Open in new window / Try shogun cloud
--- Log opened Thu Apr 05 00:00:19 2012
gsomixsonney2k, there is bug in strings. i already fixed it, but not posted code.00:00
gsomixsonney2k, 800MB. gcc 4.5.200:00
gsomixstrange00:00
@sonney2k?00:00
@sonney2khow that?00:00
gsomixsonney2k, are you asking about python3 or memory?00:02
gsomix*or about memory00:02
gsomixsonney2k, can you see my proposal?00:09
gsomixat that time I can make corrections in python3 interface.00:10
-!- gsomix [~gsomix@83.234.169.131] has quit [Quit: ????? ? ?? ??? (xchat 2.4.5 ??? ??????)]00:13
-!- gsomix [~gsomix@83.234.169.131] has joined #shogun00:14
-!- emrecelikten [~Anubis@92.44.132.147] has quit [Quit: Leaving.]00:15
@sonney2kblackburn_, 2700:17
@sonney2knite00:17
blackburn_sonney2k: nite00:17
blackburn_newton svm seems to take a while on my data00:17
n4nd0good night00:17
blackburn_still training00:17
gsomixsonney2k, oh, no. all the changes I'll post tomorrow.00:21
gsomixi fall asleep =___=00:21
gsomixgood night, guys00:23
blackburn_nite00:23
n4nd0good night00:24
-!- blackburn_ [5bde8018@gateway/web/freenode/ip.91.222.128.24] has quit [Quit: Page closed]00:51
n4nd0good night people00:51
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Ping timeout: 260 seconds]00:57
-!- av3ngr [av3ngr@nat/redhat/x-epzrsrwbdomgshvs] has joined #shogun01:32
-!- maximus [86544bc9@gateway/web/freenode/ip.134.84.75.201] has joined #shogun02:56
-!- maximus is now known as Guest7949402:56
-!- Guest79494 [86544bc9@gateway/web/freenode/ip.134.84.75.201] has quit [Client Quit]02:57
-!- flxb [~cronor@g231230100.adsl.alicedsl.de] has quit [Quit: flxb]02:57
-!- mohit [86544bc9@gateway/web/freenode/ip.134.84.75.201] has joined #shogun03:05
mohitHi03:05
mohitis there anybody from shogun here?03:05
mohitI have a query regarding gsoc?03:07
-!- mohit [86544bc9@gateway/web/freenode/ip.134.84.75.201] has quit [Quit: Page closed]03:14
-!- pluskid [~chatzilla@li400-235.members.linode.com] has joined #shogun03:18
-!- pluskid_ [chatzilla@111.120.23.144] has joined #shogun04:46
-!- pluskid [~chatzilla@li400-235.members.linode.com] has quit [Ping timeout: 276 seconds]04:50
-!- pluskid_ is now known as pluskid04:50
-!- pluskid [chatzilla@111.120.23.144] has quit [Quit: ChatZilla 0.9.88.1 [Firefox 11.0/20120312181643]]05:39
-!- Johannes_ [55e5c1c3@gateway/web/freenode/ip.85.229.193.195] has joined #shogun05:49
Johannes_Morning, do I have to install examples seperatly? Been trying all morning to run one :)05:59
Johannes_Or is it perhaps as simple as that it's only included in the source code version, not from the package.06:01
Johannes_yup, was as simple as that.06:07
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun07:05
-!- Afan [~Afan@c-76-29-50-209.hsd1.il.comcast.net] has joined #shogun07:22
@sonney2kJohannes_, use this to get them from git git submodule update --init08:08
Johannes_Thank you, downloaded the source code instead, might still try that though.08:09
Johannes_still can't run the java-modular examples though, probably need to install something additional because I couldn't locate modshogun.jar08:11
@sonney2kJohannes_, did it compile the modshogun.jar08:16
@sonney2k?08:16
@sonney2kif so it should be somewhere under shogun/java_modular08:16
@sonney2kand you need to have CLASSPATH set ...08:16
@sonney2kmorning n4nd008:17
Johannes_not sure if it compiled the modshogun.jar, used the package, wasn't so specific at install.08:23
-!- flxb [~cronor@g231228114.adsl.alicedsl.de] has joined #shogun08:23
@sonney2kuse find ./ -name '*.jar' to figure out if it did08:25
@sonney2kgtg08:25
@sonney2kl8r08:25
Johannes_yea tried that already, doesen't look like it installed it08:29
Johannes_thanks for the help.08:29
gsomixsonney2k, hi08:31
n4nd0sonney2k: good morning08:42
-!- Afan [~Afan@c-76-29-50-209.hsd1.il.comcast.net] has quit [Ping timeout: 246 seconds]08:46
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds]09:05
-!- av3ngr [av3ngr@nat/redhat/x-epzrsrwbdomgshvs] has quit [Quit: That's all folks!]09:08
