IRC logs of #shogun for Sunday, 2012-09-23

« Saturday, 2012-09-22 Index Monday, 2012-09-24 »
--- Log opened Sun Sep 23 00:00:17 2012
CIA-31shogun: Sergey Lisitsyn master * r9a174ee / (2 files in 2 dirs): Fixed compressed string features crasher and a few more potential memory errors - http://git.io/H2c2nA00:03
-!- blackburn1 [~blackburn@31.28.33.88] has quit [Quit: Leaving.]00:15
shogun-buildbot_build #542 of deb3 - modular_interfaces is complete: Failure [failed test ruby_modular]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/542  blamelist: Sergey Lisitsyn <lisitsyn.s.o@gmail.com>00:49
-!- ptizoom_ [55d2543d@gateway/web/freenode/ip.85.210.84.61] has quit [Ping timeout: 245 seconds]02:45
shogun-buildbot_build #109 of nightly_default is complete: Failure [failed test]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_default/builds/10904:19
-!- ptizoom [55d2543d@gateway/web/freenode/ip.85.210.84.61] has quit [Ping timeout: 245 seconds]07:03
-!- hoijui [~hoijui@dslb-088-074-107-101.pools.arcor-ip.net] has joined #shogun10:07
-!- blackburn [~blackburn@31.28.33.88] has joined #shogun11:45
wikingblackburn: yo11:46
blackburnwiking: hey! how are you?11:47
wikingblackburn: sorry i haven't seen your message on time the other day11:47
blackburnah we are all busy nevermind :)11:47
wikingyeah i'm recovering now from my wedding :)11:47
wikingbut now i'm more or less back11:47
blackburnoh so we are congratulating you?11:48
blackburn:)11:48
wikinghave u seen michal around lately?11:48
wikingyeye :)))11:48
blackburnno he rather disappear11:48
wiking:(11:48
blackburnis it your first time? :)11:48
wikinghehehe first and last i hope11:48
wikingi have some problems with the dualsolver11:48
wikingmaybe i should email him :S11:49
blackburnbmrm?11:49
wikingyeps11:49
blackburnwhat's wrong?11:49
wikingwell11:49
wikingthe whole thing starts with w = 011:49
wikingand basically when i give it a problem to solve11:49
wikingit doesn't converge to anywhere :(11:49
wikingw is zero11:49
wikingand after 2 iteration the whole thing stops11:50
blackburndid you tune lambda?11:50
wikinghaven't touched yet that11:50
wikingbut then again11:50
wikingit's weird11:50
blackburnoh that's the reason then11:50
wikingmmm let me try11:50
wikingdo u have a good lambda?11:50
blackburnno idea, just try some small/large11:51
blackburnthe reason you get w=0 is that you overregularize that11:51
blackburnso the algorithm prefers w=0 rather than avoid any errors on train set11:51
wikinglemme check what was the lambda11:52
wikingok so lambda =  0.100000 is not good11:52
wikingi'll try now with 1011:52
blackburnlambda is 1/C for svms11:54
blackburnIIRC11:54
wikingyeps11:54
blackburnno idea if #count counts here as well11:54
wikingok so neither 0.001 nor 10.0 does anything11:54
blackburntry some extreme values like 1e-10 and 10e1011:55
wikingi think i still have some other problem11:55
wikingbasically w keeps being 011:55
blackburnwhat is the model?11:55
wikingok lemme copy-paste the risk function11:56
wiking\lambda||\mathbf{w}||^2 + \sum_{i=1}^n \left( \max_{(\hat{y},\hat{h}) \in YxH} \left[ \mathbf{w}  \cdot \Psi(x_i, \hat{y}, \hat{h})+\Delta(y_i, \hat{y}, \hat{h}) \right ] \right )  - \sum_{i=1}^n \mathbf{w} \cdot \Psi(x_i, y_i, h^*_i)11:57
wikingput it here for easy reading11:57
wikinghttp://www.codecogs.com/latex/eqneditor.php11:58
blackburnso latent thing again :)11:58
wikingbasically the last term is = 0 because of w=011:58
wikingand the 2nd term is -24511:58
wikingand that's it11:58
wikingit doesn't move anywhere11:58
blackburncan you paste  the thing BMRM output?11:59
wiking[DEBUG] risk: -245.000000 0.000000 0: tim=0.545, Fp=-245.000000, Fd=-inf, R=-245.00000011:59
wiking[DEBUG] risk: -245.000000 0.000000 1: tim=0.550, Fp=-245.000000, Fd=-0.000000, (Fp-Fd)=-245.000000, (Fp-Fd)/Fp=1.000000, R=-245.000000, nCP=1, nzA=0, QPexitflag=111:59
blackburnI find that confusing to have exactly 245.012:00
wikingit's because of \Delta(y_i, \hat{y}, \hat{h})12:01
wikingbuuuut12:01
wikingi think i'm stupid :))))12:01
wikingyes12:01
blackburnah it is hamming-like?12:01
blackburnI see12:01
wikingyeah12:01
wikingbut12:01
wikingi did12:01
wikingw*psi - delta12:02
wikinginstead of12:02
wikingw*psi + delta12:02
wikingin the 2nd term12:02
blackburnheh12:02
wikingdunno why12:02
blackburntry brand new thing then :)12:02
wikingtrying with 0.01 lambda12:03
wikinghahahah much better12:03
wikingit's already at the 42th iteration12:03
wiking5512:03
wikingand counting12:03
blackburnlets hope it converge12:03
wikingwell now fp went down to 812:04
wikingfrom 245.00000012:04
wiking;)12:04
blackburnwiking: so something promising now?12:10
wikingiteration 450 ;)12:10
blackburnehh what's dual gap?12:10
