Open in new window / Try shogun cloud
--- Log opened Tue Mar 12 00:00:11 2013
-!- FSCV [~FSCV@206.225.135.24] has quit [Quit: Leaving]00:58
shogun-buildbotbuild #275 of nightly_none is complete: Failure [failed compile]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_none/builds/27503:01
-!- blackburn1 [~blackburn@188.168.14.162] has joined #shogun06:39
-!- blackburn [~blackburn@83.234.54.98] has quit [Ping timeout: 245 seconds]06:39
blackburn1sonney2k: nice song06:52
blackburn1:D06:52
blackburn1sonney2k: git://github.com/airblade/vim-gitgutter.git07:02
blackburn1(if you are still using vim)07:02
blackburn1:)07:02
-!- blackburn1 [~blackburn@188.168.14.162] has quit [Quit: Leaving.]07:21
-!- n4nd0 [~nando@n177-p170.kthopen.kth.se] has joined #shogun07:56
-!- blackburn [~lisitsin@212.46.246.30] has joined #shogun08:20
sonne|workblackburn: I am still undecided...09:01
blackburnsonne|work: a nice feature anyway09:04
sonne|workblackburn: yes indeed... I am already used to that from eclipse btw09:08
blackburnsonne|work: have you heard hinton is joining google?09:32
sonne|workinteresting. I guess he wants to create google brain(tm)09:34
blackburnsonne|work: yeah infinitely deep network09:38
sonne|worka c'mon a few million layers should be enough for everybody :)09:39
-!- sumit_ [73f91219@gateway/web/freenode/ip.115.249.18.25] has joined #shogun09:39
blackburnsonne|work: 640K09:40
sonne|worktrue that09:40
blackburnas a hipster teenager I should tweet that09:42
blackburnor may be I should write it down to my moleskine and instagram it with my iphone..09:43
blackburnsad I don't have either hah09:43
-!- heiko [~heiko@nat-168-237.internal.eduroam.ucl.ac.uk] has joined #shogun10:59
-!- sumit_ [73f91219@gateway/web/freenode/ip.115.249.18.25] has quit [Ping timeout: 245 seconds]11:36
blackburnheiko: okay that MDS thing must be changed12:12
heikoblackburn: ok will do tonight12:13
blackburnheiko: I will :)12:13
heikoblackburn:  just dicovered the coolness of Eigen3 :)12:13
blackburnheiko: hah congraatz12:13
heiko(compared to boost lina classes :)12:13
blackburnheiko: that's kind of late discover :)12:14
heikoblackburn:  well so far I did not need it12:14
heikoc++ array were sufficient12:15
blackburnheiko: explicit SSE is somewhat important speedup I think12:15
heikowhat is that?12:15
blackburnheiko: when you compute norm your CPU can handle two doubles per op12:16
heikoblackburn:  I see, thats good stuff12:16
blackburnheiko: eigen does all vectorization explicitly (when compiler can be afraid of doing that)12:16
heikoblackburn: what does that mean vectorization?12:17
blackburnheiko: SIMD (single instruction multiple data), i.e. SSE12:19
blackburnheiko: (at least SSE)12:19
heikoblackburn: I see12:19
heikocool12:19
heikogood stuff! :)12:19
blackburnheiko: some more magic of eigen is in template expression though12:20
heikoyeah this is what I meant12:20
heikothis is really cool and seems flexible12:20
blackburnheiko: a = b + c + d; means no copy12:20
blackburnas operator+ returns some thing that can be added to other vector and finally it is evaluated to a[i] = b[i] + c[i] + d[i] (but vectorized)12:21
n4nd0do you guys know if // means something special in python? I have found it in some code as an operator between two integers12:22
n4nd0I have seen it does the same as / for simple cases12:22
n4nd0i.e. integer division12:22
heikon4nd0: no idea :)12:23
heikon4nd0: but python has strange operators sometimes x**2 for example12:23
heikoto I guess // is related to that12:23
blackburnn4nd0:    3//1 => 312:23
n4nd0floor division!12:23
blackburnha12:23
blackburn   3//2.0 => 1.012:24
blackburnn4nd0: ^12:24
n4nd0I see, thanks!12:24
n4nd0so actually if both of them are integers, it doesn't make any difference wrt using /12:24
n4nd0more explicit though12:25
blackburnheiko: I took a look at your statistics code - pretty well organized12:30
