06
heiko1argh, always this messy stuff ;)18:07
yoon4nd0: yes I am testing most of the training machines and get used to those18:07
-!- blackburn [~blackburn@] has joined #shogun18:08
-!- blackburn2 [~blackburn@] has joined #shogun18:08
n4nd0heiko1: I have downloaded the tar.gz and it is empty...18:09
-!- yoo_ [2eda6d52@gateway/web/freenode/ip.] has joined #shogun18:09
-!- blackburn1 [~blackburn@] has quit [Ping timeout: 240 seconds]18:09
heiko1blackburn, did the script work for you?18:10
heiko1I get tar error messages18:10
heiko1Think I am too stupid to use it18:10
n4nd0but isn't it the tar.gz empty?18:11
blackburn2heiko1: well I used it last year but no I didn't try yet18:11
heiko1blackburn, I see18:11
heiko1n4nd0 which download?18:11
heiko1I use the once from sergeys mail18:11
n4nd0aah I was checking the one from the link he sent18:12
-!- yoo [2eda6d52@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]18:12
n4nd0I don't have that one from the mail18:12
heiko1Ill send it18:12
-!- blackburn [~blackburn@] has quit [Ping timeout: 244 seconds]18:12
n4nd0it seems that I completely ignored that mail :O18:13
blackburn2heiko1: it must be the same18:13
heiko1blackburn, any updates on the release? I feel bad for having so little time currently ... :(18:14
blackburn2heiko1: I would say we are mostly ready18:14
heiko1blackburn, nice, I will add a few things today which I discussed with Arthur ....18:15
heiko1still the serialisation is missing right?18:15
blackburn2heiko1: I am missing its status to be honest18:15
blackburn2what is wrong with serialization?18:16
heiko1blackburn, memory leaking :)18:16
heiko1reference counts not being stored18:16
heiko1examples failing18:16
blackburn2super :D18:16
heiko1(leaking=failing in this case)18:16
yoo_have to go, see you later18:23
heiko1n4nd0, blackburn got it :)18:24
heiko1(the git extract)18:24
heiko1the script fails if the git log ist is empty18:24
heiko1this can be checked with git log --author= ....18:24
-!- yoo_ [2eda6d52@gateway/web/freenode/ip.] has quit [Quit: Page closed]18:25
n4nd0heiko1: how did you do?18:25
n4nd0I get18:25
heiko1I had a typo in my name before18:25
n4nd0tar: *: Cannot stat: No such file or directory18:25
n4nd0tar: Exiting with failure status due to previous errors18:25
n4nd0test.tar.gz created18:25
n4nd0so we have to use our git name18:26
heiko1git log --author='Heiko Strathmann <>'18:26
heiko1if this works, plug it into the script18:26
heiko1and you have to delete line 35 in the script18:27
heiko1this causes a deadlock18:27
blackburn2haha sorry I thought it is working18:27
heiko1Ill send an example to the list18:29
heiko1n4nd0, blackburn, see mail18:31
n4nd0heiko1: thanks!18:32
n4nd0heiko1: are you going to upload that as code sample?18:33
heiko1n4nd0, just asking myself that :)18:34
heiko1there have been many emails on the list18:34
heiko1but I deleted all of them :)18:34
n4nd0it looks like some people just submitted this18:35
n4nd0I think I will submit this18:35
n4nd0and some simpler example18:35
blackburn2probably nobody takes care of that18:35
n4nd0like the things we prepare for python_modular18:35
heiko1oh n4nd018:35
heiko1I did read your question wrong18:35
heiko1yes you can submit that18:35
heiko1We all did that last year18:36
heiko1but where?18:36
heiko1(thought you were asking that)18:36
n4nd0heiko1: in melange18:36
n4nd0in the page of your project18:36
n4nd0just press edit there18:36
n4nd0heiko1: the script didn't create the tar.gz for me directly, I had to give it an output like /home/....../gsoc.tar.gz18:37
heiko1n4nd0, edit?18:37
heiko1ah found it18:38
-!- blackburn2 [~blackburn@] has quit [Quit: Leaving.]18:39
n4nd0exactly :)18:40
-!- n4nd0 [] has quit [Quit: leaving]19:04
-!- blackburn [~blackburn@] has joined #shogun20:25
-!- blackburn1 [~blackburn@] has joined #shogun20:42
-!- blackburn [~blackburn@] has quit [Ping timeout: 256 seconds]20:44
CIA-52shogun: Heiko Strathmann master * r3ec005c / (13 files in 5 dirs): Merge pull request #760 from karlnapf/master (+6 more commits...) -
CIA-52shogun: iglesias master * r68061ea / (2 files): * fix compile errors in libshogun hmsvm mosek examples -
CIA-52shogun: Sergey Lisitsyn master * rc35b5ed / (2 files): Merge pull request #759 from iglesias/master -
-!- heiko1 [] has left #shogun []21:22
-!- n4nd0 [] has joined #shogun21:34
-!- blackburn1 [~blackburn@] has quit [Quit: Leaving.]21:46
--- Log closed Wed Aug 29 00:00:17 2012