Open in new window / Try shogun cloud
--- Log opened Thu Jun 21 00:00:41 2012
-!- emrecelikten [~emre@176.41.10.108] has quit [Ping timeout: 246 seconds]03:11
-!- emrecelikten [~emre@176.40.230.182] has joined #shogun03:27
-!- emrecelikten [~emre@176.40.230.182] has quit [Quit: Leaving.]04:38
-!- emrecelikten [~emre@176.40.230.182] has joined #shogun04:38
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]05:18
-!- emrecelikten [~emre@176.40.230.182] has quit [Quit: Leaving.]05:23
-!- uricamic [~uricamic@2001:718:2:1634:5cfb:a4b3:604d:78a] has joined #shogun08:48
-!- emrecelikten [~emre@176.40.230.182] has joined #shogun08:54
@sonney2kuricamic: could n4nd0 directly compare his .mc-so code with your example ?09:03
uricamicI don't know if he tried, but it should be possible now09:04
@sonney2kI mean you do mc via the so  framework but with bmrm right,?09:04
uricamicyes09:04
uricamicI told him that as soon as it will be finished, he can try to compare09:04
uricamicbtw I have some problems with that pull request09:04
@sonney2kI assume you cannot have a true bias term?09:05
uricamicis it possible to select only some commits for the PR?09:05
uricamicno09:05
@sonney2kno what?09:05
uricamicthe true bias term09:06
@sonney2kwell you need to create a new branch09:06
@sonney2kso no bias term?09:06
uricamicI see, ok, and then it is possible to move commits to the new branch?09:06
uricamicyes, we are not using the bias term09:07
@sonney2kin the worst case with git format-patch09:07
uricamicok, thanks, I will look on that09:08
@sonney2kalright gtg09:08
-!- pluskid [~pluskid@202.130.113.141] has joined #shogun09:51
pluskidhi sonney2k, sorry, just see the email09:52
-!- emrecelikten [~emre@176.40.230.182] has quit [Quit: Leaving.]09:55
-!- emrecelikten [~emre@176.40.230.182] has joined #shogun10:10
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun10:25
@sonney2kuricamic, did you manage to create a new pull request?12:21
@sonney2kpluskid, I just wanted to say that it is Ok to merge - but we should use the shogun core functions to not have duplications everywhere12:22
@sonney2kpluskid, but I guess it is better you have some working example before refactoring the code...12:23
pluskidsonney2k: you mean memory allocation?12:26
@sonney2kpluskid, yes and the vector functions like \alpha x etc12:26
@sonney2kpluskid, I would want to see if the sse optimized stuff brings any gain (which I don't expect)12:27
@sonney2kthe compiler usually does the same12:27
pluskidsonney2k: the lbfgs doc require that we use lbfgs_malloc to allocate memory, I still haven't look into its implementation to see what it has done, if it is just plain malloc, I think we can replace that with SG_MALLOC12:28
@sonney2kpluskid, I looked at the code - they have their own malloc function and it uses malloc12:28
pluskidsonney2k: and by "\alpha x" you mean which code?12:28
@sonney2kactually it does malloc + zeros the stuff12:28
@sonney2kso SG_CALLOC would be the replacement12:29
pluskidsonney2k: OK, thanks!12:29
@sonney2khttps://github.com/shogun-toolbox/shogun/pull/590/files12:29
@sonney2klook at arithmetic ansi12:29
pluskidsonney2k: ah, I see, you mean we replace those with SGVector stuff?12:31
@sonney2kpluskid, yes12:31
pluskidI might be able to try to do this after I made ShareBoost working12:31
@sonney2kor take their code if theirs is better12:31
@sonney2kpluskid, it is very little amount of work12:32
@sonney2kbut yes12:32
@sonney2kfirst some reference12:32
@sonney2kotherwise it is hard to detect if one breaks things12:32
