IRC logs of #shogun for Saturday, 2011-10-08

« Friday, 2011-10-07 Index Sunday, 2011-10-09 »
--- Log opened Sat Oct 08 00:00:19 2011
CIA-3shogun: Sergey Lisitsyn  * r5a4f2de / (10 files): Finally added missing libshogun tests for DR preprocessors (no leaks) - http://git.io/tOuhTw00:15
-!- blackburn [~blackburn@85.114.187.69] has quit [Quit: Leaving.]00:45
-!- sonne|work [~sonnenbu@194.78.35.195] has quit [Ping timeout: 240 seconds]08:29
-!- sonne|work [~sonnenbu@194.78.35.195] has joined #shogun09:55
-!- blackburn [~blackburn@85.114.187.69] has joined #shogun11:18
CIA-3shogun: Sergey Lisitsyn  * r2382e0c / src/shogun/classifier/GaussianNaiveBayes.h : Added necessary Math.h include to GNB - http://git.io/jGHYyA12:35
CIA-3shogun: Sergey Lisitsyn  * re927318 / (src/shogun/machine/Machine.cpp src/shogun/machine/Machine.h): Moved code from .h to .cpp at CMachine - http://git.io/nzUQFw12:35
blackburnsonney2k: just tried to integrate covertree and it is fucking broken itself14:34
blackburnwrong15:43
-!- alesis-novik [~alesis@cpat001.wlan.net.ed.ac.uk] has joined #shogun16:09
@sonney2kblackburn, ?19:02
@sonney2kJL's code you mean?19:02
blackburnsonney2k: I was wrong, it is correct19:03
@sonney2kok19:03
blackburn(but it is ugliest thing I've ever seen19:03
@sonney2kblackburn, do you recall the C5.0 code?19:04
@sonney2kI would hope it is better than that19:04
blackburnwell19:05
blackburnI don't know19:05
blackburnthey are awful19:05
@sonney2kblackburn, I am still restoring my harddrive - I have basically random files that are corrupted19:06
@sonney2kthats like the worst that could happen19:06
blackburndamn19:06
@sonney2kI never know which file is $%*(Y$19:06
@sonney2kwrong here means anything from wrong permissions to garbage content19:07
@sonney2kblackburn, ahh btw I had some advanced java programming lessons last week19:07
@sonney2kso I had to use eclipse19:07
blackburnsonney2k: ahaha and how it was?19:08
@sonney2kit is pretty nice (if it doesn't crash)19:08
blackburnyou should try IDEA :D19:08
@sonney2kwell java is well weird :)19:08
@sonney2ksome things that trigger crashes in c/c++ are just hidden problems there19:08
@sonney2kpoor students that start with java as their first language19:09
@sonney2kthey will have a hard time finding out what is really going on underneath19:09
blackburnsure19:09
blackburnthere are a lot of pitfalls19:09
@sonney2kbut the course was nice - java thread programming19:09
@sonney2ksome things are *very* elegant19:10
@sonney2ksome things are very ugly19:10
blackburnlike?19:10
@sonney2k(like the synchronized stuff)19:10
blackburnah19:10
blackburnyes19:10
@sonney2know you have locks and sync'd sections19:10
@sonney2kand since Object has wait / signal or so19:10
@sonney2kyou easily mix up everything like await / wait19:11
@sonney2ksignal / notify19:11
blackburnI feel depressed with some things available in java but not in our shogun :)19:11
@sonney2kanyway going to reboot now19:11
@sonney2kblackburn, well we can improve as much as is possible withing C++ :)19:12
@sonney2kbrb (hopefully)19:12
blackburnoka19:13
blackburny19:13
blackburnsonney2k: why you changed v_array with shashwat?20:00
-!- alesis-novik [~alesis@cpat001.wlan.net.ed.ac.uk] has quit [Ping timeout: 258 seconds]20:27
-!- alesis-novik [~alesis@cpat001.wlan.net.ed.ac.uk] has joined #shogun20:30
@sonney2kblackburn, ?20:31
blackburnsonney2k: I guess it was struct?20:31
@sonney2kI am back - what about v_array?20:31
@sonney2kwehre is it?20:31
blackburncovertree uses it too20:31
blackburnshogun/lib/v_array.h20:31
blackburnbut covertree uses it as structure20:31
blackburnI guess I have to make it unified but I have no time to rewrite covertree :(20:33
@sonney2kblackburn, why not use the one in shogun also in covertree?20:34
@sonney2kI guess it doesn't occur at too many locations20:34
blackburnI hope20:34
blackburnsonney2k: they even didn't consider deletion of resulting tree!20:37
blackburnsome people should not program :D20:37
@sonney2kblackburn, well it is research code right?20:37
@sonney2kso it is just some throwaway stuff20:37
