--- Log opened Mon Apr 02 00:00:13 2012 | ||
--- Day changed Mon Apr 02 2012 | ||
harshit_ | okay.doing that .. | 00:00 |
---|---|---|
n4nd0 | sonney2k, blackburn new pull request :) | 00:00 |
blackburn | it seems I stare to github whole day long hah | 00:00 |
blackburn | n4nd0: I'll review it in a min | 00:01 |
n4nd0 | blackburn: thank you man, but no hurries, it is kind of late already | 00:01 |
@sonney2k | harshit_, I am starting to fall asleep - but I want to see the updated output before I do so please do it now if you can | 00:01 |
n4nd0 | blackburn: just tell me a general opinion if so | 00:01 |
blackburn | n4nd0: do you prefer commenting code here or at github? ;) | 00:02 |
@sonney2k | n4nd0, cool figures | 00:02 |
n4nd0 | blackburn: let's do it at github better | 00:02 |
blackburn | ok | 00:02 |
n4nd0 | sonney2k: thank you :D | 00:02 |
n4nd0 | sonney2k: did you see the one I prepared for multiclass svm? | 00:03 |
@sonney2k | no | 00:03 |
n4nd0 | sonney2k: http://dl.dropbox.com/u/11020840/shogun/multiclass_svm.png | 00:04 |
n4nd0 | sonney2k: I didn't expect that flower shape haha | 00:04 |
@sonney2k | cool | 00:04 |
@sonney2k | I should check whether my son recognizes this as flower :D | 00:05 |
@sonney2k | saying bluuu bluu (flower == blume in german) | 00:05 |
n4nd0 | haha | 00:05 |
n4nd0 | blomma in Swedish | 00:05 |
n4nd0 | there are similarities | 00:06 |
blackburn | sonney2k: your son is kind of talking already?! | 00:06 |
blackburn | n4nd0: are you sure it works only for euclidean? | 00:07 |
@sonney2k | might be an illusion but I think he starts to recognizing flowers - at least every time I got close to one he is making this weird bluuu bluee sound | 00:08 |
n4nd0 | blackburn: I didn't check it :S | 00:08 |
blackburn | sonney2k: cool | 00:08 |
harshit_ | sonney2k: hey figured that out .. it was not the problem in the values , it was the problem with the way i was printing them | 00:09 |
harshit_ | the actual problem was something else | 00:09 |
harshit_ | :) | 00:09 |
harshit_ | :) | 00:09 |
-!- PhilTillet [~Philippe@tillet-p42154.maisel.int-evry.fr] has joined #shogun | 00:10 | |
@sonney2k | harshit_, so enlighten us! | 00:10 |
harshit_ | actually problem was that i was resetting value of t to 0 in function line_search_linear | 00:13 |
harshit_ | where as it needed to be calculated by using its previous value | 00:14 |
harshit_ | Just need some more testing then i'll send you the final test results on both c++ and matlab | 00:15 |
shogun-buildbot | build #652 of libshogun is complete: Failure [failed compile] Build details are at http://www.shogun-toolbox.org/buildbot/builders/libshogun/builds/652 blamelist: sonne@debian.org, shelhamer@imaginarynumber.net | 00:18 |
blackburn | ah yes btw it fails | 00:19 |
blackburn | :D | 00:19 |
blackburn | I'll take care | 00:24 |
CIA-64 | shogun: Soeren Sonnenburg master * ra2a559c / testsuite/python_modular/tester.py : add -m option to tester to show only missing tests - http://git.io/4kFzJA | 00:27 |
@sonney2k | nite folks | 00:27 |
blackburn | sonney2k: nite | 00:27 |
n4nd0 | good night | 00:27 |
@sonney2k | harshit_, good to hear - I think we will need to compare newton svm from C++ to some bigger data set once you are confident again that it works | 00:28 |
@sonney2k | cu | 00:28 |
harshit_ | sonney2k : wait | 00:28 |
harshit_ | for 2 min | 00:28 |
harshit_ | see here : http://snipt.org/ugiic5 | 00:28 |
blackburn | no way :D | 00:28 |
harshit_ | is he gone ? | 00:29 |
blackburn | yes | 00:29 |
harshit_ | shit | 00:29 |
harshit_ | then please you have a look | 00:29 |
n4nd0 | I had enough for today too guys | 00:30 |
n4nd0 | good night | 00:30 |
blackburn | hah | 00:30 |
-!- Vuvu [~Vivan_Ric@115.248.130.148] has quit [Remote host closed the connection] | 00:30 | |
-!- romi_ [~mizobe@187.74.1.223] has quit [Quit: Leaving] | 00:30 | |
blackburn | n4nd0: nite | 00:30 |
harshit_ | good bye n3nd0 | 00:30 |
n4nd0 | harshit_: idk if I can help you with that, what is the problem? | 00:31 |
harshit_ | problem is with precision of float64_t | 00:31 |
n4nd0 | mmm no idea what's going on there | 00:32 |
blackburn | yeah hard to digest | 00:32 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has left #shogun [] | 00:32 | |
-!- Vuvu [~Vivan_Ric@115.248.130.148] has joined #shogun | 00:34 | |
blackburn | harshit_: hey but h's are different! | 00:36 |
harshit_ | I think i have messed up a lil with the label values | 00:37 |
harshit_ | in sleep :( | 00:37 |
blackburn | hah no worries | 00:38 |
harshit_ | dont knw what i am doing, I think i should continue tmrw | 00:38 |
harshit_ | maybe everything requires a clean test again | 00:38 |
CIA-64 | shogun: Sergey Lisitsyn master * ra17ad85 / (3 files): Fixes for SG_ADD conversion - http://git.io/-T635w | 00:38 |
harshit_ | blackburn: hey do you think liblinear + lbp features + wiking preprocessor will make a good proposal | 00:39 |
harshit_ | vs | 00:39 |
harshit_ | c5.0 + liblinear | 00:39 |
blackburn | c5.0 is much more time demanding and important for me | 00:39 |
harshit_ | may be i should submit both , what say ? | 00:40 |
blackburn | feel free if you want to do that :) | 00:41 |
-!- av3ngr [av3ngr@nat/redhat/x-mtbcichayakoymjc] has joined #shogun | 00:42 | |
harshit_ | btw would you also apply for gsoc ? | 00:42 |
harshit_ | if so then what project would you be working on ? | 00:45 |
blackburn | yes, for multitask learning | 00:48 |
-!- PhilTillet [~Philippe@tillet-p42154.maisel.int-evry.fr] has quit [Ping timeout: 252 seconds] | 00:48 | |
harshit_ | transfer learning is really nice field :) would be nice to see it in shogun | 00:52 |
harshit_ | good bye, going to sleep. | 00:53 |
-!- harshit_ [~harshit@182.68.67.61] has quit [Remote host closed the connection] | 00:55 | |