pluskidbut I guess keeping the original code untouched will be easier for future upgrade (to newer version of liblbfgs), what do you think?12:32
@sonney2kpluskid, the last update is years back12:35
@sonney2kso no12:35
@sonney2kthis lib is basically unmaintained12:35
pluskidok12:35
uricamicsonney2k: I have done it right now, had some troubles with moving the commits, but hope now it is ok12:45
@sonney2kuricamic, "This pull request cannot be automatically merged."12:45
@sonney2kso no it is not...12:46
uricamichmm, I have done git cherry-pick, could this be because of that?12:46
@sonney2kuricamic, my suggestion would be to update to the state of master12:46
@sonney2kand then in a branch just apply the missing patches12:46
@sonney2kuricamic, no idea - you have files in there that are already merged12:47
@sonney2kfor example I moved things from so -> structure a day ago or so12:47
@sonney2kseems like you do the same thing now12:47
uricamicI know, I have done some commit and then before PR it introduced the conflict, so I resolved it, but it seems it was not good12:48
uricamicI will try to fix it12:49
@sonney2kuricamic, look at https://github.com/shogun-toolbox/shogun/pull/593/commits12:50
@sonney2kthere is 'indention fixes'12:50
@sonney2kwhich is long merged12:50
@sonney2kuricamic, as I said I would suggest to do a git pull --rebase on your master branch12:56
@sonney2kthen resolve the conflicts12:56
@sonney2k(git add the conflicting files afterwards)12:56
@sonney2kand continue the merge12:57
n4nd0sonney2k: is there any reason why DynamicObjectArray is not includid in Library.i for the modular interfaces?13:17
n4nd0sonney2k: I have found it in ModelSelection.i but for some reason to do get_name from python fails with it, I want to include in Library and do the rename and so13:18
@sonney2kn4nd0, yes please fix13:20
n4nd0ok13:20
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: Lost terminal]13:49
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun13:50
@sonney2kuricamic, did you get it to work?13:51
uricamicsonney2k: working on it, I think I have found the source of the problem, so I have decided to introduce new commits from scratch13:55
uricamicsoon there will be a PR13:56
@sonney2kuricamic, ok13:56
uricamicsonney2k: done. I hope now it will work14:02
CIA-18shogun: Michal Uricar master * r2cbb216 / (2 files): libbmrm fixes - http://git.io/IVx0MA14:04
CIA-18shogun: Michal Uricar master * r6e12559 / (3 files): minor RiskFunction change + MulticlassRiskFunction implementation - http://git.io/mh_HQg14:04
CIA-18shogun: Michal Uricar master * rd8c48bf / (2 files): DualLibQPBMSOSVM fixes due to changes in libbmrm and RiskFunction - http://git.io/J8BFaw14:04
CIA-18shogun: Soeren Sonnenburg master * ra53beb2 / (7 files): Merge pull request #594 from uricamic/BM_SOL2 - http://git.io/yWso-w14:04
@sonney2kuricamic, alright14:04
@sonney2kthanks14:04
@sonney2kuricamic, so could n4nd0 use this for comparison to his MC-SO mosek based thing?14:04
uricamicI thank you for patience :)14:04
uricamicI think he can14:04
@sonney2kuricamic, btw do you have the value objective at the end?14:04
@sonney2kthen you could both compare nicely...14:05
uricamicwell now it is only written on output, I think, but it is there14:05
uricamicoh, I just realized, I haven't post the example to this PR14:06
n4nd0yes yes, I want to test my stuff! testing testing :D14:06
n4nd0sonney2k: btw, is there any way to compile and just show the warnings?14:09
n4nd0any flag or so14:09
n4nd0otherwise I will redirect the output and grep or something like that14:10
@sonney2kn4nd0 ... I use vim and :mak it will directly stop at all warnings/errors14:10
