--- Log opened Sun Jun 17 00:00:41 2012 | ||
blackburn | Ok finally first multitask regressor read | 00:04 |
---|---|---|
CIA-18 | shogun: Sergey Lisitsyn master * rf621fdf / (14 files in 4 dirs): Updated TaskGroup, added train routines for MultitaskLS regression and fixed some bugs in SLEP ported code - http://git.io/Cb8x1A | 00:07 |
CIA-18 | shogun: Sergey Lisitsyn master * rafe730c / (14 files in 4 dirs): Merged SLEP updates - http://git.io/yJ0Baw | 00:07 |
n4nd0 | nice | 00:14 |
blackburn | n4nd0: I'm now porting algo that does SVD each iteration (sic!) | 01:12 |
n4nd0 | blackburn: and how is it going? | 01:15 |
blackburn | going ok but SVD! | 01:15 |
blackburn | each iteration! | 01:15 |
blackburn | :D | 01:15 |
n4nd0 | too much cost right? | 01:16 |
blackburn | N^3 | 01:17 |
shogun-buildbot_ | build #632 of octave_modular is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/octave_modular/builds/632 | 01:17 |
blackburn | well the only exciting thing about svd is its robustness | 01:17 |
blackburn | with cond=1 it is very stable | 01:18 |
n4nd0 | ok | 01:18 |
-!- emrecelikten [~emre@176.40.254.205] has joined #shogun | 01:32 | |
shogun-buildbot_ | build #608 of lua_modular is complete: Failure [failed test_1] Build details are at http://www.shogun-toolbox.org/buildbot/builders/lua_modular/builds/608 blamelist: blackburn91@gmail.com | 02:06 |
shogun-buildbot_ | build #633 of octave_modular is complete: Failure [failed compile] Build details are at http://www.shogun-toolbox.org/buildbot/builders/octave_modular/builds/633 blamelist: blackburn91@gmail.com | 02:12 |
-!- blackburn [~blackburn@62.106.111.221] has quit [Ping timeout: 244 seconds] | 02:25 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Ping timeout: 246 seconds] | 02:44 | |
-!- emrecelikten [~emre@176.40.254.205] has quit [Quit: Leaving.] | 06:35 | |
-!- emrecelikten [~emre@176.40.254.205] has joined #shogun | 08:05 | |
-!- emrecelikten [~emre@176.40.254.205] has quit [Quit: Leaving.] | 09:33 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 10:35 | |
-!- blackburn [~blackburn@62.106.111.221] has joined #shogun | 10:52 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Ping timeout: 240 seconds] | 11:29 | |
-!- blackburn [~blackburn@62.106.111.221] has quit [Ping timeout: 248 seconds] | 12:13 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 12:39 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: Lost terminal] | 13:53 | |
-!- emrecelikten [~emre@176.40.254.205] has joined #shogun | 17:42 | |
-!- blackburn [~blackburn@62.106.111.221] has joined #shogun | 18:37 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 19:42 | |
blackburn | n4nd0: hey any progress with your bug? | 19:48 |
n4nd0 | blackburn: I changed a couple of things and the results improved, but I don't find satisfactory yet | 20:02 |
blackburn | I see | 20:05 |
n4nd0 | blackburn: but the way, let me ask you something | 20:07 |
n4nd0 | general SVM theory :D | 20:07 |
blackburn | I've been waiting for this moment whole my life ;) | 20:07 |
n4nd0 | haha | 20:08 |
n4nd0 | I have seen that in the paper they write | 20:08 |
n4nd0 | min_w,\xi 1/2 ||w||^2 + C/n sum_i \xi_i | 20:08 |
n4nd0 | http://www.cs.cornell.edu/people/tj/publications/finley_joachims_08a.pdf | 20:08 |
n4nd0 | if you prefer to look at it ^ | 20:08 |
blackburn | so the question is C/n? | 20:09 |
n4nd0 | page 2, equation 1, near to the end of the first column | 20:09 |
n4nd0 | it is 1/n in particular | 20:09 |
blackburn | yes found | 20:09 |
n4nd0 | I am not using that currently | 20:09 |
n4nd0 | I don't know if it should make a difference | 20:09 |
blackburn | it doesn't make difference - just changes C implicitly | 20:10 |
blackburn | what changes the game is \sum_i C_i \xi_i | 20:10 |
n4nd0 | why C_i? | 20:10 |
blackburn | it can be done with differents C's | 20:11 |
n4nd0 | one for each example | 20:11 |
blackburn | yes | 20:11 |