blackburn | good night | 00:57 |
-!- PhilTillet [~Philippe@npasserelle10.minet.net] has joined #shogun | 01:00 | |
-!- PhilTillet [~Philippe@npasserelle10.minet.net] has quit [Ping timeout: 245 seconds] | 01:13 | |
-!- PhilTillet [~Philippe@npasserelle10.minet.net] has joined #shogun | 01:17 | |
-!- blackburn [~qdrgsm@31.28.44.229] has quit [Quit: Leaving.] | 01:20 | |
shogun-buildbot | build #654 of libshogun is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/libshogun/builds/654 | 01:31 |
-!- PhilTillet [~Philippe@npasserelle10.minet.net] has quit [Ping timeout: 265 seconds] | 01:36 | |
-!- flxb [~cronor@e178169137.adsl.alicedsl.de] has left #shogun [] | 01:40 | |
-!- Vuvu [~Vivan_Ric@115.248.130.148] has quit [Quit: Leaving] | 06:06 | |
-!- harshit_ [~harshit@182.68.67.61] has joined #shogun | 06:43 | |
-!- harshit_ [~harshit@182.68.67.61] has quit [Ping timeout: 246 seconds] | 07:39 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 07:40 | |
-!- harshit_ [~harshit@182.68.67.61] has joined #shogun | 07:48 | |
-!- menonnik [b4953181@gateway/web/freenode/ip.180.149.49.129] has joined #shogun | 08:00 | |
@sonney2k | harshit_, for liblinear/ocas not a lot of work (maybe close to nothing) might be necessary - so implementing other dotfeatures might be sth to focus on and of course if you intend to work on trees in general - also fine | 08:10 |
harshit_ | sonney2k: So would you prefer liblinear + c5.0 | 08:11 |
harshit_ | or liblinear + lbp + some other features | 08:12 |
@sonney2k | dotfeatures + c5.x | 08:12 |
@sonney2k | or other trees | 08:12 |
harshit_ | ohk got it, so trees are in demand. | 08:13 |
harshit_ | sonney2k : for Newton SVM , it was definitely not precision issue | 08:13 |
@sonney2k | well we don't have any trees in there - so it would be nice to have any | 08:13 |
harshit_ | there is some other small problem, which i am searching for last 3hrs | 08:14 |
harshit_ | will get you on it,As soon i'll find it | 08:14 |
@sonney2k | but other stuff is also fine - we will have to see how many slots we get and how we can divide work then so a little flexibility is necessary | 08:14 |
@sonney2k | harshit_, thanks! | 08:14 |
harshit_ | so in trees would c5.0 will be the best to implement, the c5.0 released under gnu license | 08:15 |
harshit_ | or you have something else in your mind ? | 08:15 |
harshit_ | Also i dont want to waste the current time, So want to start working on LBP features | 08:16 |
harshit_ | Saw opencv code: it looks good | 08:17 |
CIA-64 | shogun: Evan Shelhamer master * r4f1e9e5 / src/README.developer : Cleanup doc whitespace and word wrap - http://git.io/7Afz_Q | 08:17 |
CIA-64 | shogun: Evan Shelhamer master * r787d681 / src/README.developer : (log message trimmed) | 08:17 |
CIA-64 | shogun: Revise whitespace and versioning scheme sections of the developer readme | 08:17 |
CIA-64 | shogun: - include notice of newline convention (LF) and how to enforce it automatically | 08:17 |
CIA-64 | shogun: through git settings | 08:17 |
CIA-64 | shogun: - rephrase trailing whitespace caution and suggest means to automatically | 08:17 |
CIA-64 | shogun: strip trailing whitespace for emacs and vim | 08:17 |
CIA-64 | shogun: - update versioning section with info on github and suggested workflow | 08:17 |
CIA-64 | shogun: Soeren Sonnenburg master * r142a356 / src/README.developer : | 08:17 |
CIA-64 | shogun: Merge pull request #411 from shelhamer/readme-whitespace-and-versioning | 08:17 |
CIA-64 | shogun: Developer Readme whitespace and versioning - http://git.io/ss8t6w | 08:17 |
-!- menonnik [b4953181@gateway/web/freenode/ip.180.149.49.129] has quit [Quit: Page closed] | 08:26 | |
harshit_ | sonney2k: ^ | 08:26 |
-!- gsomix [~gsomix@188.168.4.80] has quit [Read error: Operation timed out] | 08:33 | |
-!- sonne|work [~sonnenbu@194.78.35.195] has joined #shogun | 08:51 | |
-!- mridul [~Adium@61.12.27.114] has joined #shogun | 08:59 | |
-!- gsomix [~gsomix@188.168.4.80] has joined #shogun | 09:04 | |
-!- harshit_ [~harshit@182.68.67.61] has quit [Ping timeout: 248 seconds] | 09:09 | |
-!- harshit_ [~harshit@182.68.67.61] has joined #shogun | 09:10 | |
sonne|work | moin gsomix | 09:10 |
harshit_ | sonney2k: https://gist.github.com/2281396, have a look at line 17,18,19 in C++ output and 18,19,20 in Matlab output | 09:29 |
harshit_ | n4nd0 : if you are around, please have a look at https://gist.github.com/2281396 | 09:30 |
n4nd0 | harshit_: so what's the problem? | 09:31 |
harshit_ | actually precision is the problem , | 09:31 |
harshit_ | as you can see in line 19 of C++ | 09:32 |
harshit_ | and in line 21 of matlab | 09:32 |
harshit_ | g = gpart1- gpart2 | 09:33 |
n4nd0 | aham | 09:33 |
harshit_ | g after subtraction comes out to be of order e-17 | 09:33 |
n4nd0 | there is a little error around here | 09:33 |
-!- mridul [~Adium@61.12.27.114] has left #shogun [] | 09:34 | |
n4nd0 | but what's g? is it the difference between the result you get with matlab and C++? | 09:34 |
harshit_ | yeah and thats bocz of high precision in matlab | 09:34 |
harshit_ | no g is just a variable used for getting the results | 09:34 |
n4nd0 | ok | 09:35 |
n4nd0 | but the difference between C++ and matlab is ~e-17?? | 09:35 |
-!- stephenlee [da18b3c4@gateway/web/freenode/ip.218.24.179.196] has joined #shogun | 09:36 | |
harshit_ | yeah its a very small difference but the problem is that in next line of code i compute g/h | 09:36 |
harshit_ | as h is also of order e-17 , g/h turn out to be some thing large | 09:37 |
harshit_ | and here the difference becomes prominent | 09:37 |
n4nd0 | aham | 09:38 |
n4nd0 | I see | 09:38 |
harshit_ | do you think i should move on with it .. | 09:41 |
harshit_ | coz it really doesnt make any big impact on final result | 09:42 |
