IRC logs of #shogun for Monday, 2011-06-13

« Sunday, 2011-06-12 Index Tuesday, 2011-06-14 »
--- Log opened Mon Jun 13 00:00:21 2011
-!- in3xes_ [~in3xes@180.149.49.227] has joined #shogun00:00
-!- in3xes [~in3xes@180.149.49.227] has quit [Ping timeout: 246 seconds]00:04
@sonney2ki can imagine...00:13
-!- in3xes_ is now known as in3xes00:15
serialhexthe insanity that some people code... though this is kinda interesting: http://github-high-scores.heroku.com/shogun-toolbox/shogun/high_scores/04:59
-!- in3xes_ [~in3xes@180.149.49.227] has joined #shogun06:02
-!- in3xes [~in3xes@180.149.49.227] has quit [Ping timeout: 276 seconds]06:06
-!- in3xes_ is now known as in3xes06:17
-!- in3xes [~in3xes@180.149.49.227] has quit [Ping timeout: 260 seconds]07:15
-!- in3xes [~in3xes@180.149.49.227] has joined #shogun07:27
-!- alesis-novik [~alesis@78-56-236-240.static.zebra.lt] has quit [Ping timeout: 244 seconds]09:41
-!- alesis-novik [~alesis@78-56-236-240.static.zebra.lt] has joined #shogun09:58
-!- alesis-novik [~alesis@78-56-236-240.static.zebra.lt] has quit [Ping timeout: 240 seconds]10:04
-!- alesis-novik [~alesis@78-56-236-240.static.zebra.lt] has joined #shogun10:17
-!- alesis-novik [~alesis@78-56-236-240.static.zebra.lt] has quit [Ping timeout: 276 seconds]10:29
-!- alesis-novik [~alesis@78-56-236-240.static.zebra.lt] has joined #shogun10:47
-!- alesis-novik [~alesis@78-56-236-240.static.zebra.lt] has quit [Ping timeout: 240 seconds]10:59
-!- alesis-novik [~alesis@78-56-236-240.static.zebra.lt] has joined #shogun11:06
CIA-18shogun: Soeren Sonnenburg master * r78b5538 / src/.generate_link_dependencies.py :11:18
CIA-18shogun: test if doxygen is installed and only then add the *_doxygen.i files as11:18
CIA-18shogun: deps - http://bit.ly/lmg2Fb11:18
-!- alesis-novik [~alesis@78-56-236-240.static.zebra.lt] has quit [Ping timeout: 246 seconds]11:28
CIA-18shogun: Soeren Sonnenburg master * r9ec240c / examples/undocumented/java_modular/ClassifierLibsvmMinimalModular.java : modified example to work with jblas DoubleMatrix - http://bit.ly/m38Zt511:37
-!- alesis-novik [~alesis@78-56-236-240.static.zebra.lt] has joined #shogun11:40
-!- in3xes_ [~in3xes@180.149.49.227] has joined #shogun15:07
-!- in3xes [~in3xes@180.149.49.227] has quit [Ping timeout: 240 seconds]15:11
-!- alesis-novik [~alesis@78-56-236-240.static.zebra.lt] has quit [Quit: Leaving]16:29
-!- blackburn [~blackburn@188.122.224.102] has joined #shogun17:16
blackburnhttp://dl.dropbox.com/u/10139213/share/image.png17:26
blackburnmds faces example ^ :)17:27
-!- in3xes_ is now known as in3xes17:47
@sonney2kblackburn, heh :)19:53
@bettyboo<:*)19:53
blackburnsonney2k: that picture rocks http://dl.dropbox.com/u/10139213/share/image2.png19:53
blackburnunfortunately I can't integrate it into shogun19:54
blackburnbut I have an idea19:54
blackburngoing to make my own dataset I can integrate19:54
blackburnsonney2k: is it ok for you?19:54
@sonney2kno not ok19:55
@sonney2kfantastic!19:55
blackburnsonney2k: I think I'll make photos of my girlfriend :D19:55
blackburnjust as these ones19:55
@sonney2kblackburn, she will probably be very pleased19:56
blackburnsonney2k: I hope so :)19:56
blackburnlet's try isomap on these pics19:57
@sonney2kblackburn, what similarity measure are you using?19:58
blackburnjust euclidian distance between matrices :)19:59
@sonney2kimpressive results...19:59
blackburnin a minute we will see what isomap makes for this dataset20:00
blackburnand after I'll try LLE20:00
blackburnoh isomap made it pretty similar20:02
blackburnlle turn20:02
-!- f-x [~user@117.192.221.133] has joined #shogun20:04
blackburnsonney2k: which version of matplotlib do we require for shogun?20:05
@sonney2kblackburn, at least 0.99.3 I would say20:05
blackburnIdk exactly but now I'm using some thing (OffsetImage) that isn't in 0.9920:05
blackburnat least it wasn't in default ubuntu repo matplotlib20:06
blackburnI had to install 1.0.1 manually20:06