n4nd0ok, thank you!14:11
shogun-buildbotbuild #1082 of libshogun is complete: Failure [failed compile]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/libshogun/builds/1082  blamelist: uricar.michal@gmail.com14:11
@sonney2kuricamic, http://www.shogun-toolbox.org/buildbot/builders/libshogun/builds/1082/steps/compile/logs/stdio ?14:15
@sonney2kseen the error?14:15
uricamicthis is because ot that changes, I will fix it14:16
shogun-buildbotbuild #1083 of libshogun is complete: Success [build successful]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/libshogun/builds/108314:16
uricamicbut this is strange14:19
uricamicI just compiled it on my computer and everything works14:20
@sonney2kuricamic, well you committed something that doesn't compile14:22
@sonney2kand then later fixed it14:22
@sonney2kand committed again14:22
@sonney2kuricamic, next time please use git commit --amend14:23
@sonney2kto create just one *but working* commit14:23
uricamicso it is ok now?14:23
@sonney2kuricamic, yes ^ buildbot above said so14:34
uricamicaha, ok, thanks, next time I will be more careful14:35
uricamicI am trying to push also the example I forgot to push it into the last PR, but it seems something is wrong with my connection14:36
uricamicI have sent the PR with the example14:56
-!- uricamic [~uricamic@2001:718:2:1634:5cfb:a4b3:604d:78a] has quit [Quit: Leaving.]15:25
-!- pluskid [~pluskid@202.130.113.141] has quit [Ping timeout: 246 seconds]15:57
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving]16:10
-!- pluskid [~pluskid@111.120.60.54] has joined #shogun16:16
-!- blackburn [d557894d@gateway/web/freenode/ip.213.87.137.77] has joined #shogun16:25
blackburnhey there16:25
blackburngot bsc :)16:26
-!- ckwidmer [8ca3fe9d@gateway/web/freenode/ip.140.163.254.157] has joined #shogun16:27
-!- pluskid [~pluskid@111.120.60.54] has quit [Ping timeout: 264 seconds]16:49
-!- pluskid [~pluskid@202.130.113.141] has joined #shogun16:50
-!- blackburn [d557894d@gateway/web/freenode/ip.213.87.137.77] has quit [Ping timeout: 245 seconds]17:05
-!- pluskid [~pluskid@202.130.113.141] has quit [Quit: Leaving]17:15
-!- gsomix [~gsomix@188.168.14.46] has joined #shogun17:40
-!- alexlovesdata [c25faea9@gateway/web/freenode/ip.194.95.174.169] has joined #shogun18:29
@sonney2kgsomix, how is it going? recovered from the tornado?18:48
gsomixsonney2k, huh. yes. minimal destruction.18:48
@sonney2kgsomix, any updated regarding gsoc besides reconstruction works?18:49
gsomixsonney2k, not many, only drafts of examples. I'm little busy by last exams now. However, I'll finish with examples this week.18:51
@sonney2kgsomix, when are your exams?18:52
gsomixsonney2k, last on 25 June. is there some job related to director classes except examples for machines and dotfeatures?18:53
gsomixand I have passed optics exam at 19 June.18:56
@sonney2kgsomix, well I can think of sparse features, string features, multiclass machines, preprocessors, structured output machines, converters, loss, labels18:58
gsomixwow!18:58
@sonney2kbefore midterm ...18:58
@sonney2kthen after that the other cool feature18:59
@sonney2kpython buffer protocol18:59
@sonney2kthat is not so easy to get to work18:59
@sonney2kanyway, what it does is it will enable you to work with shogun densefeatures etc as if they are numpy matrices19:00
@sonney2kso one can from python do x + y (with x being a shogun dense feature obj and y being a numpy array)19:00
@sonney2kand even x+=y19:00
@sonney2ksuch stuff19:01