wiking 489: tim=1.278, Fp=2.828893, Fd=2.449111, (Fp-Fd)=0.379781, (Fp-Fd)/Fp=0.134251, R=0.379781, nCP=223, nzA=214, QPexitflag=112:11
blackburnuh12:12
blackburntight12:12
wikingok maybe i should change lambda12:14
blackburnwiking: I expect either problem is hard to solve or you've got hard to solve parameters12:14
-!- hoijui [~hoijui@dslb-088-074-107-101.pools.arcor-ip.net] has quit [Quit: Leaving]12:43
wikingwtf12:45
wikingthis shit took about 7gigs of ram12:45
blackburnwow13:03
wikingyeah i think i should valgrind the shit out of it ;)13:50
CIA-31shogun: Sergey Lisitsyn master * re0a93ad / (3 files in 2 dirs): A few more stringfeatures improvements - http://git.io/bHbEEA15:29
wikingyey i was leaking15:31
wikingand am still but not that much at least15:31
blackburnwiking: what is leaking?15:32
blackburnyour code?15:32
wikingyeah15:32
wikingmmm when should this converge...ever? :)15:36
blackburnshould converge in 200-300 iters in a normal way15:37
blackburnas I believe I mean15:37
blackburnwell15:37
blackburnthat's all a bit strange15:37
wikingyeah i'm at iter 90015:37
wikingtrying with different lambda15:40
wikingblackburn: something for reading: http://www.computer.org/portal/web/csdl/doi/10.1109/TPAMI.2012.197 :)15:44
blackburnpdf unavailable heh15:45
wikingmmm15:45
blackburnI guess you've got some university access for that?15:46
wikinghaven't tried15:46
wikingi've just read the abstract ;)15:46
wikingi'll check it15:46
blackburnwiking: I'm now attacking stereo matchin' problem15:51
wikinghiih15:52
blackburnsurprisingly that's kind of hard to be real-time thing15:52
wikingyep15:52
blackburnbut I have to make it realtime15:52
wikingoh bmrm why u no converge!15:52
blackburnbmrm can haz converge I am sure :)15:53
wikingyeah i think i need something like this15:53
wikingbut with bmrm15:53
wikinghttp://images.memegenerator.net/instances/500x/9090678.jpg15:53
wiking:D15:53
blackburnso risk is decreasing and decreasing?15:53
blackburnheh15:53
blackburnand decreasing once again15:53
wikingyeah indeed it does15:53
wiking0.16879815:54
wikingthis is the risk atm :D15:54
blackburnhow much it decreases iteration by iteration?15:54
wiking0.0115:54
blackburnhmm15:54
wikingit's basically moving around 0.15-0.1715:54
wikingfor the last 10 iterations15:54
blackburnmay be you could try wider tolerance?15:55
blackburnjust to check it converges to something consistent15:55
wikinghow? which parameter?15:55
wikingas now i'm using everything default15:55
wikingexcept lambda15:55
blackburnlemme check15:55
wikinglambda is 0.0115:55
blackburnwiking: is your class built on top of dual lib qp bm so svm?15:56
wikingsortof15:56
blackburnset_TolRel15:56
blackburn1e-3 by default15:56
wikingso i guess 0.1 would be nicer than :)15:57
blackburntry that15:57
wikingok doin'15:57
wikingok risk is still going down with 1.015:58
wikingmmm15:59
wikingstill nada16:00
wikingi have like 122 cutting planes16:00
wikingthe risk is moving between 1.3-1.416:01
shogun-buildbot_build #543 of deb3 - modular_interfaces is complete: Failure [failed test ruby_modular]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/543  blamelist: Sergey Lisitsyn <lisitsyn.s.o@gmail.com>16:13
wikingoh woah16:15
wikingafter iter  896: tim=1.641, Fp=2.754775, Fd=2.489348, (Fp-Fd)=0.265427, (Fp-Fd)/Fp=0.096352, R=0.265427, nCP=272, nzA=265, QPexitflag=116:15
wikingit converge16:15
wikingd16:15
blackburnwiking: anything consistent or?16:29
shogun-buildbot_build #110 of nightly_default is complete: Failure [failed test]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_default/builds/11016:58
-!- blackburn [~blackburn@31.28.33.88] has quit [Quit: Leaving.]17:20
-!- heiko [~heiko@host86-180-121-44.range86-180.btcentralplus.com] has joined #shogun18:09
heikosonney2k, around?18:10
-!- geenoc [9c389a0c@gateway/web/freenode/ip.156.56.154.12] has joined #shogun18:24
-!- geenoc [9c389a0c@gateway/web/freenode/ip.156.56.154.12] has quit [Client Quit]18:26
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun18:48
-!- heiko [~heiko@host86-180-121-44.range86-180.btcentralplus.com] has left #shogun []18:49
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Ping timeout: 248 seconds]20:36
CIA-31shogun: Sergey Lisitsyn master * r2250ce5 / src/interfaces/ruby_modular/swig_typemaps.i : Fixed ruby string typemaps - http://git.io/T1gxxQ21:04
shogun-buildbot_build #544 of deb3 - modular_interfaces is complete: Failure [failed test lua_modular]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/544  blamelist: Sergey Lisitsyn <lisitsyn.s.o@gmail.com>21:47
-!- blackburn [~blackburn@188.168.2.50] has joined #shogun22:27
--- Log closed Mon Sep 24 00:00:17 2012
« Saturday, 2012-09-22 Index Monday, 2012-09-24 »

Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!