heikoblackburn:  thanks :) what did you look at in particular?12:30
blackburnheiko: glanced over classes12:31
heikolinear time mmd is the best one ! :)12:31
blackburnheiko: pretty big12:32
heikoyes, it does a lot of things12:32
heikoremember: it can handle infinite amounts of data12:32
heiko(and do automatic optimal kernel selection for that)12:33
blackburnheiko: I still do not understand how to use it but that's nice12:33
heikoblackburn: on streams12:33
heikochangepoint detection for example12:33
heikobut to be honest, it is not that widely used :)12:34
blackburnheiko: what is changepoint detection?12:34
heikoyou have a stream of data and want to know when it changes12:34
heikoso imagine a brain-scanner12:34
heikoand you change your thought :)12:34
heiko(roughly)12:34
heikotest whether a distribution changes12:34
blackburnheiko: what would I need to pass to that algorithm?12:34
heikofirst one needs a change point detection algo12:35
heikothe MMD does the following12:35
heikoyou pass samples from two distributions p and q12:35
heikoand it tells you whether p!=q12:35
blackburnbut how is that changepoint?12:35
heikoyou can also use that to select kernels for classification since it will tell you which kernel maximises the difference12:36
heikoblackburn: you have to embed it into other things12:36
heikothere is a paper somewhere12:36
heikohow to do that12:36
heikoyou have like groups and test one against others12:36
blackburnheiko: so I first collect some normal data?12:36
blackburnheiko: I see some 'business' case12:37
heikoyes12:37
blackburnimagine I have a website12:38
blackburnand collect visitors data12:38
blackburnthat's P12:38
blackburnand then I change something in my adverstisement and collect more data12:38
blackburnthat's Q12:38
heikoyep thats a thing you could do12:39
blackburnso I define some kernel on visiting data and check if distribution of P and Q12:39
blackburnare equal12:39
blackburnquite interesting12:39
heikoblackburn:  yeah its a bit like classification12:39
blackburnheiko: I was recently asked about possible way to solve similar problem12:39
blackburnheiko: but P>Q12:39
heikobut more the qeustion: are there differences12:39
heikoyou mean number of samples?12:40
blackburnheiko: not really12:40
blackburnheiko: not correct formulation though :)12:40
blackburnheiko: guy needed to check if site with # of clicks P has more efficient adv than site with # of clicks Q12:41
blackburnor so12:41
blackburnheiko: he actually had only means so I said that's quite impossible if you just have means12:41
heikothats quite little information :)12:42
blackburnas he needed to do that with 0.95 confidence heh12:42
blackburnno way to compute anything with confidence without probability distribution I guess12:42
blackburnor at least variance..12:43
heikoyou need samples then you can do that12:44
blackburnheiko: I am actually unsure what is sample here13:02
blackburn# of clicks per some period may be13:03
-!- heiko [~heiko@nat-168-237.internal.eduroam.ucl.ac.uk] has quit [Quit: Leaving.]14:30
sonne|workblackburn: you seem to be so very excited about sse stuff in eigen - note that the standard gcc compiler will vectorize your simple for loops too so no need to be overly excited14:54
blackburnsonne|work: I am actually more excited about it gives you an API to do that stuff15:01
sonne|workyeah that is the benefit of it15:01
blackburnsonne|work: good code contains almost no loops15:01
blackburnsonne|work: it is not true sometimes - gcc don't care about aliasing15:02
-!- heiko [~heiko@nat-168-237.internal.eduroam.ucl.ac.uk] has joined #shogun15:02
blackburnsonne|work: so it could fail with that15:02
blackburnsonne|work: I am really unsure if gcc will be able to SIMDize some function which takes double* as a parameter15:04
sonne|workI would bet so15:04
blackburnsonne|work: anyway you can't say it does that in all cases that eigen does15:05