CIA-64shogun: pluskid master * r580d1f8 / (3 files): Use the built-in get_unique_labels instead of wrting a new function. - http://git.io/6whL1A09:23
CIA-64shogun: pluskid master * r49b6e2e / (5 files): Refine the code according to coding convention. - http://git.io/A1opPw09:23
CIA-64shogun: Soeren Sonnenburg master * r874ebc5 / (9 files in 3 dirs):09:23
CIA-64shogun: Merge pull request #420 from pluskid/normalized-mutual-info09:23
CIA-64shogun: Normalized mutual information - http://git.io/XYUlvA09:23
CIA-64shogun: pluskid master * r57d869d / (6 files in 3 dirs): Added normalized mutual information evaluation for clustering. - http://git.io/nLyPtg09:23
CIA-64shogun: pluskid master * r9473a5d / examples/undocumented/python_modular/evaluation_clustering.py : Remove debug code. - http://git.io/dqoG3w09:23
-!- sonne|work [~sonnenbu@194.78.35.195] has joined #shogun09:24
sonne|workgsomix: one further idea - you could replace/remove the current CSet implementation with your hash set one09:32
sonne|workgsomix: and even one more: in shogun we have SGArray/SGArray2 etc - these all can be removed and replaced by SGVector/SGMatrix/SGNDArray09:33
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]09:35
-!- gsomix [~gsomix@83.234.169.131] has quit [Ping timeout: 260 seconds]09:36
n4nd0sonne|work: one doubt about the design related to covertree in KNN09:43
n4nd0sonne|work: it would be really nice if CDistance would support queries for distances between vectors in the same features object09:44
n4nd0sonne|work: now we can do distance(a, b) that gives the distance between vector a in lhs and vector b in rhs09:44
sonne|workwhat is this needed for?09:45
n4nd0sonne|work: what I mean is that it would be useful if we can do distance_lhs(a, b)09:45
n4nd0so in KNN09:45
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun09:45
n4nd0we initilize the distance with train vectors09:45
sonne|workyou need it for covertrees + KNN?09:45
n4nd0so we have distance(train, train)09:46
n4nd0yes09:46
n4nd0but if we later apply we will have distance(train, test)09:46
n4nd0so if for any reason train is called again, it will fail09:46
sonne|workyeah but you can do that in the train() function09:46
sonne|workthere you have distance(train,train)09:47
n4nd0aham ok09:47
n4nd0and when it finishes, should it come back to the initial state09:47
n4nd0e.g. if train is called after apply, then at exit it should be again distance(train, test)09:48
n4nd0or is it not necessary?09:48
sonne|workbut apart from that I don't mind any distance_lhs() / distance_rhs or kernel_lhs/rhs functions09:48
sonne|worknot necessary09:48
sonne|workit usually is KNN.train()09:48
sonne|workthen only KNN.test09:48
sonne|workapply I mean09:49
n4nd0well I say it because if is done something like09:50
n4nd0KNN.train(train1)09:50
n4nd0KNN.apply(test1)09:50
n4nd0KNN.train(train2)09:50
n4nd0KNN.apply()09:50
n4nd0then it would make the second apply taking train2 as the vectors to classify09:51
n4nd0gtg now, bye!09:54
sonne|workn4nd0: yes exactly that is how it is supposed to be10:06
n4nd0sonne|work: ok no problem then, no need to have this distance_lhs and distance_rhs I talked about10:19
sonne|workof course if this behavior is not logical - we have to change things10:20
sonne|workit has been so since the dawn of days so I might be too blind already to see the obvious10:20
-!- uricamic [9320543b@gateway/web/freenode/ip.147.32.84.59] has joined #shogun10:42
sonne|workhi uricamic11:10
uricamichi11:10
n4nd0well guys I am going to travel now11:18
n4nd0talk to you later!11:19
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving]11:19
-!- blackburn [5bdfb203@gateway/web/freenode/ip.91.223.178.3] has joined #shogun11:31
blackburnwazzup11:31
sonne|workblackburn: what was up with newtonsvm?11:52
blackburnsonne|work: crashed with not enough memory actually, I should try different data11:52
sonne|workhmmhh memory leak or other bug?11:52
-!- PhilTillet [~Philippe@tillet-p42154.maisel.int-evry.fr] has joined #shogun11:53
blackburnsonne|work: no, can be caused by lack of memory actually11:54
