Open in new window / Try shogun cloud
--- Log opened Wed Jul 11 00:00:02 2012
--- Day changed Wed Jul 11 2012
blackburnnothing more00:00
@sonney2kblackburn, I thought you did say this...00:00
@sonney2kmaybe it is a bot after all00:00
blackburnI'm pretty sure he did00:00
@sonney2kreal name 'The rieck'  - hah :))00:00
@sonney2kbut with a goettingen login (where he is now a professor)00:01
@sonney2khmmh I hope I can catch gsomix tomorrow...00:02
@sonney2kanyways nite everyone...00:02
blackburnbad hope00:02
blackburnnot today :)00:02
blackburntomorrow probably00:02
n4nd0sonney2k: good night00:03
blackburnah btw I will be online tomorrow's night only - will leave city for a day to wildprogram :D00:03
wikingsonney2k: yes it did00:07
wikingsonney2k: that that'll take some minor redesign to have it for latent svm00:07
wikingsonney2k: gnite00:07
wikingthis CDualLibQPBMSOSVM  and CPrimalMosekSOSVM support is giving me a hard time...00:14
n4nd0wiking: tell me if I can help you with the CPrimalMosekSOSVM ;)00:22
wikingwell it's more like supporting both at the same time00:23
wikingthat's giving me a headache atm00:23
n4nd0good night guys00:29
-!- n4nd0 [] has quit [Quit: leaving]00:29
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]00:40
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has joined #shogun00:55
-!- av3ngr [] has joined #shogun02:04
-!- av3ngr [] has quit [Client Quit]02:04
shogun-buildbotbuild #18 of nightly_none is complete: Failure [failed compile]  Build details are at
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has quit [Quit: Page closed]03:27
-!- av3ngr [av3ngr@nat/redhat/x-jrdbfjtehidklwda] has joined #shogun04:55
-!- av3ngr [av3ngr@nat/redhat/x-jrdbfjtehidklwda] has quit [Read error: Connection reset by peer]04:55
-!- romi_ [~mizobe@] has quit [Quit: Leaving]06:15
-!- n4nd0 [] has joined #shogun07:39
-!- n4nd0 [] has quit [Ping timeout: 245 seconds]08:42
-!- n4nd0 [] has joined #shogun08:43
-!- uricamic [~uricamic@2001:718:2:1634:5d8f:b4ba:9ffa:dd84] has joined #shogun08:47
-!- pluskid [] has joined #shogun09:21
-!- n4nd0 [] has quit [Ping timeout: 265 seconds]09:35
-!- n4nd0 [] has joined #shogun09:42
-!- hoijui [] has joined #shogun10:55
-!- n4nd0 [] has quit [Quit: leaving]11:29
-!- foo__ [2eda6d52@gateway/web/freenode/ip.] has joined #shogun11:48
foo__question about using mkl classifier: I am not very familar with kernel combination and I would like to interpret a little bit my results: when I print the weights I get values in descending order and the first weight is equal to 0.95 and all the other < 0.01.11:51
foo__is the first weight corresponding to the first kenel : I mean first ker->append(sub_ker) ?11:52
foo__When I evaluate each kernel separatly I get accuracy btw 0.5 and 0.8 then I dont get why the first kernel have a such big weight.11:54
foo__maybe it is dependent of the number of MKL iterations I choose. How to choose it ?11:55
@sonney2kfoo__, did you normalize your kernels?12:03
@sonney2kyes weights are in order12:03
-!- foo__ [2eda6d52@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]12:06
-!- heiko [] has joined #shogun12:10
-!- pluskid [] has quit [Quit: Leaving]12:25
-!- heiko [] has quit [Quit: Leaving.]13:37
-!- n4nd0 [] has joined #shogun14:17
-!- heiko [] has joined #shogun14:43
heikosonney2k, around?15:58
heikoof anybody else who knows how again to activate debug messages in python?15:59
-!- n4nd0 [] has quit [Read error: Operation timed out]16:03
-!- heiko [] has left #shogun []16:04
-!- hoijui [] has quit [Quit: Leaving]16:25
-!- alexlovesdata [] has joined #shogun16:29
-!- foo__ [2eda6d52@gateway/web/freenode/ip.] has joined #shogun16:54
-!- gsomix [~gsomix@] has joined #shogun17:02
foo__sonney2k: thx for your answer. I didn't normalize the kernels.17:03
gsomixsonney2k, hi. :) all news at leta evening17:03
gsomixneed to hard work17:03
foo__Is there any SG method for that in the kernel framework ?17:03
foo__ok I found CKernelNormalizer object17:05
foo__in the kernel constructor there is a kernelnormalizer initialized. how to use it ?17:11
alexlovesdataviking: you did not do a PR for the interface merger about latent/latentstruct yet?17:14
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has joined #shogun17:25
-!- os252 [] has joined #shogun17:43
os252hi puffin44417:43
puffin444hello os25217:43
os252how is it going?17:44
puffin444better. I finally submitted the pr yesterday.17:44