n4nd0 | so the final result is ok? | 09:42 |
harshit_ | yeah in most cases, the difference is of order of 1e-5 at most | 09:44 |
harshit_ | in final result | 09:44 |
n4nd0 | in my opinion, that is ok | 09:44 |
n4nd0 | but you should ask sonney2k just in case | 09:45 |
harshit_ | have a look here : http://snipt.org/ugjp4 | 09:46 |
harshit_ | thats one of the big differences i have encountered | 09:47 |
harshit_ | do you think thats acceptable | 09:47 |
n4nd0 | I see | 09:47 |
n4nd0 | idk, I mean, there's probably no general answer to that | 09:47 |
n4nd0 | I guess it depends on application | 09:47 |
harshit_ | yeah, right may be i need to do some real tests | 09:48 |
harshit_ | on it | 09:48 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: Changing server] | 09:50 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 09:53 | |
n4nd0 | harshit_: have you already thought of any particular example to test it? | 09:54 |
harshit_ | not yet, why ? do you have any in your mind ? | 09:55 |
n4nd0 | mmm I would try with classification of toy data first | 09:57 |
n4nd0 | if you get good results / similar to MATLAB | 09:57 |
n4nd0 | it will be ok I guess | 09:57 |
harshit_ | ohk will do it tday, lets hope for best :) | 09:58 |
-!- flxb [~cronor@141.23.80.206] has joined #shogun | 10:03 | |
-!- flxb [~cronor@141.23.80.206] has quit [Quit: flxb] | 10:10 | |
-!- flxb [~cronor@141.23.80.206] has joined #shogun | 10:15 | |
-!- flxb [~cronor@141.23.80.206] has left #shogun [] | 10:16 | |
-!- flxb [~cronor@fb.ml.tu-berlin.de] has joined #shogun | 10:16 | |
-!- blackburn [5bdfb203@gateway/web/freenode/ip.91.223.178.3] has joined #shogun | 10:21 | |
blackburn | oh good old sonne|work haven't seen you there for a while ;) | 10:22 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Ping timeout: 244 seconds] | 10:25 | |
sonne|work | heh | 10:26 |
blackburn | sonne|work: do you think development and master line separating is a good idea? I do | 10:28 |
-!- vikram360 [~vikram360@117.192.173.130] has quit [Ping timeout: 265 seconds] | 10:30 | |
sonne|work | I don't understand why it is a good idea... | 10:32 |
sonne|work | in my brain (at least currently) master is always the most up-to-date thing | 10:32 |
blackburn | sonne|work: something like debian stable as master and debian unstable as development? :) | 10:34 |
sonne|work | but where is master in all this? | 10:35 |
sonne|work | I mean what is master needed for then? | 10:35 |
sonne|work | we currently have branches for shogun 1.X etc | 10:35 |
sonne|work | but never use them (unfortunately) | 10:35 |
sonne|work | so we don't really fix bugs and release 1.1.1 / 1.1.2 ... | 10:36 |
blackburn | sonne|work: what is master for you? | 10:36 |
sonne|work | but always work on master | 10:36 |
sonne|work | master is development branch for me - the thing that is most recent | 10:36 |
blackburn | in this approach current master should go to developmen | 10:36 |
sonne|work | of course it makes sense to have branches for new complicated features | 10:37 |
blackburn | other way is to create stable branch | 10:37 |
sonne|work | like the (rotten) c5.0 branch | 10:37 |
blackburn | one example | 10:37 |
blackburn | we wanted to release | 10:37 |
blackburn | so we would separate branch from some march revision | 10:37 |
blackburn | before this new features | 10:37 |
blackburn | and merge fixes | 10:37 |
sonne|work | yeah I agree to that | 10:40 |
blackburn | in fact I think we are way too fast to have releases | 10:40 |
blackburn | :D | 10:40 |
sonne|work | I don't think so :D | 10:40 |
blackburn | I can hardly imagine somebody would want to install old shogun | 10:41 |
sonne|work | but that is the reality! | 10:42 |
sonne|work | people only install release versions! | 10:42 |
blackburn | sonne|work: what I do not like is to create branches for each fix | 10:42 |
blackburn | sonne|work: yes that'd right if we had .deb | 10:42 |
sonne|work | we should create automagically created debs / releases every night | 10:43 |
sonne|work | only when the test suite passes | 10:43 |
blackburn | it would be costly for me to get into .deb packaging ;) | 10:43 |
blackburn | will you have time to manage this? | 10:44 |
blackburn | however we have mighty gsomix! | 10:45 |
blackburn | :D | 10:45 |
blackburn | gsomix: would you like to set it up later? ;) | 10:45 |
blackburn | hey that's kind of good idea to add to your proposal | 10:46 |
blackburn | sonne|work: do you like it? | 10:46 |
blackburn | gsomix: do not afraid to be overloaded with it, it would be more interesting than hunt for covertree memleaks | 10:48 |
blackburn | gsomix: ok even two extensions of your proposal: auto deb (with ppa probably) and auto tests improvement | 10:53 |
-!- n4nd0 [~nando@n179-p53.kthopen.kth.se] has joined #shogun | 10:57 | |
gsomix | blackburn, ou. | 11:08 |
gsomix | blackburn, I'll read it later. But, I agree to everything. | 11:09 |
-!- PhilTillet [~Philippe@ASte-Genev-Bois-152-1-52-104.w82-121.abo.wanadoo.fr] has joined #shogun | 11:09 | |
blackburn | gsomix: and again start writing proposal, proposal would never be wrote by itself ;) | 11:10 |
gsomix | blackburn, ok. Once I get back from the gym. | 11:11 |
blackburn | are you used to visit gym? you never told me that before :) | 11:12 |
gsomix | blackburn, in university. | 11:12 |
blackburn | ah that kind of | 11:12 |
gsomix | Classes in physical culture. | 11:12 |
gsomix | Bjjj. | 11:12 |
blackburn | lolz | 11:13 |
blackburn | you have 17 minutes more, get the python4 support done! | 11:13 |
gsomix | More defines! | 11:14 |
gsomix | #defines, i mean | 11:14 |
blackburn | ah yes one more funny thing you can deal with | 11:14 |
blackburn | is to add common lisp typemaps | 11:14 |