blackburnthe reason why - it seems to be the only way to place images on graph plot20:06
blackburnwow, LLE had segfault20:07
blackburnbad bad20:07
@sonney2kblackburn, well you could do a test in your example for matplotlib 1.0.x and then complain if it is not there20:09
blackburnah okay20:10
@sonney2kdoesn't matter too much for this example I would say20:10
blackburnsad news - LocallyLinearEmbedding is bad now20:11
blackburngoing to debug20:11
@sonney2kf-x, any news regarding the benchmarking?20:13
f-xsonney2k: using fread(), should i try to read the whole file in one go?20:14
@sonney2kwell with sth. like filesize == 3times your memory20:15
@sonney2kit will be hard20:15
@sonney2kjust read huge chunks20:15
@sonney2kuntil eof20:15
blackburnhaha found error20:16
f-xi think that can be done with a small modification in the io_buf code from vw20:16
@bettyboogrin20:16
f-xjust changing the default buffer size from 64 KB to something like 10 MB20:16
blackburnsonney2k: I forgot that get_feature_vector *creates* new vector20:16
blackburnsonney2k: so with 4000d examples it wastes memory a little ;D20:17
@sonney2kblackburn, wait get_feature_vector creates new vector?20:23
blackburnyes20:23
@sonney2kwell it doesn't necessarily20:23
blackburnsonney2k: I just used feature_matrix20:24
blackburnwith SGMatrix™20:24
blackburnso problem disapperead20:24
blackburntfu20:24
@sonney2kblackburn, yeah but then it doesn't create a new vector / matrix etc...20:24
blackburnwith SGMatrix no20:24
blackburnbut with get_feature_vector - yes20:24
@sonney2kahh20:24
@sonney2kI understand20:24
blackburnsonney2k:20:25
blackburn*dst=(ST*) SG_MALLOC(vlen*sizeof(ST));20:25
blackburnmemcpy(*dst, vec, vlen*sizeof(ST));20:25
blackburnthat's how it does20:25
@sonney2kthat code will die soon20:25
@sonney2kat least for python I've converted all the typemaps20:25
blackburnSimpleFeatures.h:26120:25
blackburnI really like new SGMatrix20:25
blackburnI'm now using it where possible20:26
@sonney2know I need to do it for octave / R and then I can start replacing these copy crap things20:26
@sonney2kdouble** stuff20:26
blackburnsonney2k: the only thing - I don't use it if I need some matrices to multiply, subtract and so on20:26
blackburnwhat do you think about it? should sgmatrix be used everywhere where matrices are used?20:27
blackburnsegfault again! damn20:28
@sonney2kblackburn, it makes sense to go that route...20:29
blackburnsonney2k: with sth like SGMatrix.dot(A,B) and so on?20:29
@sonney2kmultiply20:30
@sonney2kyes20:30
blackburnI see20:30
@sonney2kA.multiply(B)20:30
@sonney2kbut only the things that you need - we don't want to compete with the many matrix / vector libraries here20:31
@sonney2ksigh20:31
@sonney2ksploving is not here and I haven't seen him for a long time20:31
blackburnsonney2k: I just want ARPACK :)20:31
@sonney2kas you wish...20:32
blackburnnothing more :)20:32
blackburnbecause Lanzcos is very crucial20:32
@sonney2kjust don't forget you have another exam20:32
blackburnah yeap20:32
blackburnbut now I want to find out why LLE faults20:33
blackburnoh20:36
blackburnI guess I was drunk writing it20:36
f-xsonney2k: which functions are you looking to have reimplemented in StreamingAsciiFile? the get_*_vector() functions?20:36
* sonney2k compiles kernel 3.0rc220:37
blackburnaren't you like stable? :)20:37
@sonney2kblackburn, I though Russians are non-operational without their Vodka ;-)20:39
blackburnyeah something like that :)20:40
blackburnthere are so many matrices that I use gnote for checking if I not delete some20:41
@sonney2kblackburn - but Linus' kernels are always stable20:43
@sonney2k:)20:44
blackburnexactly but I mean I thought you use old&stable soft :)20:44
@sonney2kblackburn, the desktop environment is much more complex than the kernel20:46
@sonney2kf-x, yeah that are the ones we currently need20:47
f-xsonney2k: okay, so i guess i don't need to change much in the code which is currently StreamingFile.cpp20:47