os252yes, great. I am just looking through the PR17:45
puffin444I wanted to at least include some unit tests, but some stubborn memory bugs ate up my time.17:46
os252Yes, it's a quite a bit of code that came together by now!17:47
os252do you have some python demos or similar with plots?17:47
puffin444Not in the current pr, but it would be easy to add them.17:49
os252yeah... I guess best if we discuss what to add before moving on to the icing on the cake.17:50
os252What had you in mind?17:50
puffin444Well as you said the unit tests17:50
puffin444I don't think it should be too much of a hassle to add those17:51
os252Great, I think it's worth it and if you don't add them now they'll never make in there ;-)17:51
puffin444This approximates the derivative, correct?17:51
os252Yes, it is a numerical approximation that allows for comparison with the analytic one.17:51
os252And in terms of kernels?17:52
os252I see the Gaussian, linear, sum and product which is a very solid set to build on.17:52
os252Have you looked at the ARD variants?17:52
puffin444Yes. Again it should be an almost trivial task at this point to add other ones.17:52
puffin444For linear kernels?17:52
os252Linear + Gaussian would be good.17:53
os252Agree, it's just mainly cleaning up this firs part such that it can be useful in practical applications already.17:54
os252So that makes ARD variants of lin+gauss, unit tests for gradients of kernels and demos.17:54
puffin444What else do you think needs to be done before I move on to sparse approx/laplacian inference?17:55
os252I think those points are quite relevant at this stage.17:57
os252+ basic demos that show their use. At the end of the GSOC period it would be great to really push for a couple of very nice demos but perhaps it's best to do this once all the functionality is on the plate.17:57
os252Thus, I'd suggest to now create some basic python demos that solve regression tasks with a standard Gaussian kernel + sum kernel and create the typical sausage plots.17:58
os252You could then towards the end try to get some examples on bigger datasets in which would be really cool to have!17:59
os252once sparse things work, etc.17:59
os252What do you think, how much time does this take?18:00
puffin444For the demos/unit tests, I don't think more than a day or two.18:01
os252ok, great.18:02
puffin444The ARD variants might add a day or two, because ARD itself is not implemented yet.18:02
os252I'd just add 2 new kernels.18:02
os252GaussianARD and LinearARD18:02
puffin444I am sorry that the sum/product kernel gradients took so long. There was a lot more code to change around than I thought.18:03
os252Yes, I do realize. Underestimated the efforts to get this into shogun myself. I am not that deep in the shogun code. Looks like it's not in very good shape.18:04
os252So that makes 2-3 days, let's say 4.18:04
os252Meaning beginning of next week you could start seriously looking into the sparse things.18:04
os252How about we initiate discussions to make plans in parallel?18:04
os252I'll be traveling beginning of August for 14 days, 7 of which I'll be difficult to reach. Perhaps best to try and make use of the time before then to take the key decisions.18:05
os252Have you read this:18:05
os252That's the most comprehensive review on the sparse stuff.18:06
puffin444I also have the GPML examples to for comparison.18:06
foo__Question: how to normalize a kernel (c++) ? There is a kernelnormlizer initialized with the kernel constructor. How to use it ?18:06
os252In which context?18:07
os252You should not normalize really as the hyper parameters are optimized.18:07
puffin444The GPML has sparse approximation methods18:07
foo__I trying to work with mkl machine and I should normalize the kernels before "append" method18:08
foo__how to ?18:08
os252Yes, the question is which ones to add first, etc.18:08
os252I honestly don't know the shogun functions for kernel normalization but still don't get the motivation18:09
os252The MKL is just an approach to learn the weighting of the combinator (sum) kernel.18:09
os252In the GP world you'd achieve the same thing by learning the hyper parameters through optimization of the marginal likelihood.18:09
puffin444os252, any more questions?18:12
os252not from my side - but would be curious to hear what you planed with the MKLmachine18:12
puffin444that's foo__, not me :)18:13
alexlovesdatafoo__: you could use custom kernels and normalize them outside in your preferred style. This is what I do18:14
