Open in new window / Try shogun cloud
--- Log opened Wed Oct 31 00:00:17 2012
@sonney2kwiking, what's that?00:05
@sonney2kwiking, where is you gcc faq?00:06
wikingnowhere yet00:21
wikingi'm still looking into how to redesign the faq page00:21
wikingah ok so we dont have sitemap00:21
wikingi'll generate one then00:21
@sonney2kwiking, we might be able to do that automagically based on django urls00:25
@sonney2kwiking, it is a bit of work but possible
@sonney2kahh what is badly missing from our website is the feature matrix00:31
@sonney2kthis one here
@sonney2kit is created based on some google spredsheet00:32
-!- n4nd0 [] has quit [Quit: leaving]00:51
-!- ptizoom [~christian@] has quit [Quit: Ex-Chat]01:09
-!- ptizoom_ [~christian@] has joined #shogun01:09
blackburnIT WORKS01:47
blackburnwiking: 1.63s to solve 15k x 15k eigenproblem01:51
blackburnthat means when I get covertree for nearest neighbor queries I can get to quasilinear embedding cost like 5s for enormous number of vectors01:52
blackburnstability issues though :D01:53
-!- ptizoom_ is now known as ptizoom07:55
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]08:15
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has joined #shogun08:45
ptizoomsonney*: shall /perl directory (from shogun root) be used for building and exporting a perl-shogun package (to CPAN)?10:10
blackburnsonne|work: let me cite you a VW code line10:29
blackburnfor (vector<string>::iterator i = all.pairs.begin(); i != all.pairs.end();i++)      {10:29
sonne|workptizoom: shogun/src or?10:33
blackburnsonne|work: I want to set VW up to date10:37
blackburnsonne|work: do you think I should discuss with john?10:37
sonne|workhe would probably be happy about that10:38
sonne|workbut he is NY10:38
blackburnsonne|work: well hurricane will end someday10:38
sonne|worknot yet awake I mean10:38
blackburnwell I use email usually10:39
blackburnso you say he would be happy?10:39
blackburnI am not really in what you have done with shashwat10:40
blackburnit'd be really cool if we just depend on vw10:40
sonne|worklonger project then10:40
blackburnsonne|work: is there any reason to really include the codE?10:41
blackburnmay be I could help them to create an API I don't know10:41
blackburnsonne|work: they have they own development line - so hard to track it10:42
blackburnliblinear is not being really developed anymore, vw is10:42
sonne|workno feel free10:44
sonne|workbut lot of work10:44
blackburnsonne|work: okay10:45
blackburnwill write a mail later10:45
blackburnjohn's APIs are pretty cumbersome for me though10:46
blackburnI remembed all these struggles with the covertree10:46
blackburnI never managed to get it working, only fernando did :D10:46
sonne|workwell then maybe better choose some other fun project :D10:47
blackburnsonne|work: well I got interested10:48
blackburnsonne|work: btw I kinda developed a new method for eigendecomposition10:48
blackburnsonne|work: you've seen that probably in last night log10:48
blackburnit was working already but it took 3 days to realize that kernel distance is k(x,x) + k(y,y) - 2 k(x,y)10:49
blackburnand not k(x,y)10:49
blackburnso we will get some fun stuff soon10:50
blackburnboth in standalone and shogun versions10:50
blackburnsonne|work: can you stand 3 headers with high std and eigen usage?10:51
blackburnflexibility was one of my intentions so I have to live with it10:51
blackburneigen bloats code a little btw10:52
blackburnbut it is like you have intel MKL with no intel MKL10:52
-!- n4nd0 [] has joined #shogun11:54
ptizoomsonney*: from root...together with data,doc,examples,src, application... then perl! unless you find a good place to stuff a Makefile.PL and a t directory?13:15
sonne|workblackburn: can't you avoid using eigen/std in headers?13:26
sonne|workptizoom: well I guess at some point we should just move the src/Makefile and configure and everything one layer up but for now13:27
sonne|workin src13:28
ptizoomsonne*: okey dookey, I build a /src/perl ; typically I guess one put directories like src, perl, python, swig, debian etc... in the root.13:32
n4nd0can someone tell me why #include <eigen3/Eigen/...> doesn't work?13:37