blackburnthe data is rather big11:54
blackburnI'll ceck again11:54
blackburncheck* :)11:58
-!- PhilTillet [~Philippe@tillet-p42154.maisel.int-evry.fr] has quit [Ping timeout: 276 seconds]11:58
-!- pythonroar_ [d2198538@gateway/web/freenode/ip.210.25.133.56] has joined #shogun12:06
-!- PhilTillet [~Philippe@npasserelle10.minet.net] has joined #shogun12:23
-!- Varagrawal [0e8b7952@gateway/web/freenode/ip.14.139.121.82] has joined #shogun13:12
VaragrawalHey guys13:13
VaragrawalSorry I wasn't active here for the past few days. Had my GRE this last week13:13
VaragrawalI am interested in helping out for the Fast Newton Based SVM solver and the ECOC porting projects13:17
VaragrawalI am quite new to statistical machine learning, but I have some basic to intermediate experience with it as well as I am quite familiar with Matlab and Octave13:18
VaragrawalIs Soeren sir around?13:18
blackburnhi13:19
blackburnnewton based svm has been recently merged13:19
blackburnif you want to work on ECOC you'd have to hurry up, gsoc proposal deadline is tomorrow13:25
-!- PhilTillet [~Philippe@npasserelle10.minet.net] has quit [Ping timeout: 244 seconds]13:41
Varagrawalrecently merged?13:50
VaragrawalYou mean it's been done already??13:50
Varagrawal:(13:50
-!- shogun-toolbox [~shogun@7nn.de] has quit [Ping timeout: 245 seconds]13:50
--- Log closed Thu Apr 05 13:50:34 2012
--- Log opened Thu Apr 05 13:50:42 2012
-!- shogun-toolbox [~shogun@7nn.de] has joined #shogun13:50
-!- Irssi: #shogun: Total of 13 nicks [1 ops, 0 halfops, 0 voices, 12 normal]13:50
-!- Irssi: Join to #shogun was synced in 7 secs13:50
Varagrawalanybody out there?13:55
blackburnVaragrawal: yes, newton svm is done13:58
Varagrawal:(13:59
Varagrawalcool, anything else that might seem a good fit for me?13:59
blackburnwhat do you prefer?14:00
blackburnwe always welcome new examples14:01
Varagrawalwell I am not very experienced in ML as I am self-taught and still an undergrad14:02
Varagrawalbut when Shogun was selected for GSoC, this was the only project I had my eye one.14:02
-!- gsomix [~gsomix@188.168.14.72] has joined #shogun14:03
gsomixhi.14:03
blackburnhey there14:04
gsomixtoday's lecture about philosophy was interesting.14:07
blackburnVaragrawal: strange you have chosen shogun then :)14:08
VaragrawalI am interested in Machine Learning and want to become as good as you guys14:08
VaragrawalI just got killed coz I had my midsem exams in mid-march and then my GRE the previous week :'(14:09
blackburnI still think to develop some nice graphical example is a good way to learn14:09
-!- harshit_ [~harshit@182.68.94.221] has joined #shogun14:11
sonne|workgsomix: seen the 2 other ideas I had?14:11
VaragrawalI am going through the tutorial pages of shogun14:11
VaragrawalI had cloned the repo14:11
Varagrawalbut then I ws getting this weird error for one application. Nando know about it and he had absolutely no clue14:12
gsomixsonne|work, CSet -> hash set?14:12
blackburnVaragrawal: we are on the way to fix applications, you should rather check examples14:12
blackburnapplications are pretty complex and domain specific14:12
VaragrawalI was just running them to check how breaking some code would affect the application14:13
sonne|workgsomix: yes and SGArray removal14:13
sonne|work(replace by SGVector)14:13
sonne|workgsomix: but now I really think this is enough...14:14
sonne|workexcept if you intend to work 60+ hrs :D14:14
blackburnsonne|work: he will be able!14:15
blackburnsonne|work: 90 is ok for him14:15
gsomixsonne|work, i dont have 60+ hrs. :) i think such changes can be done if there is free time.14:15
gsomixblackburn, >:314:16
blackburngsomix: you do! I recall you wanted to work for a month w/o sleeping14:16
sonne|workI am sure that with the help of vodka you both will manage14:17
gsomixblackburn, okay, master. =___=14:17
blackburnsonne|work: the vodka14:17
sonne|workyes vodking?14:17
sonne|workha :D14:17
blackburnvod-king?14:18
blackburngsomix: I have some small task for you14:18
blackburnall you need is to implement distributed linear svm core faster than liblinear14:19