blackburn | :D | 11:14 |
blackburn | yeah why not | 11:14 |
gsomix | haskell | 11:14 |
blackburn | IIRC there is no haskell support for swig | 11:14 |
blackburn | who would ever need this shit | 11:15 |
gsomix | There is part in swig documentation "Extending SWIG to support new languages". | 11:15 |
blackburn | nevertheless I can hardly imagine this stuff in haskell | 11:16 |
-!- stephenlee [da18b3c4@gateway/web/freenode/ip.218.24.179.196] has quit [Ping timeout: 245 seconds] | 11:17 | |
harshit_ | blackburn : good news :) every thing is working perfectly now in Newton SVM | 11:28 |
harshit_ | hurray :D | 11:28 |
harshit_ | now just gonna test it on toy dataset | 11:28 |
blackburn | harshit_: that's nice | 11:29 |
harshit_ | thanks for your support blackburn and n4nd0 | 11:29 |
blackburn | oscar award speech ;) | 11:30 |
harshit_ | lol | 11:30 |
harshit_ | cant wait for my first open source contribution to be accepted | 11:31 |
blackburn | is it your first PR? | 11:31 |
harshit_ | PR ? | 11:33 |
blackburn | pull request | 11:33 |
harshit_ | to an open source organization - yes. | 11:34 |
blackburn | I see | 11:34 |
harshit_ | but had some PR on private git hub repos | 11:34 |
blackburn | I just thought you commited some things before | 11:34 |
-!- harshit_ [~harshit@182.68.67.61] has quit [Ping timeout: 260 seconds] | 11:51 | |
-!- PhilTillet [~Philippe@ASte-Genev-Bois-152-1-52-104.w82-121.abo.wanadoo.fr] has quit [Quit: Leaving] | 12:11 | |
-!- PhilTillet [~Philippe@ASte-Genev-Bois-152-1-52-104.w82-121.abo.wanadoo.fr] has joined #shogun | 12:11 | |
-!- nickon [~noneedtok@dD5774105.access.telenet.be] has joined #shogun | 12:13 | |
-!- n4nd0 [~nando@n179-p53.kthopen.kth.se] has quit [Ping timeout: 246 seconds] | 12:14 | |
-!- nickon [~noneedtok@dD5774105.access.telenet.be] has quit [Client Quit] | 12:16 | |
-!- nickon [~noneedtok@dD5774105.access.telenet.be] has joined #shogun | 12:16 | |
-!- n4nd0 [~nando@n179-p53.kthopen.kth.se] has joined #shogun | 12:27 | |
-!- vikram360 [~vikram360@117.192.161.93] has joined #shogun | 12:30 | |
-!- PhilTillet [~Philippe@ASte-Genev-Bois-152-1-52-104.w82-121.abo.wanadoo.fr] has quit [Read error: Connection reset by peer] | 12:33 | |
n4nd0 | blackburn: regarding one of the comments in th PR | 12:38 |
blackburn | yes | 12:38 |
n4nd0 | blackburn: you mean that it should work with sparse features too? | 12:38 |
blackburn | why not? | 12:38 |
blackburn | all you need is distance so you don't have to check it | 12:38 |
n4nd0 | yes, it looks reasonable for me | 12:38 |
n4nd0 | blackburn: but we are just talking about simple or sparse features right? | 12:40 |
n4nd0 | blackburn: I am checking some of the other types of CDotFeatures though | 12:41 |
blackburn | n4nd0: let the distance check it :0 | 12:42 |
blackburn | :) | 12:42 |
n4nd0 | blackburn: do you mean that it shouldn't be checked here? | 12:43 |
blackburn | yes, distance should I think | 12:43 |
n4nd0 | in case it cannot be done, the distance will lock it when doing get_distance_matrix | 12:44 |
blackburn | it should throw SG_ERROR on init | 12:44 |
n4nd0 | in init yeah | 12:44 |
n4nd0 | I got to that part right now :D | 12:44 |
n4nd0 | blackburn: btw, I think I am not really familiarized with the idea of test suite and I probably should | 12:49 |
n4nd0 | blackburn: I understand that it's something we use to check the behaviour of the build right? | 12:50 |
n4nd0 | blackburn: that everything is working fine | 12:50 |
blackburn | n4nd0: it compares output with previous one | 12:50 |
n4nd0 | blackburn: ok | 12:51 |
n4nd0 | blackburn: we should always be sure that the previous one is ok then :P | 12:51 |
blackburn | if they are equal - life is good | 12:51 |
blackburn | yes | 12:51 |
n4nd0 | ok | 12:51 |
n4nd0 | this might make me look kind of stupid but ... | 12:52 |
n4nd0 | I think I have never run this testsuite | 12:52 |
blackburn | you don't have to | 12:52 |
n4nd0 | is it something we should ourselves? or done in the buildbot? | 12:52 |
blackburn | well we have to | 12:53 |
n4nd0 | so? | 12:54 |
n4nd0 | do I have to run it or not? | 12:54 |
blackburn | hah if you want to check something - yes | 12:54 |
blackburn | I mean in any case I run it sometimes and Soeren does | 12:55 |
n4nd0 | ok | 12:55 |
blackburn | try to run it anyway | 12:56 |
blackburn | just cd to testsuite/python_modular | 12:56 |
blackburn | and run tester.py | 12:56 |
n4nd0 | oh | 12:56 |
n4nd0 | should all of them be ok?? | 12:56 |
n4nd0 | maybe I'm missing a library or sth, I got quite a few of ERROR, some exceptions and it ended with a seg fault :O | 12:58 |
blackburn | they should but they are not ok :D | 12:58 |
n4nd0 | haha ok | 13:00 |
n4nd0 | is the seg fault normal?? | 13:00 |
n4nd0 | last file was | 13:00 |
n4nd0 | kernel_sparse_gaussian_modular.py setting 1/2 ERROR | 13:00 |
blackburn | it could be related to some serialization stuff we need to fix still | 13:00 |
n4nd0 | ok | 13:01 |
n4nd0 | I have also noted distance_mahalanobis_modular.py setting 1/1 NO TEST | 13:01 |
n4nd0 | I guess I should provide one | 13:01 |
n4nd0 | same for QDA | 13:01 |
blackburn | yes we need to generate tests for new ones | 13:02 |
blackburn | it is pretty straightforward btw | 13:02 |
blackburn | generator.py examplename.py | 13:03 |
n4nd0 | distance_mahalanobis_modular.py setting 1/1 OK | 13:07 |
n4nd0 | :D | 13:07 |
n4nd0 | blackburn: about the other comment at github | 13:14 |
n4nd0 | blackburn: the one related to the embed_distance idea | 13:14 |
blackburn | yes | 13:14 |
blackburn | just do as it was done in mds and isomap | 13:14 |
n4nd0 | blackburn: do you have time for it a moment now? | 13:14 |
blackburn | yes a little | 13:15 |
n4nd0 | blackburn: didn't notice that it's done like that in the others | 13:15 |