n4nd0I have /usr/local/include/eigen313:37
n4nd0and also /usr/include/eigen3 although I don't really know why both of them...13:38
n4nd0I made a symbolic link of /usr/local/include/eigen3/Eigen in /usr/local/include and then I got to compile with the directive #include <Eigen/...>13:38
n4nd0but still, I do not get why #include <eigen3/Eigen/...> fails13:38
sonne|workn4nd0: maybe eigen3 is including sth with <Eigen/xxx> too?13:48
sonne|workptizoom: if it doesn't make sense we will change it later13:48
n4nd0sonne|work: what do you mean?13:48
n4nd0eigen3 is directory in /usr/local/include that includes another directory called Eigen13:49
sonne|workyeah but /usr/local/include/eigen3 is not in the include path14:36
sonne|workso when you include <eigen3/Eigen/sth>14:36
sonne|workand sth includes <Eigen/whateve> it cannot find whatever14:36
-!- n4nd0 [] has quit [Quit: leaving]14:47
wikingah shit17:17
wikingbut yeah he is doing it wrong17:17
wikinghe should include -issytem /user/local/include/eigen317:17
wikingand then #include <Eigen...17:17
-!- zxtx [] has joined #shogun17:50
-!- blackburn_ [d5578acb@gateway/web/freenode/ip.] has joined #shogun17:51
blackburn_sonne|work: I can hide them to cpp adapter17:52
blackburn_sonne|work: but there are templates17:52
blackburn_not specialized so can't put realization anywhere17:53
blackburn_sonne|work: that's all negotiable though but I still like functors more than crazy function pointers17:55
blackburn_so not really sure I can avoid templating17:55
blackburn_sonne|work: what I am sure is that all libraries should use reverse interface17:59
-!- fmailhot [] has joined #shogun18:01
fmailhotHi all...what's the appropriate Python interface for multiclass logistic regression trained via SGD? (assuming Shogun has this)  I'd like to do some Shogun/scikit-learn comparisons...18:09
blackburn_fmailhot: I have bad news - your assumption is a little wrong :)18:13
blackburn_fmailhot:but  modular is the only interface one should use18:13 it.18:14
fmailhotAnd...multiclass SGD-trained LR?18:14
blackburn_fmailhot: you need exactly that method?18:15
blackburn_fmailhot: I can suggest you to try MulticlassLibLinear18:16
blackburn_L2 regularized SVM Crammer-Singer that is18:16
fmailhotI'm hoping for something that can be trained online, in a streaming data kind of scenario...18:18
blackburn_fmailhot: LaRank18:18
fmailhotOh sorry, and most to the point, I'd like well-calibrated probabilities as outputs (hence my initial request for LR)18:19
blackburn_fmailhot: ah I see18:19
blackburn_fmailhot: if you can still live with batch mode I have one more suggestion18:20
* fmailhot is all ears18:20
blackburn_a minute18:21
blackburn_forgot the name of class I wrote :D18:21
blackburn_fmailhot: it is tree group lasso multiclass logistic regression18:22
blackburn_yeah it is funny we have such complex algorithms and have no SGD18:22
blackburn_fmailhot: I think I can implement basic multiclass logistic regression in next 2-3 days if it is required18:23
fmailhotCool, thanks. I'll have a look at that. And if you want to implement multiclass LR, I'd certainly be happy about it!  :D  (obviously I understand that you & other contributors are busy, though)18:24
blackburn_fmailhot: well I can find time if I know what to do - usually I am a little lost18:24
blackburn_fmailhot: IIRC basic regression is just two lines patch for that tree one18:25
fmailhotOK, cool...I'll do some experimenting with that and come back w/ questions as they arise. Thanks again!18:25
blackburn_fmailhot: I just checked - yes I'd have to change regularizer18:28
blackburn_fmailhot: but as it would be highly experimental I give no warranty :D18:28
fmailhot:)  Fair enough18:29
blackburn_fmailhot: I'll try, test it a little, commit and let you know once I get home18:30
blackburn_so 2-3 hours18:30
fmailhotNo need to knock yourself's only 10:30AM here in California.18:31
fmailhotAnd tomorrow is fine, too.18:31
fmailhotThanks very much for this!18:31
blackburn_fmailhot: I will forget in other case :D18:31
blackburn_fmailhot: 21:31 here heh18:31
fmailhotAh...heh. OK then.18:32