os252Oops. Sorry foo__ puffin444: this IRC client has a very suboptimal color scheme assigning both of you the same non-legible color.18:15
os252ok puffin444, do you want to catch up some time to discuss the sparse methods? I suggest you take a look at that review and based on this and the existing code base you should get a good overview.18:16
-!- foo__ [2eda6d52@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]18:16
os252Easiest would be some subset of regressors sparse model first, which essentially ignores some of the training data, choosing an "intelligent" subset.18:16
os252That's probably the thing to start with.18:17
puffin444Would Friday/Saturday be a good time to discuss this?18:17
os252yes, both good but we need to email. My father is in hospital and I'll go and visit this Friday. I can email tomorrow when I have a better clue how my schedule looks like, ok?18:18
puffin444Ok, sounds good.18:18
-!- uricamic [~uricamic@2001:718:2:1634:5d8f:b4ba:9ffa:dd84] has quit [Quit: Leaving.]18:24
os252I'll email. Talk soon puffin44418:35
-!- os252 [] has left #shogun []18:35
-!- os252 [] has joined #shogun18:35
-!- os252 [] has quit [Quit: os252]18:40
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]18:45
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has joined #shogun18:51
-!- n4nd0 [] has joined #shogun19:06
-!- gsomix [~gsomix@] has quit [Ping timeout: 248 seconds]19:27
puffin444Hey sonney2k, does my pull request still say it's not rebased?19:34
-!- gsomix [~gsomix@] has joined #shogun19:42
puffin444Hey gsomix, how do you tell if pull requests are able to be merged (i.e, they are rebased)?19:55
puffin444*how can19:56
@sonney2kpuffin444, now it looks good20:18
puffin444okay good.20:18
@sonney2kpuffin about sgvector / sgmatrix20:18
@sonney2kyou just need to do SGVector<float64_t> x = y;20:18
@sonney2kno need for ptrs, freeing of memory etc20:18
puffin444oh where do I do that?20:23
@sonney2kpuffin444, ?20:24
@sonney2kthat was meant as a general hint20:24
puffin444oh okay20:24
puffin444I typically do not do that20:24
@sonney2kI don't know if you know how to use them...20:24
@sonney2kor if you even use them20:24
-!- n4nd0 [] has quit [Quit: leaving]20:25
puffin444There was some untested stuffin modelselection that I removed right now.20:25
puffin444*stuff in20:26
puffin444regarding vectors. It will be back in soon.20:26
-!- blackburn [~blackburn@] has joined #shogun20:31
blackburnhey what's up20:32
@sonney2kpuffin444, so do you think your patch can be merged?20:34
@sonney2kpuffin444, as I understood you left out the 'experimental' sgvector/matrix stuff20:34
@sonney2kblackburn, please also have one last look20:34
puffin444I made the fixes you prescribed. I left out that experimental sgvector/matrix stuff20:34
@sonney2kpuffin444, ok so then good from my side20:35
@sonney2kblackburn, any comments?20:35
@sonney2kpuffin444, btw did you submit your evaluation?20:35
puffin444Yes. Thanks for reminding me. For some reason I thought they came next week.20:35
@sonney2kpuffin444, np - it is my job to make sure no one fails due to administrative issues20:36
blackburnsonney2k: a sec20:39
blackburnaham dictionary parameters20:40
blackburnsonney2k: did cwidmer send?20:40
puffin444what's wrong with the dictionary parameters?20:41
blackburnpuffin444: nothing wrong at all - I was just curious how will you approach your problem20:41
puffin444The dictionary builds a mapping of parameters to the objects that own them. This includes the current sgobject and any sgobjects which it owns as parameters20:42
blackburnsonney2k: I glanced over puffin444 PR - looks ok20:42
@sonney2kblackburn, everyone did send :))20:42
blackburnsonney2k: cool20:43
CIA-18shogun: Soeren Sonnenburg master * r179b38e / (41 files in 9 dirs): Merge pull request #638 from puffin444/master (+7 more commits...) -
shogun-buildbotbuild #20 of bsd1 - libshogun is complete: Failure [failed configure]  Build details are at  blamelist: puffin444 <>20:49
puffin444wonderful. I checking it now...20:50
@sonney2kpuffin444, not your fault20:51
@sonney2kblackburn, I think I finally understand when this happens20:52
@sonney2kwhen we change anything in configure20:52
@sonney2kbut someone already committed stuff locally long before20:52
@sonney2k(the required change to configure)20:52
@sonney2kthen kaboom20:52
blackburnsonney2k: do you have any clue how to fix it?20:53
@sonney2kblackburn, not possible20:56