@sonney2kf-x, well we need efficient function though20:48
@sonney2kf-x so your buffered line reader is highly needed20:48
f-xsonney2k: that will be used, though i cannot think of something more efficient than the readto() from VW for now20:50
blackburnhehe N*m_k -> N*dim20:52
blackburnoh I finally fixed it20:58
blackburnsonney2k: please merge pull request with fixes :)21:03
blackburnit is bug-free now hehe21:03
-!- f-x [~user@117.192.221.133] has quit [Remote host closed the connection]22:12
@sonney2kRe22:26
@sonney2kblackburn, I think it would be more readable to define some float64_t* matrix=feature_matrix.matrix22:34
@sonney2kand then use this22:34
blackburnwhy?22:34
blackburnit is used only twice22:34
blackburnsonney2k: well then if we will have 3-4 matrices it doesn't make sense do it this way22:37
blackburnso I think it is more readable22:38
@sonney2kI understand you have many matrices :)22:41
@sonney2kI am just not sure if feature_matrix.matrix[neighborhood_matrix[j*N+i]*dim+k]; is optimized22:42
blackburnwhy not?22:43
blackburnwe should check because in the case it is not optimized SGMatrices shouldn't be used internally22:44
blackburnsonney2k: how can I check it?22:46
@sonney2kwell you could just measure the speed22:46
@sonney2kwell SGMatrices can of course be used22:47
@sonney2kbut the dereferencing feature_matrix.matrix could be costly22:47
@sonney2kI am just not sure22:47
blackburnokay will check it in a minute22:47
@sonney2khttp://www.fefe.de/know-your-compiler.pdf does only mention an aliasing problem22:49
@sonney2kthanks22:50
blackburnsonney2k: how can I get assembly code?22:51
@sonney2kblackburn, I would just measure speed22:51
blackburnyes I would rather too but just curious22:51
blackburnyou have experience I guess :)22:52
@sonney2kblackburn, http://www.delorie.com/djgpp/v2faq/faq8_20.html22:52
-!- blackburn [~blackburn@188.122.224.102] has quit [Quit: Leaving.]22:52
-!- blackburn [~blackburn@188.122.224.102] has joined #shogun22:53
* sonney2k reboots into the new kernel...22:56
@sonney2kbrb *hopefully*22:56
blackburnsonney2k: making shogun + testing lle + music playing + etc = fire22:56
blackburnhehe22:56
blackburngood luck22:56
blackburnfire on your applet :)22:56
blackburnsonney2k: clock() is unreliable a bit, can you suggest me a better way to measure time? googling didn't help23:15
blackburnwell with float64_t* - average is 60000 clocks23:19
blackburnsame without23:20
blackburnhave to check assembly23:25
@sonney2kblackburn, epic fail23:28
blackburnwhat's up?23:28
@sonney2kI am now back to a version before the version I used to run before23:28
@sonney2knvidia drivers don't understand 3.0.x23:29
blackburnah23:29
CIA-18shogun: Sergey Lisitsyn master * r9552aef / src/libshogun/preprocessor/LocallyLinearEmbedding.cpp : Crucial fixes for drunk-made mistakes in LocallyLinearEmbedding - http://bit.ly/koZjxa23:32
blackburnsonney2k: I haven't ever thought that gcc makes so sophisticated assembly code23:32
blackburnsonney2k: about time measure - any ideas?23:32
@sonney2kblackburn, use shogun's CTime23:33
@sonney2kget_curtime()23:33
@sonney2kresoluition is microseconds23:33
blackburnit uses clock()23:34
blackburnjust like I did23:34
@sonney2knope23:34
@sonney2kit uses gettimeofday23:34
blackburnah23:34
blackburnI see23:34
blackburnwill try23:34
blackburnsonney2k: about ~0.055 without additional pointer23:37
blackburnabout 0.055 with23:37
blackburnsonney2k: so seems no diffrerence23:38
@sonney2kvery good23:38
blackburnhehe it produced all sources into .s and .ii files23:41
blackburnwith assembly code23:41
@sonney2k:)23:41
@sonney2kfor some reason 1 of my cpu cores vanished?!23:42
blackburnhow? :)23:49
@sonney2kno idea all slow now23:50
blackburnwhich kernel do you use now?23:53
@sonney2ksome .3823:59
--- Log closed Tue Jun 14 00:00:23 2011
« Sunday, 2011-06-12 Index Tuesday, 2011-06-14 »

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