Open in new window / Try shogun cloud
--- Log opened Mon May 28 00:00:41 2012
gsomixsonney2k, blackburn good night00:04
-!- blackburn [~blackburn@] has quit [Ping timeout: 244 seconds]02:06
-!- cheng [80fa500f@gateway/web/freenode/ip.] has joined #shogun02:48
-!- cheng [80fa500f@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]05:33
-!- gsomix [~gsomix@] has quit [Quit: Ex-Chat]06:38
-!- gsomix [~gsomix@] has joined #shogun06:39
@sonney2kgsomix, nice ok06:41
gsomixsonney2k, good morning :)06:42
-!- cheng [80fa500f@gateway/web/freenode/ip.] has joined #shogun06:48
-!- n4nd0 [] has joined #shogun08:02
-!- gsomix [~gsomix@] has quit [Ping timeout: 244 seconds]08:06
-!- n4nd0 [] has quit [Ping timeout: 240 seconds]08:24
-!- n4nd0 [] has joined #shogun08:38
-!- uricamic [~uricamic@2001:718:2:1634:d981:6e8d:30f8:8a27] has joined #shogun08:48
-!- cheng [80fa500f@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]09:00
-!- n4nd0 [] has quit [Ping timeout: 245 seconds]09:55
-!- n4nd0 [] has joined #shogun10:11
-!- n4nd0 [] has quit [Ping timeout: 245 seconds]10:20
-!- dfrx [] has joined #shogun10:41
-!- dfrx [] has quit [Quit: Leaving.]10:42
-!- heiko [] has joined #shogun11:10
-!- heiko1 [] has joined #shogun11:25
-!- heiko [] has quit [Ping timeout: 245 seconds]11:26
-!- blackburn [~blackburn@] has joined #shogun11:29
-!- blackburn [~blackburn@] has quit [Client Quit]11:30
-!- blackburn [~blackburn@] has joined #shogun11:32
-!- n4nd0 [] has joined #shogun11:42
n4nd0sonne|work, hey! are you ok if we just link to our blog for our weekly reports in the mailing list?11:55
-!- blackburn [~blackburn@] has quit [Ping timeout: 260 seconds]12:41
-!- heiko [] has joined #shogun12:45
-!- heiko1 [] has quit [Ping timeout: 260 seconds]12:48
-!- heiko [] has quit [Ping timeout: 245 seconds]13:29
-!- n4nd0 [] has quit [Quit: Ex-Chat]14:12
-!- heiko [] has joined #shogun14:57
-!- n4nd0 [] has joined #shogun16:35
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has joined #shogun17:25
n4nd0sonney2k, around?17:51
n4nd0sonne|work, ^?17:51
-!- uricamic [~uricamic@2001:718:2:1634:d981:6e8d:30f8:8a27] has quit [Quit: Leaving.]18:00
-!- n4nd0 [] has quit [Ping timeout: 244 seconds]18:31
-!- gsomix [~gsomix@] has joined #shogun18:38
-!- heiko [] has quit [Ping timeout: 245 seconds]18:44
-!- blackburn [~blackburn@] has joined #shogun18:48
gsomixblackburn, yo18:58
-!- heiko [] has joined #shogun19:03
@sonney2kgsomix, don't forget about the weekly report...19:04
gsomixsonney2k, thanks, I remember :)19:05
@sonney2kgsomix, heh19:05
@sonney2kgsomix, so what is the directorkernel state now?19:05
@sonney2kdo you have a patch I can try?19:05
gsomixsonney2k, compiling...19:06
@sonney2kn4nd0, just post an email to the ML with a pointer to your report for the week on your blog.19:10
* sonney2k checks whose report is missing19:14
@sonney2klooks like gsomix's n4nd0's and wikings...19:15
gsomixsonney2k, check github, please.19:19
-!- heiko [] has quit [Ping timeout: 245 seconds]19:27
CIA-113shogun: Evgeniy Andreev master * r1a1a30c / (2 files in 2 dirs): added missing virtual function to DirectorKernel -
CIA-113shogun: Evgeniy Andreev master * r92b1eec / src/shogun/kernel/Kernel.h : deleted spaces -
CIA-113shogun: Soeren Sonnenburg master * r2dc985c / (3 files in 2 dirs): Merge pull request #557 from gsomix/director_kernel -
gsomixsonney2k, works? =___=19:44
@sonney2kgsomix, why is get_kernel_row() not needed?19:45
@sonney2kI mean it is also defined virtula in ckernel?19:45
@sonney2kgsomix, compiling...19:45
gsomixsonney2k, oops.19:45
@sonney2kgsomix, ?19:47
@sonney2knot needed or needed?19:47