shogun-buildbotbuild #21 of bsd1 - libshogun is complete: Success [build successful]  Build details are at
@sonney2kit is because the buildbot builds the first commit20:56
@sonney2kso if that one is before the required configure change - it cannot work20:56
@sonney2kblackburn, but it is clear that this is a temporary thing..20:57
-!- blackburn [~blackburn@] has quit [Read error: Connection reset by peer]20:57
@sonney2k(well it can happen for every configure change...)20:57
-!- blackburn1 [~blackburn@] has joined #shogun20:58
@sonney2kgsomix, any progress on the buffer proto stuff?21:00
shogun-buildbotbuild #125 of deb3 - modular_interfaces is complete: Failure [failed compile csharp_modular]  Build details are at  blamelist: puffin444 <>21:05
@sonney2kblackburn1, actually21:06
@sonney2kno - this should only happen if someone forgot to pull --rebase21:06
@sonney2kpuffin444, did you do git pull --rebase all the time?21:06
puffin444No I tend to use merge.21:07
@sonney2kpuffin444, but with --rebase ?21:07
@sonney2kargh not possible21:07
puffin444No. I would execute git merge upstream/master21:07
blackburn1only pull with rebase is21:08
@sonney2kblackburn1, see that explains it21:08
-!- blackburn1 is now known as blackburn21:08
-!- rieck [] has quit [Remote host closed the connection]21:08
@sonney2kpuffin444, then please do git rebase upstream/master from now on21:08
shogun-buildbotbuild #116 of deb2 - static_interfaces is complete: Failure [failed test libshogun]  Build details are at  blamelist: puffin444 <>, Soeren Sonnenburg <>21:08
puffin444How did that fail? That should be fixed.21:09
puffin444Investigating now...21:11
@sonney2kpuffin444, when you do a merge instead of a rebase all commits that are now in upstream/master will be put on top of what you did21:11
@sonney2kpuffin444, so this explains all the crap we see now21:14
@sonney2kall failures we've seen - we will see again now21:14
@sonney2kpuffin444, so stop investigating until it is at the very last commit21:14
puffin444sonney2k, very sorry about this. Will be using rebase command for now on.21:15
@sonney2kpuffin444, don't worry too much everyone her made this mistake21:17
blackburnsonney2k: yyeah I received that mail :)21:18
@sonney2kWe regard that code [the code done during GSoC] as potentially throwaway code!21:19
@sonney2k^ citing from the above21:19
blackburnwhy do you emphasize that?21:19
@sonney2kblackburn, isn't that a suprise to you?21:20
blackburnthey are much bigger project I'd say21:20
@sonney2kthat doesn't explain this to me21:21
blackburnand moreover21:21
blackburntheir student level is lower21:21
@sonney2knot sure what you mean...21:22
blackburnwe have pretty good students21:22
@sonney2kahh ok yes21:22
blackburnthat are able to understand both math and programming issues21:22
@sonney2kit is probably easier to do packaging work than ML21:23
@sonney2kblackburn, any idea where gsomix is?21:25
blackburnsonney2k: it is night here so he is probably not in his garden21:26
@sonney2kblackburn, what?21:27
@sonney2kapple attack21:28
* sonney2k probably not the wisest idea to sit under an apple tree coding21:28
@sonney2kblackburn - look at that
blackburnsonney2k: are you sitting under a tree? :)21:29
blackburnsonney2k: well not a surprise definitely21:30
@sonney2kblackburn, I moved a bit21:30
@sonney2kafter the apple attack *lol*21:30
@sonney2klots of other interesting stuff in there21:30
@sonney2kblackburn, e.g. communication via IRC makes it more likely that people to succeed21:31
@sonney2kprivate mails rather cause failures21:32
blackburnsonney2k: I am sure that's better than mails21:32
blackburnnot the privacy thing though I believe21:32
blackburnit is just slower21:32
@sonney2kno it really is the privacy thing21:33
blackburnblue thing is funny21:34
-!- alexlovesdata [] has left #shogun []21:34
@sonney2kit is from this page21:34
@sonney2kso it is even more important that we get mentors/students all to stay in IRC...21:36
gsomixsonney2k, hey21:36
shogun-buildbotbuild #126 of deb3 - modular_interfaces is complete: Failure [failed test libshogun]  Build details are at  blamelist: puffin444 <>, Soeren Sonnenburg <>21:41
@sonney2kgsomix, hey - so how is it going? did you make progress on the buffer protocol stuff?21:41
gsomixsonney2k, yep, I managed to how hack swig, hehe. I'm writing some stuff now, results will be li-i-i-i-ttle later.21:41
@sonney2kgsomix, ok21:42
@sonney2kgsomix, btw this is not yet with the 'avoid copying' stuff right?21:42