blackburnsonney2k: I prefer good code whenever it research or not :)20:38
@sonney2kresearch is very fast paced20:38
@sonney2kso the time you waste on writing clean throwaway code is sth you should better spend on writing papers20:39
@sonney2kbut yes I agree20:39
blackburnwriting better code will help later anyway20:40
@sonney2kHowever, my lessons learned is that you will always re-use your code at some point later20:40
@sonney2kyes20:40
@sonney2kso writing clean code and a small test is *never* wasted20:41
blackburnthat's my point too20:41
blackburnsonney2k:20:51
blackburnfloat *new_p;20:51
blackburn      posix_memalign((void **)&new_p, 16, p.index*sizeof(float));20:51
blackburnthat's crazy20:51
blackburnthey would better work on different things20:51
blackburnthan some tricky memalign and so on20:51
@sonney2kwell aligned data can be faster...20:52
@sonney2kI mean fit completely in cache if done right20:52
@sonney2knon-trivial though20:52
blackburnthey did it fast in complexity means20:52
blackburnno need to do these things20:53
blackburnwhy not to write in assembler then?20:53
@sonney2kblackburn, some consider C to be a big macro assembler :D20:55
blackburnsonney2k: http://www.agner.org/optimize/20:56
blackburnpage to bookmark :)20:56
@sonney2knice link20:58
@sonney2kyes done already :)20:58
blackburnsonney2k: did I show some scikits vs shogun LLE benchmark?21:00
@sonney2kprobably - but I forgot21:00
@sonney2kagain please :)21:00
blackburnthis one21:01
blackburnhttp://dl.dropbox.com/u/10139213/shogun/bench.png21:01
blackburnthe only problem - it become vice versa if k is big21:01
blackburncause they have balltree21:01
-!- alesis-novik [~alesis@cpat001.wlan.net.ed.ac.uk] has quit [Read error: Operation timed out]21:01
blackburnso that's why I'm working on covertree integration21:02
blackburnsonney2k: if you will have some time to install scikits I could send benchmark to you to check how it is on your machine21:04
-!- alesis-novik [~alesis@cpat001.wlan.net.ed.ac.uk] has joined #shogun21:05
blackburnalesis-novik: how are you buddy? :)21:05
@sonney2kblackburn, what does 1T / 2T stand for?21:05
blackburnsonney2k: threads21:05
@sonney2kk21:06
alesis-novikblackburn, doing well. Started my PhD, organising a conference, tutoring some courses, using Shogun in my research :)21:06
alesis-novikWhat about you?21:06
blackburnalesis-novik: started working, wanna die integrating covertree :D21:06
blackburnhave you tried my fancy algos? ;)21:06
alesis-novikLTSA21:07
blackburnbtw I came up with an idea of kernel LTSA21:07
blackburnit is in shogun already21:07
alesis-novikCool. I'm mostly doing finance now :D21:08
@sonney2kalesis-novik, so when can we expect your next patch :)21:10
alesis-noviksonney2k, whenever I get some time to be useful :)21:10
alesis-novikHopefully soon21:10
alesis-novikHopefully I can get up to date and potentially mentor next year21:11
blackburnwe are really need to have not only two contributors :)21:11
blackburnsonney2k: btw I have updated shogun wiki page a little21:12
blackburnhttp://en.wikipedia.org/wiki/Shogun_(toolbox)21:12
blackburnsonney2k: and I need your help with multiclass svm :)21:13
@sonney2kwhats wrong with it?21:13
blackburnI need you to recommend papers to me21:13
blackburn:D21:13
@sonney2kblackburn, for true multiclass or one-vs-one / one-vs-rest?21:25
blackburnsonney2k: true multiclass21:29
@sonney2kthen the cramer singer paper21:33
@sonney2kand vojtech's21:33
-!- alesis-novik [~alesis@cpat001.wlan.net.ed.ac.uk] has quit [Read error: Operation timed out]21:41
-!- alesis-novik [~alesis@cpat001.wlan.net.ed.ac.uk] has joined #shogun21:43
blackburnftp://cmp.felk.cvut.cz/pub/cmp/articles/franc/franc-multiSVM02.pdf sonney2k: this one?21:44
@sonney2kyes21:44
blackburnsonney2k: thanks21:46
blackburnsonney2k: changing float to double leads covertree to die :D21:52
blackburnthat's %$%#$ !21:53
@sonney2kwell covertree is for performance so just stay with floats :)21:53
@sonney2kblackburn, do you also have these warnings? :21:55