blackburnsonne|work: for example ICC vectorization is known to be more powerful (so there is a room)15:05
-!- heiko [~heiko@nat-168-237.internal.eduroam.ucl.ac.uk] has quit [Ping timeout: 252 seconds]15:07
blackburnsonne|work: about alignment issues: http://gcc.gnu.org/projects/tree-ssa/vectorization.html#assume-aligned15:07
blackburnI think there won't be such thing if it could determine that in 100% cases15:08
-!- FSCV [~FSCV@78.129.244.119] has joined #shogun15:44
wikingblackburn: i'm going to russian for 5 days... :)15:52
-!- blackburn [~lisitsin@212.46.246.30] has quit [Ping timeout: 245 seconds]16:00
-!- blackburn [~lisitsin@mxs.kg.ru] has joined #shogun16:14
-!- shogun-notifier- [~irker@7nn.de] has joined #shogun16:36
shogun-notifier-shogun: Sergey Lisitsyn :master * b6c25aa / examples/undocumented/python_modular/converter_multidimensionalscaling_modular.py: https://github.com/shogun-toolbox/shogun/commit/b6c25aa6ddcc3dd07f1d94f2970fb331bc9a870216:36
shogun-notifier-shogun: An attempt to get MDS example make more sense16:36
-!- tom____ [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has joined #shogun16:38
tom____hi!16:38
tom____I would like to use the sosvm framework in shogun16:38
tom____Is there any example of usage with sequential data ?16:39
-!- FSCV [~FSCV@78.129.244.119] has quit [Quit: This computer has gone to sleep]16:39
n4nd0hey tom____ !16:45
n4nd0nice to hear you want to use the sosvm framework :)16:45
tom____n4nd0: hey n4nd016:46
n4nd0tom____: take a look to examples so_hmsvm_mosek.cpp and so_hmsvm_mosek_simple.cpp in examples/undocumented/libshogun16:46
tom____n4nd0: then mosek need to be installed ?16:46
n4nd0tom____: and for the python interface, structure_hmsvm_*16:46
n4nd0tom____: not really16:47
n4nd0there is a solver that makes use of mosek16:47
n4nd0for the examples with mosek in the name you need it16:47
n4nd0but we have also some bundle methods implemented16:47
n4nd0look for example structure_hmsvm_bmrm.py in examples/undocumented/python_modular16:47
n4nd0sorry for the misleading name so_ in some of them - I will rename those to structure_16:48
tom____n4nd0: ok thank you16:54
n4nd0tom____: no problem16:54
tom____n4nd0: I use most of the time c++ then I will take a lokk to CDualLibQPBMSOSVM16:54
tom____that's the bundle isn't it ?16:54
n4nd0tom____: yeah16:54
-!- shogun-t1olbox [~shogun@7nn.de] has quit [Ping timeout: 264 seconds]16:54
--- Log closed Tue Mar 12 16:54:27 2013
--- Log opened Tue Mar 12 16:54:34 2013
-!- shogun-toolbox [~shogun@7nn.de] has joined #shogun16:54
-!- Irssi: #shogun: Total of 12 nicks [1 ops, 0 halfops, 0 voices, 11 normal]16:54
!wright.freenode.net [freenode-info] help freenode weed out clonebots -- please register your IRC nick and auto-identify: http://freenode.net/faq.shtml#nicksetup16:54
-!- Irssi: Join to #shogun was synced in 7 secs16:54
n4nd0tom____: there are also examples using the bundle method in the C++ interface16:54
n4nd0tom____: examples/undocumented/libshogun/so_multiclass_BMRM.cpp16:55
n4nd0that's not dealing with sequential data16:55
n4nd0but for that part you can always check the other examples16:56
tom____such ?16:56
n4nd0the ones I mentioned before16:56
n4nd0e.g. so_hmsvm_mosek16:57
n4nd0you can exchange models and solvers as you wish16:57
n4nd0so if you are interested in the bundle methods -> check things with bmrm16:58
n4nd0if you are interested in sequential data -> things with hmsvm16:58
n4nd0tom____: see what I mean?16:58
tom____n4nd0: allright16:58
tom____n4nd0: thank you16:59
n4nd0you are welcome16:59
-!- travis-ci [~travis-ci@ec2-54-234-212-155.compute-1.amazonaws.com] has joined #shogun17:03
travis-ci[travis-ci] it's Sergey Lisitsyn's turn to pay the next round of drinks for the massacre he caused in shogun-toolbox/shogun: http://travis-ci.org/shogun-toolbox/shogun/builds/544415517:03
-!- travis-ci [~travis-ci@ec2-54-234-212-155.compute-1.amazonaws.com] has left #shogun []17:03