@sonney2kgsomix, for that you will need to write a simply python extension (stand-a-alone like for directors)19:02
@sonney2kthat just is some double* x, int len19:02
@sonney2ksome vector thingy19:02
@sonney2kand then register a new pyobject for that type implementing the buffer protocol19:03
@sonney2kif this works19:03
@sonney2kthe next step is to create a simple swig example which does the same19:03
@sonney2kthat will require a very recent version of swig though.19:04
@sonney2kif that works - do this inside shogun and add some convenience operations to shogun features19:05
alexlovesdatais it possible that MKL is broken in 1.1.0??19:05
gsomixsonney2k, swig2?19:05
@sonney2kmaybe 2.0.6 or so19:06
@sonney2knot sure19:06
gsomixok19:06
gsomix>> with shogun densefeatures etc as if they are numpy matrices | it's cool feature. I think that cpp+python+swig is one of best "platform" for programming19:07
gsomixafter lisp, haskell... huh, joke :)19:08
CIA-18shogun: Michal Uricar master * ra5c9046 / examples/undocumented/libshogun/so_multiclass_BMRM.cpp : BMRM SOL Multiclass example - http://git.io/urAKog19:08
CIA-18shogun: Soeren Sonnenburg master * r9c2be58 / examples/undocumented/libshogun/so_multiclass_BMRM.cpp : Merge pull request #595 from uricamic/BM_SOL_EXAMPLE - http://git.io/RVDxgw19:08
@sonney2kgsomix, the director stuff is not a lot of work now that we know how it works19:09
@sonney2konly dotfeatures is involving19:09
gsomixsonney2k, get it. so, thanks for further plan. you made my work easier. :)19:16
-!- ckwidmer [8ca3fe9d@gateway/web/freenode/ip.140.163.254.157] has quit [Ping timeout: 245 seconds]19:34
-!- gsomix [~gsomix@188.168.14.46] has quit [Remote host closed the connection]21:08
-!- blackburn [d557894d@gateway/web/freenode/ip.213.87.137.77] has joined #shogun23:06
blackburnhey there23:07
blackburnsonney2k: around?23:11
@sonney2kblackburn, yes23:12
@sonney2kblackburn, seen that you broke the tests with chris' stuff?23:12
@sonney2kblackburn, ahh and congrats!23:12
blackburnsonney2k: yes yes yes23:12
blackburnthanks23:12
blackburnsonney2k: I'll fix tomorrow23:12
@sonney2ktoo much vodka already?23:13
blackburnno had some headache after all23:13
blackburnso had to get some sleep :)23:13
@sonney2knaah sleep ;-)23:14
blackburnI wish I had some vodka and had headache because of that :)23:14
blackburnsonney2k: ok my plans tomorrow include fixing that23:15
blackburnand sparsevector too23:16
blackburnthe last thing to change for referenced data23:16
@sonney2kstring stuff yes23:16
blackburnsonney2k: what is purpose of cached vector?23:16
blackburnI mean how do we get any impact on performance23:17
@sonney2kblackburn, simple err dense features don't need to be in memory23:17
blackburnthat is the thing I am getting confused23:17
blackburnsonney2k: but they have matrix aren't they?23:17
@sonney2kand preprocessors can apply things just one by one23:17
@sonney2kblackburn, no23:17
@sonney2kmatrix is optional23:17
@sonney2kso things can be cached23:18
blackburnhow can we get densefeatures with no matrix but some contents inside?23:18
@sonney2kthat is why we have this get_feature_vector / free_feature_vector business23:18
blackburnfor example in python23:18
@sonney2kwith these functions23:18
@sonney2kthings are fetched/preprocessed/computed on-demand23:18
blackburnsonney2k: did guys make some great progress on SO already?23:22
blackburna lot of commits I am lost in23:22
@sonney2kblackburn, well at least some progress - framework is merged and uricamics bmrm with application to MC is in there23:23
@sonney2kn4nd0 needs to compare against it and make some progress23:23