blackburnshould be easy14:19
gsomixsonne|work, anyway I want to work on shogun after GSoC. It's approximately infinity time. :]14:19
gsomixblackburn, wut? you're at work?14:20
blackburngsomix: I am14:20
-!- PhilTillet [~Philippe@npasserelle10.minet.net] has joined #shogun14:23
gsomixprofessor of philosophy said that the Russian and German languages are better suited for metaphysics than others14:23
gsomixfun fact14:24
blackburnI guess the reason is he knows only these two ones14:24
gsomixblackburn, and a little Sanskrit and English14:26
blackburnreally? he tends to pronounce english words in a funny way14:27
blackburnhowever this fact could arise for real14:27
blackburnI mean native language determines the way of thinking often14:28
gsomixI see14:29
blackburngsomix: and even programming languages do the same14:29
-!- sidsniper [~piggy@14.139.82.6] has joined #shogun14:38
PhilTilletHello everybody14:39
blackburnhi14:39
PhilTilletwhat's up?14:39
sidsnipersky :P14:39
PhilTilletlol.14:40
PhilTilletand avian carriers?14:40
PhilTillet:D14:40
sidsniperi was wondering how many slots will shogun get this time?14:40
sonne|workwe are too14:40
blackburnhah14:40
PhilTilletrand() slots14:43
-!- sidsniper [~piggy@14.139.82.6] has left #shogun []14:43
-!- sidsnipe [~piggy@14.139.82.6] has joined #shogun14:43
blackburnno, I don't like rand stuff there14:45
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun14:46
PhilTillet:D let's just hope it will be a Dirac Distribution14:46
PhilTillet:D14:46
PhilTilletrand over dirac distribution is cool14:46
-!- siddharthk [~siddharth@14.139.82.6] has joined #shogun14:47
gsomixnumber of slots is determined by position of the planets.14:47
siddharthkerr is it too late to apply now?14:47
-!- sidsnipe [~piggy@14.139.82.6] has left #shogun []14:48
blackburnsiddharthk: due date is tomorrow14:48
siddharthkare there any free projects for which i can apply for?14:50
blackburnsiddharthk: what do you mean free? ;)14:51
siddharthkwhich has less competition :P14:51
PhilTilletsonney2k, maybe I can release a first iteration of the OpenCL patch tomorrow, but design will really not be optimal :D As computing kernel values one per one is completely inefficient (and the whole shogun design is kinda based on that :D)14:51
siddharthkprevious year me n u applied for same project :P14:51
siddharthki got axed :/14:52
blackburnsiddharthk: yeah14:52
PhilTilletoops, should have been sonne|work rather than sonney2k maybe :D14:52
blackburnsiddharthk: ok libqp and multiclass probably have less competition there14:53
siddharthki can see that there are more people now :)14:53
blackburnhowever I can't be sure14:53
siddharthkok let me check ideas page14:53
blackburnnot really last year there were 70+ appliants14:53
siddharthki mean on irc14:53
blackburnah14:54
blackburnyeah may be14:54
siddharthki hope patch will not be a problem this time14:55
-!- pythonroar_ [d2198538@gateway/web/freenode/ip.210.25.133.56] has quit [Ping timeout: 245 seconds]14:55
blackburnwhat do you mean?14:55
siddharthki mean previous year patches would count right? :P14:57
blackburnyes however I think to have this year patches would be better ;)14:59
harshit_blackburn: Can i send a pull request for shougun-toolbox/shogun-data ?15:01
blackburnharshit_: yeah probably but you would need to fork it too15:01
blackburnwhat do you need to change?15:01
harshit_I tested my octave_modular example on splice dataset, So can i add splice dataset in it15:02
harshit_its not too big15:02
harshit_and for now I have changed my octave example on toydataset15:02
blackburnthat's ok15:02
blackburnwe use toy data in examples15:02
harshit_okay ..15:03
PhilTilletblackburn, you once told me about Dimension Reduction on OpenCL but I was not here :D15:04
blackburnPhilTillet: yes that is something of interest probably15:04
PhilTilletbut, this is only simple BLAS Matrix operations right?15:04
PhilTilletI mean, for PCA at least15:04
blackburnyes15:04