blackburn_fmailhot: I am from that snowy dark country18:33
blackburn_fmailhot: solver is pretty compact18:34
fmailhotCool, thanks.18:35
-!- blackburn_ [d5578acb@gateway/web/freenode/ip.] has quit [Quit: Page closed]19:10
-!- blackburn1 [~blackburn@] has joined #shogun19:42
blackburn1fmailhot: still here?20:48
-!- fmailhot is now known as fmailhot_lunch20:50
fmailhot_lunchHi yes, but about to run out for lunch20:50
blackburn1fmailhot_lunch: just wanted to say I started implementing the method20:51
fmailhot_lunchAwesome, thanks!20:51
fmailhot_lunchWill be back online in ~1hr. But like I said, no crazy rush...tomorrow would be fine.20:52
-!- fmailhot_lunch [] has quit [Quit: Leaving.]20:53
blackburn1I know :) I just don't want to get back to it later20:55
blackburn1fmailhot: see following commits - python example inside21:32
-shogungit:#shogun- [shogun] lisitsyn pushed 2 new commits to master:
-shogungit:#shogun- shogun/master 1526e97 Sergey Lisitsyn: Introduced multiclass logistic regression21:32
-shogungit:#shogun- shogun/master 3222c06 Sergey Lisitsyn: Added python example of multiclass logistic regression21:32
-shogungit:#shogun- [shogun] lisitsyn pushed 1 new commit to master:
-shogungit:#shogun- shogun/master 5fa8df4 Sergey Lisitsyn: Updated reference to submodule tutorial21:35
-!- n4nd0 [] has joined #shogun21:57
shogun-buildbotbuild #615 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Sergey Lisitsyn <>22:17
-!- fmailhot [] has joined #shogun22:44
-!- n4nd0 [] has quit [Quit: leaving]22:46
blackburn1fmailhot: feel free to try22:47
fmailhotCool, thanks...caught up in other things at the moment, but will definitely have a look.22:48
blackburn1fmailhot: please let us know if it is working for you - sorry don't have much time for thorough testing22:49
fmailhotNo worries, I'll definitely come back w/ a report tomorrow, Friday at the latest. Thanks again very much!22:50
blackburn1fmailhot: are you a researcher btw?22:50
blackburn1oh that's even more cool22:50
fmailhotBut doing "research" for the moment...22:51
fmailhotFiguring out which libs are best for us22:51
blackburn1hah I see - we need feedback so if you 'choose' us and have time share your ideas22:51
fmailhotSure thing...I'll give you guys feedback either way.22:52
shogun-buildbotbuild #616 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Sergey Lisitsyn <>22:52
ptizoomsnow and sand22:53
blackburn1ptizoom: oh it is you with your poetry :)22:53
ptizoomwhen industry of silicates spread dust over the infinite snowy steps of the east22:55
blackburn1I have no idea what you're talking about :D22:56
ptizoomme too22:56
-shogungit:#shogun- [shogun] lisitsyn pushed 1 new commit to master:
-shogungit:#shogun- shogun/master 276a2ff Sergey Lisitsyn: Made multiclass logistic regression python example safe when shogun is compiled without eigen322:57
* fmailhot thinks he's stepped into a surrealist IRC performance art piece...22:58
ptizoomI see shogun is going well...I try to learn it also bit by bit... there are lots of things in it, incredible.22:59
blackburn1just a few algorithms23:00
ptizoomu r 2 modest23:00
ptizoomactually I wish there was already a group like this in the nooties... it would have simplified research no end...23:02
blackburn1not really sure what you mean23:03
blackburn1I failed turing test you know23:04
ptizoomwell, those ML are wonderful to organise trial and errors...and find converging algo...n'est ce pas?23:08
ptizoomand with a package like this, it gives no excuses, not to try...23:08
blackburn1well I see ML quite stupid actually :D23:08
ptizoomthat the beauty!23:09
ptizoomstupid and beautiful23:09
ptizoomhow many ll in hallowin?23:10
blackburn1come on, what's that again :D23:11
ptizoommade pictures and need to label a directory with this name...ok, no help from this channel... lets gg it.23:12
ptizoom2ll and 2ee....23:14
-!- blackburn1 [~blackburn@] has quit [Quit: Leaving.]23:24
shogun-buildbotbuild #617 of deb3 - modular_interfaces is complete: Success [build successful]  Build details are at
-!- fmailhot [] has left #shogun []23:45
--- Log closed Thu Nov 01 00:00:17 2012