@sonney2kgsomix, why are virtual functions needed?19:47
gsomixsonney2k, needed. just a moment.19:47
gsomixsonney2k, oh. I really dont know, but I have some hypothesis.19:48
-!- n4nd0 [] has joined #shogun19:48
@sonney2kgsomix, then tell me about it :D19:49
gsomixit seems, that swig "director" (verb, hehe) virtual functions from CKernel.19:49
gsomixbut... there is table with DirectorKernel's virtual methods.19:50
gsomixand there are not virtual methods from CKernel in this table.19:51
@sonney2kgsomix, ok19:51
gsomixin my "home" sample this situation is throwing exception19:51
@sonney2kgsomix, btw the has_features check is a bit questionable19:52
@sonney2kif one uses e.g. only numpy19:52
@sonney2kgsomix, so why did the endless recursion happen w/ init?19:52
@sonney2kn4nd0, (reposting what I said above:  just post an email to the ML with a pointer to your report for the week on your blog.)19:53
n4nd0sonney2k, yeah, I just saw it in the logs, thank you :)19:53
gsomixsonney2k, get_num_vec_rhs/lhs in methods is reason19:53
gsomixsonney2k, e.g in get_kernel_matrix.19:54
@sonney2kgsomix, well I couldn't train w/o doing dkernel.init(lhs,rhs)19:54
@sonney2kI would also want to set num_lhs /rhs19:54
@sonney2kand do without even setting features19:55
gsomixsonney2k, mmm?19:55
gsomixjust override has_features19:55
@sonney2kgsomix, well if I use a numpy array as in the example I send to you19:55
@sonney2kgsomix, that is too slow19:55
@sonney2kit will be checked all the time19:55
@sonney2kgsomix, I think either adding a flag to constructor19:56
@sonney2kyeah well that would be best19:56
@sonney2kgsomix, add some bool to constructor no_features_required19:56
@sonney2kif false - do like we do now19:56
@sonney2kand otherwise return true on has_features19:57
@sonney2kgsomix, btw 7 seconds for director kernel here for 1000x1000 matrix19:57
@sonney2k0.7 for shogun's19:57
gsomixsonney2k, aha, slow...19:57
blackburn10? is ok19:58
@sonney2k(for 1 thread)19:58
@sonney2kfor 2 threads 0.4 s19:58
@sonney2k10x is ok19:58
@sonney2kI think so too19:58
@sonney2kconsidering how complex the thing is19:58
blackburndevelopment can take 10000000x19:58
@sonney2kgsomix, question would be if this director stuff also works from octave etc19:59
gsomixsonney2k, ok, I'll check it.19:59
blackburnso being able to prototype even slow kernel is soo awesome19:59
@sonney2kblackburn, we now have to decide which things we want to be director-enabled19:59
blackburnfeatures, kernels, distances19:59
blackburnmulticlass strategies?19:59
blackburnreject strategies19:59
blackburnsvm especially20:00
blackburnmay be some splitting strategy or so20:01
blackburnfor xval20:01
blackburnloss for SO20:01
@sonney2kblackburn, so the way to proceed would be to have an extra class for each thing right?20:01
blackburnyes I think so20:01
@sonney2kDirectorKernelMachine etc?20:01
@sonney2kblackburn, not so sure about sparse etc features...20:02
blackburnok so DirectorDotFeatures, DirectorKernel, DirectorDistance, DirectorMulticlassStrategy, DirectorRejectStrategy, DirectorKernelMachine, DirectorLinearMachine, DirectorStructuredLoss20:02
@sonney2kblackburn, uhh ohh20:03
@sonney2kfor dotfeatures we have no way to support the api20:03
blackburnhm hmhm20:04
@sonney2kand I see lots of problems with protected etc...20:04
@sonney2kn4nd0, ok so where is your email then?20:04
* sonney2k is missing wiking :(20:04
blackburnwe need DirectorDotFeatures20:04
@sonney2kdoes anyone here have contact to him?20:05
n4nd0sonney2k, working on it right now, making ac couple of updates to the blog20:05
@sonney2kblackburn, still not at the meeting, not in irc, no email no nothing. something is wrong :(20:06