blackburn | I think apply should init distance with given features and delegate all the things to embed_distance | 13:15 |
blackburn | and embed_distance should do all the job | 13:16 |
n4nd0 | looks nice like that | 13:16 |
n4nd0 | I have to understand exactly what is the job of embed distance though | 13:16 |
n4nd0 | in one case it should do exactly as it does now | 13:16 |
blackburn | yes, the only difference is distance initialization | 13:18 |
n4nd0 | mmm I think I don't understand that clearly | 13:18 |
n4nd0 | initialization is simple m_distance->init(features, features) right? | 13:19 |
blackburn | yes | 13:19 |
blackburn | embed_distance makes sense if you have distances but do not have features | 13:20 |
n4nd0 | no ASSERT(features) then | 13:21 |
n4nd0 | ? | 13:21 |
blackburn | where? | 13:22 |
blackburn | def apply(features): | 13:22 |
blackburn | m_distance.init(features,features) | 13:22 |
blackburn | return embed_distance(m_distance) | 13:22 |
blackburn | def embed_distance(distance): | 13:23 |
blackburn | all the SPE stuff assuming distance is inited | 13:23 |
n4nd0 | aham, I see | 13:23 |
n4nd0 | but should sth else be added to handle the custom distance case you talked about? | 13:24 |
blackburn | yes if you want to precompute it prematurely | 13:24 |
blackburn | all you need is | 13:24 |
blackburn | def apply(features): | 13:24 |
gsomix | hi all | 13:24 |
blackburn | m_distance.init(features,features) | 13:24 |
blackburn | dist = CustomDistance(m_distance) | 13:25 |
blackburn | return embed_distance(dist) | 13:25 |
n4nd0 | then is it just required another apply to handle this case? | 13:26 |
blackburn | yes | 13:28 |
blackburn | just pass custom distance to embed_distance | 13:29 |
n4nd0 | blackburn: ok, I think I got it, thank you very much :) | 13:33 |
blackburn | welcome | 13:33 |
blackburn | ;) | 13:33 |
-!- vikram360 [~vikram360@117.192.161.93] has quit [Read error: Connection reset by peer] | 13:36 | |
-!- vikram360 [~vikram360@117.192.161.93] has joined #shogun | 13:36 | |
-!- av3ngr [av3ngr@nat/redhat/x-mtbcichayakoymjc] has quit [Quit: That's all folks!] | 13:54 | |
n4nd0 | blackburn: ups I got a problem with git when I tried to push the new stuff | 14:05 |
n4nd0 | error: failed to push some refs to 'git@github.com:iglesias/shogun.git' | 14:06 |
n4nd0 | To prevent you from losing history, non-fast-forward updates were rejected | 14:06 |
n4nd0 | Merge the remote changes (e.g. 'git pull') before pushing again. See the | 14:06 |
n4nd0 | 'Note about fast-forwards' section of 'git push --help' for details. | 14:06 |
n4nd0 | I think it is the same as the other day | 14:06 |
n4nd0 | I should just do git push --force?? | 14:06 |
n4nd0 | or should I pull as it suggests me | 14:06 |
blackburn | no better try to rebase your branch | 14:06 |
blackburn | or to merge | 14:07 |
n4nd0 | I did rebase before the push | 14:07 |
n4nd0 | I think that is why this is happening | 14:07 |
blackburn | yes | 14:10 |
blackburn | n4nd0: is it a branch? | 14:11 |
blackburn | n4nd0: your master is not up to date | 14:12 |
blackburn | try to push your master to github | 14:12 |
blackburn | and then push your branch once again | 14:12 |
blackburn | i.e. you rebased it locally but not on github | 14:12 |
blackburn | should work if I understand that correctly | 14:12 |
-!- wiking [~wiking@78-23-191-201.access.telenet.be] has joined #shogun | 14:13 | |
-!- wiking [~wiking@78-23-191-201.access.telenet.be] has quit [Changing host] | 14:13 | |
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun | 14:13 | |
n4nd0 | I needed to pull | 14:14 |
n4nd0 | it was in a branch yes | 14:14 |
n4nd0 | but my master is up to date, I did it a few moments ago | 14:15 |
blackburn | your local master or github one? | 14:15 |
blackburn | I checked your github fork and last commit was 3 days ago | 14:15 |
n4nd0 | ah fuck it was my local one | 14:16 |
wiking | blackburn: ey | 14:16 |
n4nd0 | maybe it was because of that | 14:16 |
n4nd0 | I did the pull and solve a couple of conflicts | 14:16 |
blackburn | wiking: hey | 14:16 |
blackburn | n4nd0: yes keep master up to date | 14:17 |
wiking | blackburn: ok so you haven't got time yet right? | 14:17 |
blackburn | both masters | 14:17 |
blackburn | wiking: hmm it seems I have actually | 14:17 |
wiking | since i haven't seen a reply by you | 14:17 |
blackburn | ahhh | 14:17 |
wiking | or maybe it slipped my eyes | 14:17 |
blackburn | I did not understand :D | 14:18 |
blackburn | wiking: did you get alex' idea | 14:19 |
blackburn | ? | 14:19 |
sonne|work | guys you should really discus on the ML | 14:19 |
wiking | noup | 14:19 |
blackburn | hah right | 14:19 |
blackburn | sonne|work: but we are talking about old mail | 14:20 |
blackburn | wiking: I'd suggest you to answer to ml this time | 14:20 |
blackburn | but I still do not understand the idea | 14:21 |
blackburn | :D | 14:21 |
sonne|work | well then invite him to this irc chat and ask him online :D | 14:21 |
blackburn | sonne|work: yeah I think that's the thing wiking should do ;) | 14:22 |
wiking | blackburn: he is on Skype rather than irc imho | 14:23 |
sonne|work | wiking: so what - send him the irc web url and grab him - tell him that I said that :) | 14:24 |
wiking | :>>>> | 14:24 |
wiking | ok | 14:24 |
n4nd0 | blackburn: I updated the PR | 14:27 |
n4nd0 | blackburn: I think I might have screwed though :S because of the merge | 14:28 |
n4nd0 | blackburn: can you take a quick look and tell if it is ok or if I should do sth? | 14:28 |
blackburn | yes | 14:28 |
n4nd0 | blackburn: thank you man | 14:28 |
sonne|work | wiking: btw did you submit your proposal? | 14:30 |
blackburn | n4nd0: allright | 14:30 |
blackburn | n4nd0: one more thing to go is to convert it to rather distance than distance matrix | 14:30 |