@sonney2kblackburn, nothing from wiking yet - no idea where he is stuck...23:24
blackburnsonney2k: isn't it easy to get basically anything to work after getting MC example to work?23:24
blackburnI do not understand that point23:24
@sonney2kthe GP stuff seems also to be progressing well23:24
@sonney2kand gsomix you know probably better23:24
blackburnyes no news from wiking23:24
@sonney2kand pluskid is working like 10 of us as usual23:25
@sonney2kblackburn, yes MC is just the PoC23:25
blackburnPoC?23:26
@sonney2kthen one 'only' needs to change the argmax23:26
@sonney2kproof-of-concept23:26
blackburnaham23:26
@sonney2kbut the argmax can be *very difficult* depending on so problem23:26
blackburndoes that mean they can integrate some HMM or anything pretty easy23:26
blackburn?23:26
blackburnahh23:26
@sonney2kHM-SVM yes23:27
blackburnso argmax is the most hard to code thing?23:27
blackburngot it23:27
@sonney2kblackburn, well HM-SVM is not so difficult23:27
blackburnI actually checked HM-SVM code23:27
@sonney2kafter all it is just viterbi algorithm23:27
blackburnI do not like it pretty much hehe23:27
@sonney2kmuch more difficult is Semi HM-SVM23:27
blackburnyour expertise :)23:27
@sonney2knon-markovian behavior on segments23:28
@sonney2kbacktracking from hell23:28
blackburnsemi-supervised?23:28
@sonney2kno23:28
blackburnsemi what then?23:28
@sonney2kthe markov property is relaxed23:28
blackburnhow 'much'?23:28
@sonney2kyou can have arbitrary look back23:29
@sonney2kso no longer p(x_t| x_{t-1})23:29
blackburnnot limited?23:29
@sonney2klimited!23:29
blackburnwith?23:29
@sonney2kbut up to whole segment length23:29
blackburnahh23:29
@sonney2kso you can model things like length distribution of segment23:29
blackburnokay nice23:29
blackburnpretty complex thing23:30
@sonney2kor frequencies of characters in segements23:30
@sonney2ketc23:30
@sonney2kwhatever you want to compute for a whole segment23:30
blackburnis it called semi markov?23:30
@sonney2kargmax on graphs is probably even worse23:30
@sonney2kyes23:30
blackburnI see23:30
@sonney2ksometimes generalized HM *23:30
blackburnit looks like you have some passion with that :)23:30
@sonney2kblackburn, recall that I wrote the HMM code23:31
blackburnsonney2k: I recall you have master thesis on markov stuff23:31
@sonney2k...which I no longer can parse since gunnar parallelized it23:31
@sonney2kblackburn, no23:31
@sonney2kstudent research project only23:32
blackburnam I wrong?23:32
blackburnahhh23:32
blackburnwhat is your master thesis about then?23:32
@sonney2ksomething before master thesis23:32
@sonney2khmm + top / fisher kernel + svm23:32
@sonney2kfor splice site recognition23:32
blackburnah more non-understable stuff :)23:32
@sonney2kso I tried a couple of things to detect splice sites23:32
@sonney2khmms23:32
@sonney2klocality improved kernel on splice sites23:33
@sonney2kthen fisher/top kernel on hmms23:33
@sonney2kand this in the end evolved into the weighted degree kernel :)23:33
blackburnit actually looks like some over-complexity for me23:33
blackburnisn't it?23:33
@sonney2kblackburn, no23:33
@sonney2ksometimes you don't know how to compare things23:34
@sonney2kbut you have some good probabilistic models23:34
@sonney2kso you derive the fisher/top kernel from it23:34
@sonney2kand use an svm and get much better results...23:34
@sonney2kthis *can* be very complex23:34
blackburnwhy to use svm if you have good prob model?23:34
@sonney2kblackburn, because it is just some OKish model23:35
blackburnI see23:36
@sonney2kblackburn, for example GNB is probably very close to poly kernel23:37