* sonney2k tries to compile the swig interface file with optimizations enabled20:09
* sonney2k waits for minutes...20:09
blackburnswig with opts?20:11
@sonney2kblackburn, to see if it helps directorkernel20:11
@sonney2kbut no change20:11
@sonney2k0.4 w/o20:11
blackburnsonney2k: try to profile it somehow :)20:11
@sonney2kblackburn, might be that even doing this from python is as slow as this20:12
blackburnsonney2k: then measure sth like 'for i in xrange(N): for j in xrange(N): kernel(i,j)'20:13
n4nd0weekly report, there you go20:23
@sonney2kblackburn, boah 10.3s!20:29
blackburnsonney2k: fors?20:30
@sonney2kblackburn, yes20:32
blackburnsonney2k: and dot(X,X.T)?20:33
blackburnor X.T, X :D20:33
@sonney2kblackburn, 0.0 s20:36
blackburnehm so we have huge overhead out there :D20:37
@sonney2kblackburn, it is not so clear if directors affect the rest of shogun...20:37
blackburnsonney2k: what is d?20:38
@sonney2kand ndata=1000 too20:38
@sonney2kI think calling a blas operation is the fastest one can do20:38
@sonney2kand that is what numpy does :)20:39
blackburn0.97 here20:41
blackburnhowever dot takes 0.620:41
blackburnsonney2k: ^ how can that be?20:42
blackburn>>> start = time.time(); x =,data); end = time.time(); print end-start20:43
blackburn>>> start = time.time(); x = k.get_kernel_matrix(); end = time.time(); print end-start20:43
@sonney2kblackburn, what do you mean?20:48
@sonney2kit is just a 1000x1000 matrix20:48
@sonney2kso 0.3s difference don't matter20:48
blackburnsonney2k: why do you have 0.0?20:48
@sonney2kmy machine is faster :D20:49
blackburnbut why I have 0.3 difference but you have 0.6 difference20:49
@sonney2kblackburn, did you compile atlas for your machine?20:50
@sonney2kthen of course20:50
@sonney2kspeedup will be factor >1020:50
blackburnwhat do numpy use?20:50
@sonney2kthe one that is in your system20:50
@sonney2kso when you compiled20:50
@sonney2katlas using the machinism provided by debian20:51
@sonney2kthen it will be a lot faster20:51
blackburnsonney2k: is it necessary to *compile* it?20:51
blackburnsonney2k: IIRC you said one in debian is b0rken20:52
@sonney2kthere are no optimized atlas packages20:52
@sonney2kin debian/ubuntu20:52
@sonney2kblackburn, no20:52
blackburnis there any reason?20:52
@sonney2kblackburn, yes too many hardware versions20:53
@sonney2kand debians is i686 specific20:53
@sonney2kso sse2/3 whatever won't be enabled20:53
@sonney2kand that makes a *HUGE* difference20:53
* sonney2k is off for a swim20:53
blackburnheh let me check20:53
@sonney2kblackburn, just run the jblas benchmark btw20:54
@sonney2kLD_LIBRARY_PATH=/usr/lib/jni java -jar /usr/share/java/jblas.jar20:55
@sonney2kRunning benchmark "Java matrix multiplication, double precision".20:56
@sonney2kn = 10   :  0.854 GFLOPS (2134302 iterations in 5.0 seconds)20:56
@sonney2kn = 100  :  1.682 GFLOPS (4206 iterations in 5.0 seconds)20:56
@sonney2kn = 1000 :  1.114 GFLOPS (3 iterations in 5.4 seconds)20:56
@sonney2kRunning benchmark "Java matrix multiplication, single precision".20:56
@sonney2kn = 10   :  0.747 GFLOPS (1868112 iterations in 5.0 seconds)20:56
@sonney2kn = 100  :  1.668 GFLOPS (4170 iterations in 5.0 seconds)20:56
@sonney2kn = 1000 :  1.334 GFLOPS (4 iterations in 6.0 seconds)20:57
@sonney2kRunning benchmark "ATLAS matrix multiplication, double precision".20:57
@sonney2kn = 10   :  0.773 GFLOPS (1931439 iterations in 5.0 seconds)20:57
@sonney2kn = 100  :  5.010 GFLOPS (12525 iterations in 5.0 seconds)20:57
@sonney2kn = 1000 :  9.299 GFLOPS (24 iterations in 5.2 seconds)20:57
@sonney2kRunning benchmark "ATLAS matrix multiplication, single precision".20:57