n4nd0 | I've never seen that, but it's ok | 20:11 |
blackburn | but while you just play with constant before the sum | 20:11 |
blackburn | I don't think it makes and difference | 20:11 |
blackburn | I think it is easier for they to prove things with 1/n | 20:12 |
n4nd0 | I am going to use how liblinear does it internally | 20:13 |
blackburn | n4nd0: does what? | 20:16 |
n4nd0 | use 1/n | 20:16 |
blackburn | they do not use 1/n | 20:16 |
blackburn | I am 99% sure they don't | 20:16 |
n4nd0 | why? | 20:17 |
blackburn | n4nd0: because finley and joachims introduced that I think :) | 20:17 |
n4nd0 | the 1/n | 20:18 |
n4nd0 | ? | 20:18 |
blackburn | standart C-S does not involve 1/n | 20:18 |
blackburn | yes | 20:18 |
blackburn | neither W-W svm does | 20:18 |
n4nd0 | ok | 20:19 |
blackburn | n4nd0: liblinears multiclass svm is just kind of gradient descent with cooling | 20:20 |
n4nd0 | cooling? | 20:20 |
blackburn | simply they just exclude alphas that zeroed | 20:21 |
blackburn | and work only with subset of alphas | 20:21 |
n4nd0 | ok | 20:22 |
blackburn | ocas does cutting plane btw | 20:22 |
n4nd0 | yes | 20:23 |
@sonney2k | n4nd0, the 1/n is kind of a relegion | 20:30 |
@sonney2k | some use it some don't | 20:30 |
@sonney2k | n4nd0, rationale is that the \sum_i \xi_i is an upper bound on the training error | 20:31 |
@sonney2k | if you divide by n it is the risk aka fraction of misclassified examples | 20:32 |
@sonney2k | bringing that on the same scale as the ||w||_2^2 makes some sense | 20:32 |
@sonney2k | however, I find it not very intuitive in practice at least: | 20:33 |
@sonney2k | C does not scale linearly with n | 20:33 |
@sonney2k | n4nd0, btw what is up with your pull request? you never answered my comment... | 20:36 |
n4nd0 | sonney2k: yeah sorry, since there are still some things that I have to correct, apart from the renaming and re-allocating of the files you told me, I thought it was clear that it was not a definitive version to merge yet | 20:39 |
n4nd0 | I don't know if you have heard about my fight against the bugs during this week ... | 20:40 |
n4nd0 | I have not really coded anything new this week, I am just testing, debugging, testing, debugging, etc | 20:40 |
n4nd0 | :S | 20:40 |
blackburn | sonney2k: mldata is not so cool! I never managed to find some popular datasets here :( | 20:40 |
n4nd0 | sonney2k: btw, you never answered either my mail on the mailing list about the issue with the modular interface :( | 20:41 |
blackburn | O! | 20:44 |
blackburn | did somebody review puffin444? | 20:45 |
blackburn | I did | 20:53 |
n4nd0 | nice | 20:54 |
@sonney2k | n4nd0, it does not have to be ready as in everything is 100% - it needs to compile and uricamic/wiking can work with the infrastructure | 20:59 |
@sonney2k | n4nd0, so it is really highly needed | 20:59 |
@sonney2k | blackburn, what dataset is popular/missing? | 21:00 |
@sonney2k | blackburn, mldata just needs more users/uploaders... | 21:00 |
blackburn | sonney2k: today I did not manage to find 'school' dataset | 21:00 |
blackburn | for multitask | 21:00 |
@sonney2k | never heard of it | 21:00 |
@sonney2k | so upload it then :D | 21:01 |
blackburn | yeah if I had | 21:01 |
blackburn | it | 21:01 |
blackburn | sonney2k: they compare on school dataset in every damned multitask paper | 21:01 |
blackburn | :D | 21:01 |
n4nd0 | sonney2k: what I commited and introduced in the pull request compiles and doesn't crash anything | 21:34 |
n4nd0 | sonney2k: but I will introduce the changes you told me aobut asap so you can merge | 21:35 |
n4nd0 | sonney2k: anyway, I am sorry to ask you again, did you have time to take a look to the mail I sent about the modular interface? | 21:35 |
n4nd0 | blackburn: I am starting to be happy with my results :) | 22:00 |
n4nd0 | blackburn: at least they are similar on average with the ones given by multiclass liblinear not using bias | 22:01 |