blackburn | this means you would have to store pointer to distance in lle point | 14:30 |
blackburn | err spe point | 14:30 |
blackburn | i.e. all distance_matrix[i*N+j] -> distance->distance(i,j) | 14:31 |
n4nd0 | aham I see | 14:31 |
wiking | sonne|work: not yet | 14:31 |
sonne|work | but you plan to right? | 14:32 |
wiking | sonne|work: deadline is on the 7th right | 14:32 |
wiking | yep | 14:32 |
sonne|work | 6th iirc | 14:32 |
sonne|work | but you might want to iterate with your potential mentor... | 14:32 |
blackburn | hurry up! | 14:32 |
blackburn | gsomix: and you as well | 14:32 |
sonne|work | (one can update until apr. 6) | 14:33 |
wiking | :>> | 14:33 |
wiking | i'll do something till tomorrow i think | 14:33 |
n4nd0 | blackburn: just in SPE_COVERTREE_POINT or also in embed_distance? the change to distance->distance(i, j) | 14:35 |
blackburn | n4nd0: everywhere probably | 14:38 |
blackburn | overhead of custom distance is rather slow | 14:38 |
blackburn | low | 14:38 |
blackburn | so it is almost the same | 14:38 |
n4nd0 | blackburn: then covertree has to be changed, instead of SGMatrix -> CDistance | 14:40 |
n4nd0 | CDistance* actually | 14:40 |
blackburn | yes exactly | 14:40 |
n4nd0 | I am with that right now then | 14:40 |
-!- gsomix [~gsomix@188.168.4.80] has quit [Ping timeout: 276 seconds] | 15:04 | |
-!- gsomix [~gsomix@83.234.169.79] has joined #shogun | 15:10 | |
-!- harshit_ [~harshit@182.68.67.61] has joined #shogun | 15:20 | |
flxb | Hello | 15:27 |
n4nd0 | flxb: hi there | 15:28 |
flxb | I have a weird error. I get Illegal instruction errors for some (!) shogun examples on some (!) nodes on our cluster. On other nodes everything works find. The nodes should have the same setup. The examples that fail have Illegal instruction in the next line in this listing: http://pastebin.com/mGWShRP1 Does anyone have any idea what this could be? | 15:29 |
n4nd0 | where does that Illegal instruction come from? | 15:32 |
n4nd0 | I mean, you redirect the output of the python scripts /dev/null right? | 15:33 |
flxb | yes | 15:33 |
flxb | but apparently illegal instruction errors are not redirected | 15:34 |
n4nd0 | mmm | 15:34 |
n4nd0 | I just executed the script and didn't get any of those Illegal instructions | 15:34 |
n4nd0 | let me check again | 15:34 |
n4nd0 | what do you get without /dev/null? | 15:35 |
n4nd0 | in particular for those that give illefal instruction | 15:36 |
n4nd0 | illegal* | 15:36 |
flxb | i get http://pastebin.com/mJFhEJam | 15:38 |
flxb | it must occur when shogun apply gets called | 15:39 |
n4nd0 | mmm | 15:42 |
n4nd0 | I think I need more hints because I cannot reproduce that illegal instruction in my machine | 15:42 |
n4nd0 | I can execute classifier_domainadaptationsvm_modular.py without problems | 15:42 |
blackburn | wtf is illegal instruction | 15:43 |
n4nd0 | I think it is something related to the processor | 15:45 |
flxb | n4nd0: it works on some nodes here too | 15:45 |
blackburn | flxb: could you please paste configure output somewhere? | 15:45 |
blackburn | it can be related to -march parameter of gcc | 15:45 |
harshit_ | n4nd0,blackburn : how can i make use of toy dataset in C++ , it doesn't seems to be in plain text | 15:46 |
blackburn | why not to use it from python? | 15:47 |
n4nd0 | harshit_: what kind of format does the file have? | 15:48 |
flxb | blackburn: configure output is here: http://pastebin.com/0wKEJjrW | 15:48 |
flxb | blackburn: maybe it's worth mentioning that i had to change swig to swig2.0 in configure file | 15:49 |
harshit_ | blackburn : my python is not very good, Actually I can test it in octave. | 15:49 |
blackburn | flxb: shouldn't cause it | 15:49 |
flxb | ok | 15:50 |
harshit_ | n4nd0: its *.dat and *.mat | 15:50 |
blackburn | flxb: actually you could try to compile with --disable-optimization | 15:50 |
n4nd0 | harshit_: binary files then? | 15:50 |
harshit_ | not really when i open DynProg in text editor it shows some weird symbols | 15:51 |
blackburn | flxb: it worths mentioning you have pretty old gcc btw | 15:51 |
n4nd0 | harshit_: then they are binary files ( != text files ) | 15:51 |
n4nd0 | harshit_: I don't think there is something in shogun to read directly .dat and .mat in C++ | 15:52 |
n4nd0 | harshit_: it's what we discussed last week | 15:52 |
harshit_ | yeah | 15:52 |
harshit_ | have to make use of octave then | 15:52 |
n4nd0 | harshit_: python works good for them too | 15:52 |
harshit_ | I remember, Also in 2011 ideas list there were some project related to this issue | 15:53 |
flxb | blackburn: should i try with Checking for GCC & CPU optimization abilities ... k8? | 15:53 |
harshit_ | project to enable shogun to support mat 7.0 files | 15:53 |
harshit_ | that seems interesting too | 15:53 |
blackburn | flxb: k8? it seems you have xeon, why k8? | 15:53 |
blackburn | I don't think we really need it - it is oneliner in python | 15:54 |
flxb | blackburn: i did --disable-optimization, but i am now on a different node. can this be the problem? | 15:54 |
blackburn | flxb: what do you mean? did you compile with disable opt? | 15:55 |
flxb | blackburn: no, i did not, i will try that now | 15:55 |
blackburn | I see | 15:55 |
flxb | but previously i compiled on another node | 15:56 |
blackburn | flxb: are they identical? | 15:56 |
blackburn | in case of optimizations enabled that could cause some troubles | 15:56 |
flxb | blackburn: yes i guess that is the problem | 15:56 |
flxb | one is Quad-Core AMD Opteron(tm) Processor 2378, the other one is Intel(R) Xeon(R) CPU | 15:57 |
blackburn | ah | 15:57 |
blackburn | yes could be a problem for sure | 15:57 |
blackburn | then just compile on each machine separately with optimizations enabled | 15:57 |