blackburnsonney2k: btw I want to know your opinion on semisupervised and transductive stuff23:37
@sonney2kI mean the fisher kernel from GNB23:37
blackburn(btw because you said semi word and I am activated now :D)23:37
@sonney2kblackburn, not really a success story so far23:38
blackburnsonney2k: I remember there is some relation between GNB and logit regression23:38
@sonney2kblackburn, fisher kernel is just gradient of log(p(x))23:39
@sonney2k(except for some normalization that no one uses)23:40
blackburnsonney2k: what do you consider as success story then?23:41
@sonney2kwell SVMs for classification23:42
@sonney2kGPs for regression23:42
@sonney2klinear methods in general23:42
@sonney2kLDA, linear SVM ...23:42
@sonney2kSO learning23:42
@sonney2knot even true multiclass23:42
blackburnsonney2k: I want to push my gf to research GPs23:42
blackburnsonney2k: true multiclass is not kind of success for you?23:43
@sonney2kwell KNN is23:43
@sonney2kand trees :D23:44
@sonney2kblackburn, almost no gain compared to OvR23:44
blackburnsonney2k: yes, I actually gain really small % of accuracy on my road sign data..23:44
@sonney2kahh and PCA :)23:47
blackburnsonney2k: what is so cool about PCA?23:48
@sonney2kthat it is often relevant in practise23:48
@sonney2kthere is lots of stuff out there that is rarely useful but overly complicated23:48
blackburnsonney2k: I never managed to use it properly..23:48
@sonney2kblackburn, to get an idea/ visualize data only I mean23:49
blackburnah23:49
blackburnLLE is cool for that too I believe23:49
blackburnand isomap/mds23:49
blackburnltsa, hlle, blabla is useless but these ones are cool23:50
@sonney2kI like stuff that works *often*23:50
@sonney2kand is not so difficult and potentially even fast23:50
blackburnmds and isomap work always23:50
@sonney2kblackburn, work as in give useful results?23:51
@sonney2kdon't you need some distance for that?23:51
blackburnsonney2k: yes always constructs some embedding23:51
blackburnyes23:51
blackburnsonney2k: I will implement t-SNE soon in shogun I think23:52
blackburnit is probably the best 'visualizer'23:52
@sonney2koh well whatever method you choose - experience tells that you can gain much more by finding the right features23:52
@sonney2kand this is actually the hard part23:53
blackburnsonney2k: http://homepage.tudelft.nl/19j49/t-SNE_files/mnist_large.jpg23:53
blackburnthat's how t-SNE works on MNIST23:53
@sonney2kimpressive23:53
blackburnyes, very impressive23:54
blackburnsonney2k: http://homepage.tudelft.nl/19j49/t-SNE_files/vis_caltech101_small.jpg one more example23:54
blackburnbut actually LLE can produce similar results23:55
@sonney2kI think the distance function matters a lot here23:56
-!- zxtx [~zv@cpe-75-83-151-252.socal.res.rr.com] has quit [Ping timeout: 246 seconds]23:56
blackburnexactly23:56
-!- shogun-buildbot [~shogun-bu@7nn.de] has quit [Quit: buildmaster reconfigured: bot disconnecting]23:56
blackburnbut still method matters too23:56
-!- shogun-buildbot [~shogun-bu@7nn.de] has joined #shogun23:57
@sonney2kblackburn, the new buildbot will create documentation on nightly_default23:57
@sonney2kif it doesn't fail of course23:57
blackburncool23:57
@sonney2kblackburn, so highest prio is to get the build back to green23:57
blackburnwill do23:58
@sonney2kblackburn, hmmhh I think I know some people on that image23:58
@sonney2kblackburn, it will be under http://www.shogun-toolbox.org/doc/en/latest/ then23:58
@sonney2k(if I didn't make a mistake :D23:59
blackburnsonney2k: know? really?23:59
blackburnit is caltech-101 dataset23:59
--- Log closed Fri Jun 22 00:00:20 2012