PhilTilletI think it will be very easy to port in a few months when ViennaCL will be more advanced :p they plan to include some eigenvalues algorithms on GPU15:05
PhilTilletthen it would be possible without writing any OpenCL code :p15:05
-!- harshit_ [~harshit@182.68.94.221] has quit [Quit: Leaving]15:05
blackburnI don't think they will include efficient sparse eigensolver15:06
PhilTillethmmm15:06
PhilTilletthis was a GSoC project last year as far as I remember15:06
blackburnLLE/LTSA/HLLE/.... are sparse15:06
blackburnreally? I'll take a look later15:06
PhilTillethmmm yes indeed15:07
PhilTilletmaybe it was not sparse15:07
PhilTilletIt was mainly Lanczos and Arnoldi's15:07
blackburnlanczos and arnoldi stuff makes use of matrix-vector product15:08
blackburnso they could be sparse15:08
PhilTilletyep15:08
PhilTilletthere is sparse matrix-vector implemented15:08
blackburnhowever in shogun I had to integrate superlu15:08
blackburnthere was need of sparse solver15:08
PhilTilletwell15:08
PhilTilletthere is GPU LU decomposition on ViennaCL too I think15:09
PhilTillet:p15:09
blackburnsuperlu is sparse (again this word) ;)15:09
PhilTillethaha15:09
PhilTilletsparsity yay15:09
PhilTilletactually i'm not an expert concerning ViennaCL, but if you have some question concerning what linalg operation is implemented, I can relay them to the project leader15:10
blackburnok thanks15:11
blackburnokay gtg15:11
-!- blackburn [5bdfb203@gateway/web/freenode/ip.91.223.178.3] has quit [Quit: Page closed]15:11
-!- gsomix [~gsomix@188.168.14.72] has quit [Ping timeout: 260 seconds]15:12
-!- harshit_ [~harshit@182.68.94.221] has joined #shogun15:26
CIA-64shogun: Harshit Syal master * r9241292 / (2 files in 2 dirs): Added brief intro to NewtonSVM and added octave_modular example - http://git.io/xpkZDQ15:36
CIA-64shogun: Harshit Syal master * rfc37445 / src/shogun/classifier/svm/NewtonSVM.h : Removed some typo - http://git.io/QQrc3g15:36
CIA-64shogun: Soeren Sonnenburg master * r39e947f / (2 files in 2 dirs):15:36
CIA-64shogun: Merge pull request #421 from harshitsyal/master15:36
CIA-64shogun: Newton SVM : improve documentation - http://git.io/wJaZLg15:36
-!- harshit_ [~harshit@182.68.94.221] has quit [Ping timeout: 260 seconds]15:40
-!- ani1 [dcff023b@gateway/web/freenode/ip.220.255.2.59] has joined #shogun15:43
-!- ani1 [dcff023b@gateway/web/freenode/ip.220.255.2.59] has quit [Ping timeout: 245 seconds]16:05
-!- gsomix [~gsomix@188.168.3.37] has joined #shogun16:07
-!- Kshiteesh [~KMH@c-24-118-111-180.hsd1.mn.comcast.net] has joined #shogun16:32
-!- harshit_ [~harshit@182.68.94.221] has joined #shogun16:32
-!- PhilTillet_ [~Philippe@tillet-p42154.maisel.int-evry.fr] has joined #shogun16:43
-!- PhilTillet [~Philippe@npasserelle10.minet.net] has quit [Ping timeout: 246 seconds]16:44
-!- sonne|work [~sonnenbu@194.78.35.195] has quit [Quit: Leaving.]16:49
PhilTillet_sonney2k, I've got some lil' problem with the ./configure16:58
PhilTillet_when I used ./consigure --cxxflags=-DUSE_OPENCL, it compiles fine but removes the -fPIC16:59
PhilTillet_so it fails at linking16:59
PhilTillet_(but at least it finally compiles, but cannot test it :D)17:02
-!- harshit_ [~harshit@182.68.94.221] has quit [Ping timeout: 264 seconds]17:09
-!- harshit_ [~harshit@182.68.94.221] has joined #shogun17:14
-!- Kshiteesh [~KMH@c-24-118-111-180.hsd1.mn.comcast.net] has quit [Ping timeout: 265 seconds]17:20
@sonney2kPhilTillet_, well then use --cxxflags="-DUSE_OPENCL -fPIC"17:20
PhilTillet_oh of course17:21
PhilTillet_i had forgotten the quotes... kinda tired :D17:21
PhilTillet_./configure --cxxflags="-DUSE_OPENCL -fPIC" => Error: Unknown parameter: -fPIC17:22
PhilTillet_tricky17:22
PhilTillet_Sounds like the configure script is trolling me.17:23
@sonney2kPhilTillet_, argh then use CXXFLAGS=-DUSE_OPENCL ./configure17:26
-!- n4nd0 [~androirc@h121n3-ksg-a11.ias.bredband.telia.com] has joined #shogun17:38
-!- harshit_ [~harshit@182.68.94.221] has quit [Ping timeout: 265 seconds]17:46
-!- flxb [~cronor@g231228114.adsl.alicedsl.de] has quit [Read error: Connection reset by peer]18:05