blackburn | should work | 15:57 |
blackburn | it is some mtune or march or other hardware-related option issue | 15:58 |
flxb | how much performance drawbacks do i have if i just disable optimization? | 15:58 |
blackburn | depends what are you doing | 15:59 |
blackburn | however I haven't seen any drawback larger than 1.5x | 15:59 |
-!- Marty28 [~Marty@158.181.76.57] has quit [Quit: ChatZilla 0.9.88.1 [Firefox 11.0/20120310010446]] | 16:00 | |
flxb | blackburn: ok. everything works now! | 16:01 |
flxb | thanks! | 16:01 |
n4nd0 | cool | 16:01 |
blackburn | whoa I don't like to work 8 hours straight :( | 16:10 |
n4nd0 | blackburn: lot of things to do at the job? | 16:12 |
blackburn | n4nd0: no, but sitting for 8 hours.. | 16:18 |
blackburn | I mean it bothers to not change activities/place :) | 16:20 |
n4nd0 | aham | 16:23 |
sonne|work | harshit_: totally fine if you come up with an octave modular example | 16:27 |
harshit_ | sonne|work : done with it ! | 16:27 |
harshit_ | test it on fm_train_real.dat | 16:27 |
sonne|work | how does it compare speed wise? | 16:27 |
harshit_ | same result on matlab | 16:27 |
sonne|work | well you can load any (bigger) data set | 16:27 |
harshit_ | which one ? dna ? | 16:28 |
sonne|work | or mnist ... | 16:28 |
sonne|work | IIRC someone has .mat files for that one around too | 16:28 |
harshit_ | ohk,but please tell me how to compare speeds | 16:29 |
sonne|work | tic; toc; in matlab :) | 16:29 |
sonne|work | or octave | 16:29 |
harshit_ | okay .. thanks | 16:30 |
sonne|work | so just load the data tic; call newton svm matlab script ; toc | 16:30 |
sonne|work | and then do the same under octave with your new shogun newton svm impl | 16:30 |
n4nd0 | sonne|work: I just submitted the application for multiclass | 16:30 |
sonne|work | seen it | 16:31 |
n4nd0 | sonne|work: could you take it a look when you get some time and tell your opinion / things to improve? | 16:31 |
sonne|work | ok | 16:31 |
n4nd0 | sonne|work: thank you :) | 16:31 |
n4nd0 | sonne|work: it is quite based on SO's one so you can save some reading ;) | 16:32 |
n4nd0 | time for me to take a rest | 16:37 |
n4nd0 | see you later guys | 16:37 |
blackburn | yeah you are too | 16:37 |
blackburn | ehm don't know right word | 16:37 |
blackburn | nervous? ;) | 16:38 |
n4nd0 | mmm no | 16:38 |
n4nd0 | why? | 16:38 |
n4nd0 | should I :P?? | 16:38 |
blackburn | no, I mean you spend too much time there | 16:38 |
n4nd0 | ah yes | 16:38 |
n4nd0 | see you later then | 16:39 |
blackburn | see you | 16:39 |
harshit_ | sonne|work : time = 8.2211e-04, for training on fm database | 16:39 |
n4nd0 | blackburn: try to stand up and stretch your legs for a while :) | 16:39 |
harshit_ | now testing it on 20 newsgroup | 16:39 |
harshit_ | will that work | 16:39 |
-!- n4nd0 [~nando@n179-p53.kthopen.kth.se] has quit [Quit: leaving] | 16:39 | |
sonne|work | harshit_: how big is the data set? | 16:40 |
sonne|work | how many examples / dims? | 16:40 |
blackburn | n4ndo yeah I do | 16:40 |
sonne|work | and please try matlab too :) | 16:40 |
sonne|work | so we can compare | 16:40 |
harshit_ | not really big | 16:41 |
harshit_ | 2*92 | 16:41 |
harshit_ | now running on 20Newgroup which is very large | 16:41 |
sonne|work | harshit_: how large? | 16:42 |
harshit_ | about 18000 examples and 6000 features in 20NEwsgroup | 16:42 |
sonne|work | ok - is this data set sparse? | 16:42 |
sonne|work | if so please don't forget to convert it to dense | 16:42 |
sonne|work | in matlab the cmd is | 16:43 |
sonne|work | double(X) | 16:43 |
sonne|work | so we can compare | 16:43 |
sonne|work | octave modular has no sparse feature matrix support yet (one of gsomix's gsoc tasks :) | 16:43 |
harshit_ | yeah that is sparse : | 16:44 |
blackburn | sonne|work: one of 100 | 16:44 |
sonne|work | 100000 :D | 16:44 |
sonne|work | gtg | 16:44 |
sonne|work | cu | 16:44 |
harshit_ | bye | 16:44 |
harshit_ | blackburn : having problems with 20Newsgroup on my problem | 16:52 |
harshit_ | its a multiclass problem | 16:52 |
harshit_ | which big dataset do you normally use | 16:53 |
blackburn | harshit_: http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary.html | 16:57 |
harshit_ | thanks, :) | 16:58 |
-!- harshit_ [~harshit@182.68.67.61] has quit [Ping timeout: 245 seconds] | 17:02 | |
-!- genix [~gsomix@188.168.5.165] has joined #shogun | 17:09 | |
@sonney2k | blackburn, well he could have just mergerd some labels... | 17:09 |
blackburn | sonney2k: Im not sure what is you are talknig about? | 17:10 |
@sonney2k | multiclass -> binary | 17:10 |
blackburn | ah yes sure | 17:10 |
@sonney2k | anyway mldata.org is probably even better for harshit - data already is in matlab format | 17:11 |
-!- gsomix [~gsomix@83.234.169.79] has quit [Read error: Operation timed out] | 17:11 | |
-!- blackburn [5bdfb203@gateway/web/freenode/ip.91.223.178.3] has quit [Quit: Page closed] | 17:32 | |
-!- PhilTillet [~Philippe@tillet-p42154.maisel.int-evry.fr] has joined #shogun | 18:19 | |
PhilTillet | Hello | 18:19 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 18:35 | |
-!- blackburn [~qdrgsm@83.234.54.186] has joined #shogun | 18:43 | |
CIA-64 | shogun: Soeren Sonnenburg master * rce6628c / examples/undocumented/python_modular/classifier_multiclasslinearmachine_modular.py : rename function to match file name - http://git.io/6uWtmg | 18:57 |
CIA-64 | shogun: Soeren Sonnenburg master * ra28591d / testsuite/python_modular/tester.py : remove pdb import from tester - http://git.io/Qamxaw | 18:57 |
n4nd0 | sonney2k: aham! so it was that pdb what changed before in the tester | 19:09 |
n4nd0 | I executed it once and was smooth and in the second I got this gdb type interface | 19:09 |
n4nd0 | I was like ... what did I do? | 19:10 |