@sonney2kn = 10   :  0.812 GFLOPS (2030799 iterations in 5.0 seconds)20:57
@sonney2kn = 100  :  9.997 GFLOPS (24993 iterations in 5.0 seconds)20:57
@sonney2kn = 1000 : 15.100 GFLOPS (38 iterations in 5.0 seconds)20:57
@sonney2kdo this for optimized / not optimized atlas20:57
@sonney2kyou will notice a big difference20:57
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]21:03
* gsomix have returned21:06
-!- blackburn [~blackburn@] has joined #shogun21:08
shogun-buildbotbuild #565 of python_modular is complete: Failure [failed test_1]  Build details are at  blamelist: gsomix@gmail.com21:32
-!- gsomix is now known as failman21:41
-!- failman is now known as fail_man21:41
fail_man>> This nickname is registered. | bebebe :(21:41
@sonney2kblackburn, so did you try?21:47
blackburnsonney2k: yeah but I think needs tweaking in configure21:48
blackburnbecame much slower :D21:48
CIA-113shogun: Soeren Sonnenburg master * r80e75c8 / examples/undocumented/python_modular/ : let example exit sanely when shogun wasn't compiled with directors -
@sonney2kblackburn, ?21:51
blackburnsonney2k: I just did configure with no options21:52
@sonney2kblackburn, I am talking about jblas...21:52
blackburnlet me try21:52
@sonney2kblackburn, yes directors have 100% overhead or so21:52
blackburnRunning benchmark "Java matrix multiplication, double precision".21:53
blackburnn = 10   :  0.411 GFLOPS (1027007 iterations in 5.0 seconds)21:53
blackburnn = 100  :  0.630 GFLOPS (1574 iterations in 5.0 seconds)21:53
blackburnn = 1000 :  0.439 GFLOPS (2 iterations in 9.1 seconds)21:53
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]21:55
@sonney2kblackburn, that is just java21:57
@sonney2kso where is the atlas part?21:58
blackburnRunning benchmark "ATLAS matrix multiplication, double precision".21:59
blackburnn = 10   :  0.368 GFLOPS (918914 iterations in 5.0 seconds)21:59
blackburnn = 100  :  0.995 GFLOPS (2488 iterations in 5.0 seconds)21:59
blackburnn = 1000 :  1.131 GFLOPS (3 iterations in 5.3 seconds)21:59
@sonney2kn = 1000 :  9.299 GFLOPS (24 iterations in 5.2 seconds)22:06
@sonney2kblackburn, what machine do you have?22:06
blackburnvendor_id: GenuineIntel22:07
blackburncpu family: 622:07
blackburnmodel: 2322:07
blackburnmodel name: Pentium(R) Dual-Core CPU       T4300  @ 2.10GHz22:07
@sonney2kblackburn, Intel(R) Core(TM)2 Duo CPU     T9900  @ 3.06GHz22:07
blackburnwell not 9 times faster22:07
@sonney2kcertainly not22:07
blackburnsonney2k: I just ran configure with no options at all22:08
blackburnI do think it can cause low performance22:08
fail_mansonney2k, ufff, I sent a little report.22:09
@sonney2kfail_man, heh22:16
@sonney2kfail_man, so lets quickly plan the next steps:22:16
@sonney2kyou should introduce these Director* classes for lots more similar classes22:17
@sonney2klike blackburn said22:17
@sonney2kkernel was a good start, maybe some linear / kernel machine should be next22:17
@sonney2kfail_man, as you know I will be away the week after next wek22:25
@sonney2kso you should one by one do more director* classes22:25
@sonney2kshould be rather easy now that we know why things didn't wotk22:26
@sonney2kfail_man, btw the amount of information in your update email is just what we want!22:32
@sonney2kso good job22:32
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has joined #shogun22:36
fail_mansonney2k, thanks.22:40
fail_mangood night guys23:02
fail_mansonney2k, I will push DirectorKernel with no_features tomorrow.23:02
n4nd0fail_man, good night!23:24
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]23:33
--- Log closed Tue May 29 00:00:41 2012