-!- n4nd0 [~androirc@h121n3-ksg-a11.ias.bredband.telia.com] has quit [Remote host closed the connection]18:07
-!- harshit_ [~harshit@182.68.94.221] has joined #shogun18:13
-!- flxb [~cronor@g231228114.adsl.alicedsl.de] has joined #shogun18:20
shogun-buildbotbuild #438 of ruby_modular is complete: Failure [failed test_1]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/ruby_modular/builds/438  blamelist: sonne@debian.org, harshit.syal@nsitonline.in18:30
-!- harshit_ [~harshit@182.68.94.221] has quit [Ping timeout: 264 seconds]18:35
PhilTillet_sonney2k, how can I add files to the build?18:37
-!- PhilTillet_ [~Philippe@tillet-p42154.maisel.int-evry.fr] has quit [Read error: Operation timed out]19:00
-!- PhilTillet_ [~Philippe@npasserelle10.minet.net] has joined #shogun19:16
-!- PhilTillet_ [~Philippe@npasserelle10.minet.net] has quit [Ping timeout: 244 seconds]19:23
-!- n4nd0 [~nando@h121n3-ksg-a11.ias.bredband.telia.com] has joined #shogun19:34
-!- PhilTillet_ [~Philippe@tillet-p42154.maisel.int-evry.fr] has joined #shogun19:36
-!- harshit_ [~harshit@182.68.94.221] has joined #shogun19:39
-!- Kshiteesh [~KMH@nat-portal-160-94-47-16.uofm.wireless.umn.edu] has joined #shogun19:40
-!- uricamic [9320543b@gateway/web/freenode/ip.147.32.84.59] has quit [Quit: Page closed]19:40
wikingPhilTillet_: well if it's a cpp file it'll just compile it19:50
wikingbut of course if you want to add it to the swig interface as well19:51
wikingthen you'll have to edit some files for includes..19:51
-!- harshit_ [~harshit@182.68.94.221] has quit [Ping timeout: 260 seconds]19:57
PhilTillet_the swig interface ? :p20:00
PhilTillet_isn't autotools used?20:00
wikingwell if you want to support modules20:12
wikinglike python_modular20:12
wikingand such then you'll want swig20:12
wikingand for that you have to do some extra work...20:12
PhilTillet_yes I know a bit of swig, headache20:13
@sonney2kno autotools all magic20:16
PhilTillet_sounds like fun20:18
PhilTillet_XD20:18
PhilTillet_oh, and some other question, I used #ifdef USE_OPENCL about everywhere, but then once you have the library you have to do #define USE_OPENCL so that you see the functions, any way to avoid this? :p20:19
@sonney2kPhilTillet_, that can be taken care of later via some configure option20:25
PhilTillet_oh yes, true20:26
PhilTillet_I hate coding with #ifdef20:26
PhilTillet_cause usually ide don't parse the #ifdef'ed code20:27
PhilTillet_><20:27
-!- harshit_ [~harshit@182.68.94.221] has joined #shogun20:27
n4nd0PhilTillet_: what IDE do you use?20:27
@sonney2kPhilTillet_, just to explain: in shogun the kernel matrix is assumed to *not* fit in memory20:27
PhilTillet_sonney2k, I know, but I cannot implement that in 1 week :D Making it work for big dataset is the whole point of the SoC20:28
PhilTillet_:p20:28
@sonney2kthat is why you need to call kernel(i,j) most of the time. actually even examples must not be in memory20:28
@sonney2kand are computed on-the-fly20:28
PhilTillet_n4nd0, KDevelop20:28
@sonney2kok20:29
PhilTillet_I mean it will require some clever design, but I am sure it is possible to do some queueing system, with flushing20:29
PhilTillet_and, even if it is not, dot product is still faster on GPU than on CPU20:29
PhilTillet_usually20:30
PhilTillet_it should suffer from bad caching though :p20:30
-!- Johannes_ [55e5c1c3@gateway/web/freenode/ip.85.229.193.195] has quit [Ping timeout: 245 seconds]20:32
@sonney2kwell you certainly can do speedups for special cases - like when simplefeatures are used and the feature matrix is in memory20:32
PhilTillet_yes20:33
PhilTillet_and when simplefeatures matrix do not fit in memory, it should be possible to transfer it chunks by chunk, for apply at least20:35
PhilTillet_even though the training might get more complicated20:35
PhilTillet_and, even for training, something clever might be to use hybrid cpu/gpu algorithm20:35
PhilTillet_but synchronization might then be a pain20:36
@sonney2kPhilTillet_, well lets focus on testing first20:39
@sonney2ktraining usually has to be done once20:39
@sonney2ktesting inf times20:39