n4nd0 | blackburn: around? | 22:11 |
blackburn | n4nd0: yes | 22:36 |
blackburn | n4nd0: that's cool | 22:37 |
n4nd0 | blackburn: I think I am going to include the possibility to include bias in LinearSOMachine | 22:37 |
blackburn | regularized bias? | 22:37 |
n4nd0 | blackburn: if the results of multiclass SO improved more (as it does with LibLienar) then I will be more or less convinced that the impl. is ok | 22:38 |
n4nd0 | blackburn: I was thinking of including bias, not regularization as this moment | 22:38 |
n4nd0 | blackburn: I think that the way it's done in SO, regularization would be included using the matrix C | 22:39 |
n4nd0 | in | 22:39 |
n4nd0 | min w'*C*w | 22:39 |
blackburn | n4nd0: is it possible to include bias in that kind of formulation? | 22:39 |
n4nd0 | why not? | 22:39 |
blackburn | well how? | 22:40 |
n4nd0 | adding one more dimension to w | 22:41 |
n4nd0 | and the same for the feature vectors | 22:41 |
n4nd0 | the one in w is to be learnt and in the feature vectors fixed to 1 | 22:42 |
blackburn | n4nd0: it is regularized bias ;) | 22:43 |
n4nd0 | where is the regularization there? | 22:43 |
blackburn | it goes to regularization term | 22:43 |
blackburn | ||w|| | 22:43 |
n4nd0 | ok | 22:43 |
blackburn | so when you optimize w you try to keep b low too | 22:44 |
n4nd0 | agree | 22:44 |
blackburn | it is not bad (Mangasarian, 1992) but it is | 22:44 |
blackburn | just is :D | 22:44 |
n4nd0 | haha | 22:44 |
n4nd0 | is what? | 22:44 |
blackburn | it is regularized I mean :) | 22:44 |
blackburn | actually our dear vojta did that in his fancy svm | 22:45 |
blackburn | GMNP one with Kesler transformation and a lot of other stuff | 22:45 |
blackburn | n4nd0: haven't you been watching por-ned? | 22:46 |
n4nd0 | blackburn: they showed germany-denmark here, not por-ned | 22:46 |
blackburn | too bad | 22:46 |
blackburn | ronaldo was fscking awesome | 22:47 |
n4nd0 | he's fucking awesome :) | 22:47 |
blackburn | portugal has very strong sprinter skills I am amazed with | 22:47 |
blackburn | I bet they do these 50-60 meters in 6-7 seconds | 22:48 |
n4nd0 | Greece-Germany in quarter finals | 22:49 |
n4nd0 | funny fact :D | 22:49 |
blackburn | hmm | 22:49 |
blackburn | Russia-Germany would be funnier | 22:49 |
n4nd0 | no way | 22:49 |
blackburn | 22th June | 22:49 |
blackburn | the day war began | 22:49 |
blackburn | :D | 22:49 |
n4nd0 | Greece - Germany with everything about the economic problems, seems fair enough to me | 22:49 |
blackburn | heh well war is more important I believe :D | 22:50 |
blackburn | 22th June, 1941 | 22:50 |
n4nd0 | aren't we talking about an old war dude? | 22:50 |
n4nd0 | some time has passed ever since | 22:50 |
blackburn | very important still for russia and germany too I think | 22:51 |
blackburn | in the end we have lost ~20M | 22:52 |
blackburn | a little less than half of spain :D | 22:53 |
n4nd0 | haha | 22:53 |
n4nd0 | Spain has nothing to do with this ;) | 22:53 |
blackburn | and two swedens | 22:53 |
blackburn | n4nd0: http://www.wired.com/wiredenterprise/wp-content/uploads//2012/06/beard-programmers-final-two.png | 22:54 |
blackburn | did you grow? | 22:54 |
n4nd0 | no ... I'm still getting shaved :) | 22:54 |
blackburn | now I do understand why I have not so outstanding skills | 22:54 |
n4nd0 | Captain Soeren was at Guru level in this moment http://videolectures.net/icml08_sonnenburg_ipsvm/ :D | 22:55 |
blackburn | yeah right | 22:56 |
blackburn | hmm | 22:57 |
blackburn | why do he call himself Kristian :D | 22:57 |
n4nd0 | who? | 22:57 |
blackburn | joke | 22:57 |
blackburn | in that video he is talking on behalf of some Kristian | 22:58 |
blackburn | :) | 22:58 |
n4nd0 | aha :) | 22:58 |
n4nd0 | good night! | 23:00 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 23:01 | |
--- Log closed Mon Jun 18 00:00:41 2012 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!