@sonney2kblackburn, btw I will be on that island again from end of this week21:42
blackburnhmm ok21:42
blackburnshould be ok according to previous experience21:43
@sonney2kI wonder what needs attention...21:43
@sonney2k...and is not too boring :D21:43
@sonney2kor awful :)21:44
blackburnsonney2k: like coding?21:44
gsomixsonney2k, hm, I just have read numpy code. it seems, that is easy21:44
@sonney2kblackburn, btw what's up with the doc/tutorial? I've seen you added an empty readme21:44
blackburnsonney2k: yeah but didn't manage to add something yet21:44
@sonney2kgsomix, ok well one step after another21:44
@sonney2kblackburn, why don't you just add the stuff you had for your dim red toolbox?21:45
blackburnsonney2k: I feel embarrassed getting back to that thing :D21:45
gsomixsonney2k, I sent midterm evaluation answers btw21:45
@sonney2kgsomix, I know :-)21:45
blackburngsomix: he knows everything about us21:45
@sonney2kso now we can all continue21:45
@sonney2kblackburn, why don't you just start with that - no one can help if you sit on that thing...21:46
@sonney2kand we really need some interaction going...21:46
gsomixsonney2k, blackburn big brothers >:]21:47
blackburnI know21:47
@sonney2kblackburn, I am actually in the mood (for that island) to write a bit...21:47
@sonney2kmaybe :D21:47
@sonney2kand drink vodka with you joining remotely21:47
blackburnsonney2k: if you want to you may cut something from you phd thesis21:47
blackburntoo bad we will not manage to make a vodka appointment21:48
blackburnin august :D21:48
@sonney2kmy kids have to become older for that ;-)21:49
blackburnsonney2k: do you want to drink vodka with your kids?21:49
@sonney2kmaybe next year if heiko manages to rescue the money21:49
blackburnif so you would have to wait 16 years more or so21:49
@sonney2kblackburn, it is all just water ;-)21:49
blackburn60% you are right21:50
blackburnsonney2k: what do we need from you is kernels description21:50
blackburnI have *no clue* what is difference between LocalAlignmentAnything and WeightedDegreeSomething21:51
@sonney2kpuffin444, now please have a look at the issues...21:51
@sonney2k^ puffin44421:51
@sonney2kblackburn, totally different indeed21:52
puffin444That's wierd. Somehow my fix disappeared. Just a minute...21:52
@sonney2kblackburn, yes I can easily do a kernels section21:52
@sonney2kblackburn, but I need the tutorial structure from you21:52
@sonney2kbefore I leave even21:52
blackburnsonney2k: give me that night or next day21:52
@sonney2kstarting to rain21:52
@sonney2kstopped again21:53
@sonney2kgsomix, little brother - please ping me when you think you have sth :)21:54
blackburnargh russian intel video drivers developer has died in sao paulo21:54
gsomixsonney2k, ok, ok21:54
@sonney2kblackburn, link?21:54
@sonney2kok serious rain now21:55
* sonney2k escapes21:55
blackburnsonney2k: evgeniy dodonov - former mandriva developer, and was in intel linux graphics team21:59
puffin444sonney2k, I have a fix for the model_selection_tree bug22:42
blackburnpuffin444: merged22:55
CIA-18shogun: Sergey Lisitsyn master * r7b220bf / (2 files): Merge branch 'master' of -
CIA-18shogun: puffin444 master * r7a90611 / (2 files): Fixed update bug in InferenceMethod -
CIA-18shogun: puffin444 master * rd5a2772 / (2 files): Fixed bug in Inference Method initialization. -
blackburnpuffin444: did you rebase this time?22:55
blackburnpuffin444: do you have anything in latex you can and want to merge into constructing tutorial?22:57
puffin444Not yet.22:57
blackburn puffin444, anyway feel free to fork and do pull request just like for shogun ;)22:58
puffin444I definitely do plan to add documentation22:59
shogun-buildbotbuild #127 of deb3 - modular_interfaces is complete: Failure [failed test libshogun]  Build details are at  blamelist: puffin444 <>23:21
CIA-18shogun: Sergey Lisitsyn master * ra46a5cf / src/shogun/kernel/Kernel.cpp : Removed unhandled case in switch warning cause -
shogun-buildbotbuild #117 of deb2 - static_interfaces is complete: Success [build successful]  Build details are at
puffin444oh okay23:29
blackburnpuffin444: take builbot easy :D23:31
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has quit [Quit: Page closed]23:41
-!- cwidmer [8ca3fe9d@gateway/web/freenode/ip.] has joined #shogun23:43
shogun-buildbotbuild #128 of deb3 - modular_interfaces is complete: Success [build successful]  Build details are at
--- Log closed Thu Jul 12 00:00:17 2012