PhilTillet_true :d20:39
PhilTillet_I mean, for testing, examples are independent, so it might be possible to do something like : "Transfer chunk 1 from CPU to GPU, compute chunk 2 on CPU, get chunk 1 result, transfer chunk 3 to GPU, compute chunk 4 on CPU20:41
PhilTillet_things like that20:41
PhilTillet_i'll spend some time again on my gsoc timeline, to focus on hybrid CPU / GPU algorithm20:42
harshit_sonney2k: Is fast k-Means algo implemented in shogun ? I saw this in issues section of github .20:45
harshit_link:https://github.com/shogun-toolbox/shogun/issues/33820:46
@sonney2kharshit_, no unfortunately not20:46
harshit_sonney2k: Its mentioned by authors that you cannot redistribute this code . but still can we port it ?20:47
n4nd0I am working in cover tree integration now20:47
@sonney2kn4nd0, for KNN right?20:47
n4nd0sonney2k: yeah20:48
@sonney2kharshit_, well the algoritm / paper is very simple20:48
@sonney2kprobably devil is in the detail but still one can compare outputs of what they get to what you would implement20:48
harshit_n4nd0 : is cover tree , in anyway  thats related to fast k-means .. i think it wound not but still confirming ..20:49
harshit_coz I am always confused about differences in KNN and K means except one is unsupervised20:50
harshit_sonney2k: yeah algo is really simple..got one more thing to work on :)20:50
@sonney2kyeah covertree would speed up k-means too - in the end you need distances for that too20:56
n4nd0it is turning out to be a bit cumbersome though :S20:57
-!- siddharthk [~siddharth@14.139.82.6] has quit []21:20
flxbIs there a possibility to save a GMNPSVM without using pickle? pickle is too slow, because the classifier is 15GB and cPickle has a memory bug.21:40
@sonney2kflxb, look at serialization_complex_example.py21:42
-!- Kshiteesh [~KMH@nat-portal-160-94-47-16.uofm.wireless.umn.edu] has quit [Ping timeout: 252 seconds]21:46
PhilTillet_sonney2k, is the kernel matrix always double precision?21:57
-!- Varagrawal [0e8b7952@gateway/web/freenode/ip.14.139.121.82] has quit [Ping timeout: 245 seconds]22:00
flxbthanks, sonney2k22:17
-!- blackburn [~qdrgsm@83.234.54.186] has joined #shogun22:18
n4nd0blackburn: hey! I got a question here22:22
blackburnyeah22:22
n4nd0blackburn: I am analizing with valgrind a python example22:22
gsomixblackburn, hello, master22:22
blackburngsomix: lol22:23
gsomix>:322:23
blackburnn4nd0: sooo22:23
n4nd0blackburn: http://snipt.org/uhaji022:23
blackburnaham22:24
n4nd0blackburn: I am giving you the code where it happens, w822:24
n4nd0https://github.com/shogun-toolbox/shogun/blob/master/src/shogun/classifier/KNN.cpp22:24
n4nd0line 235 there22:25
n4nd0it appears 296 in valgrind because I have executed it in my local version with CoverTree stuff, but this is not related22:26
n4nd0no CoverTree used there yet22:26
blackburnok  pretty strange22:26
blackburnI'd say it is a false alarm there22:26
@sonney2kn4nd0, please change the line SGMatrix<int32_t>(output,num_lab,m_k)22:26
@sonney2kto SGMatrix<int32_t>(output,num_lab,m_k,true)22:26
@sonney2kthen it should be fixed22:26
blackburnyeah had the same idea22:27
n4nd0sonney2k, blackburn: you guys are pretty good :)22:28
n4nd0fixed22:28
n4nd0sonney2k: do you have a good explanation for it?22:29
@sonney2kyeah - do valgrind only w/ debugging flags on + no optimizations22:30
@sonney2kotherwise line nr might not fit22:30
blackburnwhat is the reproducing example?22:30
n4nd0classifier_knn_python_modular.py22:30
blackburnah22:31
n4nd0sonney2k: I didn't understand ... I mean the line valgrind was right I think22:31
blackburnn4nd0: true stands for do_free22:32
@sonney2kthen I don't understand your question :)22:32
n4nd0sonney2k: because the part of memory not freed is in output = SG_MALLOC(...)22:32
n4nd0sonney2k: yeah the question is, why do we need to set explicitily this do_free?22:33
PhilTillet_sonney2k, the patch works and should be issued soon :p22:37
PhilTillet_it is in a new branch, but do you want it as a single commit?22:39