n4nd0error in travis, "no space left on device"??!?17:04
n4nd0gtg now, see you later17:06
-!- n4nd0 [~nando@n177-p170.kthopen.kth.se] has quit [Quit: leaving]17:07
-!- tom____ [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has quit [Quit: Page closed]17:11
wikinglool17:27
wikingwtf is this17:27
-!- blackburn [~lisitsin@mxs.kg.ru] has quit [Quit: Leaving.]17:32
-!- heiko [~heiko@pat-191-250.internal.eduroam.ucl.ac.uk] has joined #shogun18:06
-!- zxtx [~zv@cpe-75-83-151-252.socal.res.rr.com] has quit [Ping timeout: 264 seconds]18:52
-!- blackburn [~blackburn@188.168.13.118] has joined #shogun18:53
-!- FSCV [~FSCV@187.210.54.166] has joined #shogun19:16
-!- FSCV_ [~FSCV@204.45.132.11] has joined #shogun19:20
-!- FSCV [~FSCV@187.210.54.166] has quit [Ping timeout: 245 seconds]19:21
-!- shogun-notifier- [~irker@7nn.de] has quit [Quit: transmission timeout]19:36
-!- FSCV_ [~FSCV@204.45.132.11] has quit [Ping timeout: 240 seconds]20:20
-!- FSCV_ [~FSCV@65.19.131.247] has joined #shogun20:22
-!- heiko [~heiko@pat-191-250.internal.eduroam.ucl.ac.uk] has left #shogun []21:56
-!- sumit [73f91219@gateway/web/freenode/ip.115.249.18.25] has joined #shogun22:05
-!- sumit [73f91219@gateway/web/freenode/ip.115.249.18.25] has quit [Client Quit]22:06
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun22:09
blackburnalright GPs are broken nice22:17
n4nd0blackburn: how badly?22:21
n4nd0what is it with them??22:21
blackburnn4nd0: heiko reports it's results are non-sense22:21
blackburnsonney2k: directors work here22:21
blackburn2.0.822:37
blackburntrying 2.0.922:37
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Ping timeout: 264 seconds]23:02
-!- shogun-notifier- [~irker@7nn.de] has joined #shogun23:10
shogun-notifier-shogun: Sergey Lisitsyn :master * f4696ee / src/shogun/evaluation/DirectorContingencyTableEvaluation.h: https://github.com/shogun-toolbox/shogun/commit/f4696ee7d06ba8ed097ed45ff74ac4c78ec12a4423:10
shogun-notifier-shogun: A fix for director contingency evaluation class23:10
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun23:12
n4nd0blackburn: I also tried out a couple of director examples here this morning23:13
n4nd0they worked fine23:13
blackburnn4nd0: evaluation one was wrong and I fixed it just a mniute ago23:14
n4nd0blackburn: nice23:14
blackburnn4nd0: about results23:17
blackburn(in paper)23:17
n4nd0blackburn: tell me23:18
blackburnn4nd0: I think I've got to try to optimize it23:18
n4nd0blackburn: what is the part that concerns you?23:18
blackburnn4nd0: I'd like to show >3x speedup always (that's the reason to parallelize code)23:18
blackburnn4nd0: I am surprised with speed of lle23:19
blackburnin drtoolbox23:19
blackburnn4nd0: 8 CPUs? really?23:19
n4nd0blackburn: that's what /proc/cpuinfo tells me23:20
n4nd0I can paste it if you want to check23:20
blackburnyeah could you please23:20
blackburnI'd rather expect 2 cpus with 4 cores23:21
n4nd0http://pastebin.com/GTVJNDwk23:22
n4nd0blackburn: well it may be23:22
n4nd0I have an i7 here23:22
n4nd0an proc/cpuinfo shows me 4 processors23:22
blackburnn4nd0: does it happen to have intel compiler around?23:23
n4nd0blackburn: how can I check that?23:23
blackburnn4nd0: X3470 is 4 core23:23
blackburnn4nd0: 'icpc' may be23:23
n4nd0comman not found23:23
n4nd0command*23:23
blackburnmay be anything in /opt/intel?23:24
n4nd0no directory23:24
blackburnn4nd0: hmm alright23:25
blackburnn4nd0: what is ubuntu version?23:26
n4nd0Ubuntu 12.04.2 LTS23:27
n4nd0precise pangolin IIRC23:27
blackburnn4nd0: and matlab 2011a, right?23:27
n4nd0yep23:28
blackburnalright thanks23:30
n4nd0blackburn: a pleasure :)23:30
n4nd0blackburn: so how can we speed it up?23:31
-!- travis-ci [~travis-ci@ec2-54-234-212-155.compute-1.amazonaws.com] has joined #shogun23:35
travis-ci[travis-ci] it's Sergey Lisitsyn's turn to pay the next round of drinks for the massacre he caused in shogun-toolbox/shogun: http://travis-ci.org/shogun-toolbox/shogun/builds/545510923:35