@sonney2k/home/sonne/Documents/priv/repositories/shogun/src/shogun/mathematics/arpack_s.cpp|132| warning: ‘slu_Bv’ may be used uninitialized in this function21:55
@sonney2k/home/sonne/Documents/priv/repositories/shogun/src/shogun/mathematics/arpack_s.cpp|133| warning: ‘slu_Xv’ may be used uninitialized in this function21:55
@sonney2k/home/sonne/Documents/priv/repositories/shogun/src/shogun/mathematics/arpack_s.cpp|117| warning: ‘perm_c’ may be used uninitialized in this function21:55
@sonney2k/home/sonne/Documents/priv/repositories/shogun/src/shogun/mathematics/arpack_s.cpp|117| warning: ‘perm_r’ may be used uninitialized in this function21:55
@sonney2k/home/sonne/Documents/priv/repositories/shogun/src/shogun/mathematics/arpack_s.cpp|117| warning: ‘etree’ may be used uninitialized in this function21:55
@sonney2k/home/sonne/Documents/priv/repositories/shogun/src/shogun/mathematics/arpack_s.cpp|118| warning: ‘R’ may be used uninitialized in this function21:55
@sonney2k/home/sonne/Documents/priv/repositories/shogun/src/shogun/mathematics/arpack_s.cpp|118| warning: ‘C’ may be used uninitialized in this function21:55
blackburnehh21:55
blackburnno21:55
CIA-3shogun: Soeren Sonnenburg  * r9b410c9 / src/NEWS : update NEWS file for 1.1 - http://git.io/dvVE3A21:56
CIA-3shogun: Soeren Sonnenburg  * r884c74a / (4 files): move regression .h functions to .cpp - http://git.io/QEnhfQ21:56
blackburnsonney2k: you have no superlu I guess?21:56
blackburnno, not possible21:56
blackburnokay I'll null it21:57
@sonney2kChecking for Atlas support ... yes21:57
@sonney2kChecking for SuperLU support ... yes21:57
@sonney2kChecking for ARPACK support ... yes21:57
blackburnI don't know why I don't get it21:58
CIA-3shogun: Sergey Lisitsyn  * re13b169 / (2 files): Initialized helper arrays at arpack wrapper to avoid warnings - http://git.io/TNyJeQ22:00
CIA-3shogun: Sergey Lisitsyn  * rfa8405f / src/NEWS : Mention SUPERLU and KLTSA, KLLE - http://git.io/iu3ahQ22:02
blackburnsonney2k: how faster is float in comparison to double?22:05
@sonney2k2 times22:05
@sonney2kat least22:05
blackburnsonney2k: so may be I should transfer dimension reduction to float32_t?22:07
blackburnI don't really imagine if somebody will want double precision for visualization22:08
@sonney2knot for visualization but sometimes for data preprocessing22:12
blackburnsonney2k: making it templated is too much pain in the ass for me22:13
@sonney2kfor online learning we agreed on using float because that is really where it is all about speed22:13
@sonney2kagreed22:13
blackburnsonney2k: are you against -> float32_t?22:14
blackburn2x possible factor is really something important for this algos22:15
@sonney2kblackburn, well it is for every algorithm... not sure if it is really worth it22:17
blackburnsonney2k: what is the standard precision for numpy/scipy?22:19
@sonney2kdouble22:19
blackburnokay so let it stay double too22:19
@sonney2kI would say one has to have a strong reason to not use double22:20
@sonney2kfloat == 1/2 memory usage22:20
blackburnI can 1/2 memory usage by triangular matrices22:21
blackburnbut covertree first22:21
blackburnfucking covertree22:21
blackburnsonney2k: just got email from guy doing dimreduction at scikit-learn :)22:41
@sonney2kand what does he say?22:45
blackburnhe is curious how could benchmark look like22:46
blackburnI don't know how to integrate this god damn covertree22:47
blackburnbad code wins22:50
-!- alesis-novik [~alesis@cpat001.wlan.net.ed.ac.uk] has quit [Ping timeout: 255 seconds]22:51
@sonney2kblackburn, you mean compare scikits learn to shogun dimred stuff?23:01
@sonney2kspeed wise?23:01
blackburnyes23:01
blackburnyes23:01
@sonney2kok23:01
@sonney2kI guess you could even add this code as an example or appilcation..23:02
-!- alesis-novik [~alesis@cnat206.wlan.net.ed.ac.uk] has joined #shogun23:02
blackburnsonney2k: which code?23:07
@sonney2kthe scikits <-> shogun dimred comparison23:13
blackburnah23:14
@sonney2kg'night23:25
blackburnsee ya23:25
--- Log closed Sun Oct 09 00:00:24 2011
« Friday, 2011-10-07 Index Sunday, 2011-10-09 »

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