@sonney2kn4nd0, this do_free needs to be set whenever you alloc sth that is returned / if you just pass a ptr then it should be false22:39
@sonney2kbut this is what I would like to change with the refcounting stuff22:39
n4nd0sonney2k: ok, I understand22:44
gsomixsonney2k, there is small pull request. It's related to python3 interface. I hope I will send fixes for examples tomorrow.22:45
@sonney2kgsomix, ok22:45
@sonney2kPhilTillet_, not sure what you mean - something we can review where each commit compiles would be nice22:46
PhilTillet_ok, my question was dumb indeed22:46
PhilTillet_i'll discard the bugfix commits :D22:47
@sonney2k(additional / unrelated bugfix commits as extra pull requests please)22:48
@sonney2kn4nd0, please send a pull request with that one-liner KNN patch22:48
PhilTillet_sonney2k, what do you mean?22:48
@sonney2kbugfixes you did on the side22:48
PhilTillet_like bugfixes on my generator?22:48
n4nd0sonney2k: do you want me to prepare an example of performance improvement first?22:48
PhilTillet_oops22:49
PhilTillet_not generator22:49
PhilTillet_bugfix on my opencl support?22:49
PhilTillet_i did no additional shogun bugfix in this opencl branch if this is the question22:50
PhilTillet_:p22:50
@sonney2kn4nd0, it would be interesting to know how much the covertree helps or?22:50
n4nd0sonney2k: I was planning to do that and later issue the pull request22:51
n4nd0sonney2k: but I can open it before the example if you want so we can start fixing things I may have done wrong22:51
@sonney2kn4nd0, why not22:58
gsomixsonney2k, I have some problems with my "const char*" typemap. http://pastebin.com/tiq44rbE23:03
gsomixwho should take care of freeing memory here?23:04
gsomixthere is part in python3 doc: http://docs.python.org/py3k/c-api/bytes.html?highlight=pybytes_asstring#PyBytes_AsString23:05
@sonney2kgsomix, i don't think const char* typemaps can work23:05
@sonney2kbut I am not an expert in this either23:06
@sonney2kbut IMHO char* is automagically swig wrapped23:06
@sonney2kso anything doing char* foo()23:06
@sonney2kor void foo(char* x) should work23:06
gsomixsonney2k, i had problems in modelselection example.23:07
gsomixfor example, line: param_power_kernel_degree=ModelSelectionParameters("degree")23:07
gsomixSWIG incorrectly converted "degree" into const char*23:08
gsomixin python323:08
gsomixthen I wrote simple typemap.23:09
gsomixat now it works. but I don't know how well.23:10
@sonney2kI think we have to remove the const* stuff somewhere then23:11
n4nd0well guys, good night23:12
n4nd0sonney2k: I will issue a PR with covertree stuff tomorrow morning probably23:12
gsomixn4nd0, good night23:13
@sonney2kn4nd0, nite nite23:13
-!- n4nd0 [~nando@h121n3-ksg-a11.ias.bredband.telia.com] has quit [Quit: leaving]23:13
gsomixsonney2k, ok. while I will try to research this issue in more details.23:14
CIA-64shogun: Evgeniy Andreev master * r6ea66a8 / src/interfaces/modular/SGBase.i : fixed pickle serialization - http://git.io/c1j6-w23:18
CIA-64shogun: Evgeniy Andreev master * r2320875 / src/interfaces/python_modular/swig_typemaps.i : unicode -> const char* and fixed typemaps for int32 support - http://git.io/9K0bSQ23:18
CIA-64shogun: Soeren Sonnenburg master * r1eae5da / (2 files in 2 dirs):23:18
CIA-64shogun: Merge pull request #423 from gsomix/python3_interface23:19
CIA-64shogun: Updates on python3 interface - http://git.io/PIfgGQ23:19
gsomixoh, tomorrow classes will be about programming under linux.23:19
gsomixwhere my physics now? :)23:19
gsomixgood night, guys23:20
PhilTillet_good night23:20
@sonney2kcu23:20
PhilTillet_sonney2k, where can I add the example for using the OpenCL apply() functionnality?23:25
PhilTillet_which directory23:25
@sonney2kalright sleep time for me23:25
PhilTillet_okay, good night :p23:25
@sonney2kPhilTillet_, which $LANG?23:25
PhilTillet_C++23:25
@sonney2kexamples/undocumented/libshogun then23:25
PhilTillet_thanks23:25
@sonney2knite23:26
PhilTillet_night23:26
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]23:27
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun23:42
--- Log closed Fri Apr 06 00:00:19 2012