-!- travis-ci [~travis-ci@ec2-54-234-212-155.compute-1.amazonaws.com] has left #shogun []23:35
blackburnn4nd0: no idea actually23:35
blackburnn4nd0: what if you run scaling thing?23:35
n4nd0blackburn: what?23:36
n4nd0blackburn: ah sorry yeah23:36
blackburnn4nd0: python benchmark.py scaling23:36
n4nd0blackburn: let's do it23:36
blackburnn4nd0: I am curious what is $\infty$23:37
blackburnI'd avoid offensive claims that waffles didn't finish in 10 hours23:37
n4nd0Tapkee parallel implementations benchmark23:37
n4nd0Isomap on Swissroll with 1 threads takes 17.2111s23:37
n4nd0Isomap on Swissroll with 2 threads takes 9.5318s23:37
n4nd0Isomap on Swissroll with 3 threads takes 6.8726s23:37
n4nd0Isomap on Swissroll with 4 threads takes 5.5947s23:37
n4nd0Isomap on Swissroll with 5 threads takes 5.5965s23:37
blackburnbut it is true in the end..23:37
n4nd0blackburn: well it was not 10 hours23:38
n4nd0blackburn: or yes23:38
n4nd0blackburn: I don't actually remember it23:38
blackburnn4nd0: alright lets say 323:38
n4nd0mmm yeah yeah23:38
blackburn:D23:38
n4nd0it was 500 almost 600 minutes23:38
blackburnn4nd0: 100 dense matrix-vector products..23:40
n4nd0for each iteration23:40
n4nd0cannot be good no23:40
blackburnno I was wrong23:40
blackburnfor each dimension of target dimensions23:40
n4nd0so then it is even more than 100??23:41
blackburnI got a little lame23:41
blackburnn4nd0: 200 totally to be more precise23:41
n4nd0ah ok23:41
n4nd0isn't it a bit weird that it was taking that much then?23:41
blackburnn4nd0: but that should be computed faster than in 10 hours23:41
blackburnso I expect some even more high performance stuff is here too23:42
blackburnmay be he just waits I don't know23:42
blackburnThread.sleep23:42
n4nd0http://pastebin.com/aBvzbcYh23:42
blackburnn4nd0: hmm I see23:43
blackburnthnaks23:43
blackburnn4nd0: I'd start review of the paper23:43
n4nd0blackburn: our or waffles?23:44
blackburnn4nd0: haha our23:44
n4nd0I guess you mean ours23:44
n4nd0haha23:44
n4nd0blackburn: so should we revise the paper again and decide to send it?23:48
n4nd0blackburn: I thought we had already reviewed it :)23:48
blackburnn4nd0: just after plots and website23:48
n4nd0the review or the send?23:49
blackburnn4nd0: no, I mean we should read it carefully, meanwhile finishing website and some nice plots23:49
blackburnand then send23:49
n4nd0blackburn: agree23:50
n4nd0blackburn: did Chris contact you about something?23:50
blackburnand we should release shogun23:50
blackburn:D23:50
n4nd0haha23:50
blackburnn4nd0: no, I guess he is busy23:50
n4nd0and end with starvation in the world23:50
blackburnn4nd0: btw do you think people should stop funding africa?23:52
blackburn:)23:53
n4nd0blackburn: wtf?23:53
n4nd0:)23:53
n4nd0it is a delicate thing23:53
blackburnn4nd0: well countries spend a lot of money for nothing23:53
n4nd0nothing?23:53
n4nd0ha!23:53
n4nd0there are lot of governors there getting rich man23:53
n4nd0and what about the organizations who ask for the money?23:54
blackburnyeah that's the only thing they sponsor23:54
n4nd0:D23:54
n4nd0delicate... I remembered this23:54
n4nd0http://www.youtube.com/watch?v=qeGChOUqQuk23:54
n4nd0I couldn't stop listening this song for some years ago23:54
blackburnI know it is not really tolerant but I'd say it is not possible to pay them for not fcking all around the africa and spread aids23:56
blackburnit is a matter of culture or so23:57
blackburnn4nd0: I have heard that song back in time23:57
n4nd0blackburn: it is a hard truth23:57
n4nd0blackburn: it was in a episode of House23:57
n4nd0maybe you heard it there23:58
blackburnn4nd0: oh then I know23:58
blackburnthere were similar songs in23:58
blackburnoh let me find english name for these series23:58
blackburnahhhhh remembered23:58
blackburnscrubs23:58
n4nd0hehe, that's about doctors too23:58
--- Log closed Wed Mar 13 00:00:11 2013