n4nd0 | :D | 19:10 |
@sonney2k | n4nd0, yeah I screwed up debugging the tests | 19:17 |
@sonney2k | n4nd0, I went through most of the tests and checked where they differ | 19:18 |
@sonney2k | and most were ok but hey I screwed up again and forgot to upload the updated test files | 19:18 |
@sonney2k | which are now gone - bah! | 19:18 |
-!- siddharth [~siddharth@14.139.82.6] has joined #shogun | 19:21 | |
@sonney2k | n4nd0, I am now trying to figure out what this QDA error is about | 19:22 |
@sonney2k | QDA.cs(143,18): error CS0136: A local variable named `i' cannot be declared in this scope because it would give a different meaning to `i', which is already used in a `parent or current' scope to denote something else | 19:22 |
n4nd0 | what? I have never seen that | 19:25 |
n4nd0 | but let's check | 19:25 |
@sonney2k | n4nd0, look at the buildbot | 19:25 |
@sonney2k | I renamed i -> c and it will likely compile | 19:25 |
-!- siddharth [~siddharth@14.139.82.6] has left #shogun ["Leaving"] | 19:26 | |
PhilTillet | sonney2k, is the kernel matrix stored columnwise or rowwise? | 19:28 |
CIA-64 | shogun: Soeren Sonnenburg master * r54e89fd / src/shogun/classifier/QDA.h : rename index i to c to fix clash with csharp typemap - http://git.io/urYyOQ | 19:29 |
n4nd0 | sonney2k: were are those files like QDA.cs? | 19:29 |
@sonney2k | n4nd0, yes all good - lets hope the buildbot is happy now | 19:29 |
@sonney2k | shogun-buildbot, be happy! | 19:29 |
shogun-buildbot | What you say! | 19:29 |
n4nd0 | sonney2k: let's hope so | 19:29 |
@sonney2k | PhilTillet, everything in shogun is columnwise | 19:30 |
n4nd0 | sonney2k: ah ok, one makes the changes in the C++ source | 19:30 |
@sonney2k | PhilTillet, but kernel matrix is usually not computed as one big thing (won't fit in mem...) | 19:30 |
PhilTillet | well yes but for GPUs ... | 19:30 |
PhilTillet | :p | 19:30 |
PhilTillet | for a first prototype I mean | 19:31 |
@sonney2k | n4nd0, yes and swig generates the csharp python etc bindings | 19:31 |
-!- genix [~gsomix@188.168.5.165] has quit [Read error: Operation timed out] | 19:31 | |
@sonney2k | PhilTillet, well it is really useless to do it for kernel matrices that fit in memory | 19:31 |
PhilTillet | hmmm | 19:32 |
PhilTillet | probably | 19:32 |
@sonney2k | maybe you would rather want to speed up the compute() function? not sure what you are doing right now | 19:32 |
PhilTillet | I am trying to make some opencl implementation of my dirty code | 19:33 |
n4nd0 | sonney2k: are the bindings normal source files? I got surprised looking at the buildbot and seeing that it identifies a QDA.cs, for example | 19:33 |
PhilTillet | like in the shogun CGaussianKerel and CKernelMachine class | 19:33 |
PhilTillet | so that just have to replace svm->apply() with svm->ocl_apply() | 19:34 |
PhilTillet | it internally copies everything to GPU mem (for now with the assumption that it will fit :D) | 19:34 |
n4nd0 | sonney2k: I thought that we could use the code we do in C++ as libraries from the other languages, but that there was no generation with SWIG | 19:34 |
@sonney2k | PhilTillet, what does it copy ? the kernel matrix? | 19:35 |
@sonney2k | n4nd0, exactly | 19:35 |
PhilTillet | that was my incoming question | 19:35 |
PhilTillet | should I make another temporary matrix? | 19:35 |
PhilTillet | or should I make a gpu_kernel_matrix ? | 19:35 |
PhilTillet | like an attribute | 19:35 |
@sonney2k | if you precompute the kernel matrix - all the rest will take basically 0 time | 19:35 |
n4nd0 | sonney2k: I am reading a bit about SWIG right now to understand a bit more, thank you! | 19:35 |
PhilTillet | I know | 19:35 |
PhilTillet | the rest is just roughly a matrix vector product | 19:36 |
@sonney2k | yes | 19:36 |
PhilTillet | but the kernel matrix is computed at the "apply()" point right? | 19:37 |
PhilTillet | hmm I see | 19:38 |
PhilTillet | the kernel matrix is not necessarily features*support_vectors | 19:38 |
PhilTillet | sonney2k, did not read carefully enough your question, it copies features to gpu, and caches support vectors | 19:43 |
PhilTillet | then compute labels on gpu | 19:43 |
PhilTillet | and copies them back to cpu | 19:43 |
@sonney2k | PhilTillet, ahh so you copy just the support vectors? | 19:46 |
PhilTillet | the features too | 19:46 |
@sonney2k | which features? | 19:46 |
PhilTillet | the examples | 19:46 |
@sonney2k | which examples :) | 19:46 |
PhilTillet | wait I'm getting confused | 19:46 |
PhilTillet | XD | 19:46 |
@sonney2k | the one to compute the output for? | 19:46 |
PhilTillet | yes | 19:46 |
@sonney2k | s/one/ones | 19:47 |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun | 19:47 | |
@sonney2k | (support vectors are also examples - a subset of the training examples) | 19:47 |
PhilTillet | oh yes right | 19:47 |
@sonney2k | so test examples you mean probably | 19:47 |
PhilTillet | yes, I copy the support vectors and test examples to gpu | 19:47 |
PhilTillet | compute test labels on gpu | 19:47 |
PhilTillet | copies back to cpu | 19:47 |
PhilTillet | (test labels) | 19:48 |
@sonney2k | what would be more reasonable is to copy only SVs to gpu | 19:48 |
-!- harshit_ [~harshit@182.68.67.61] has joined #shogun | 19:48 | |
@sonney2k | and then test example by test example to GPU mem and compute output | 19:48 |
PhilTillet | I really do not think so | 19:48 |
PhilTillet | :p | 19:48 |
PhilTillet | would get way less gigaflops | 19:48 |
@sonney2k | why? | 19:49 |
@sonney2k | I would bet that it does not make any difference | 19:49 |
PhilTillet | hmmm | 19:49 |
PhilTillet | It does from a cahcing point of view | 19:49 |
@sonney2k | why? | 19:50 |
PhilTillet | I mean, it's the difference between a matrix matrix product and a matrix vector product | 19:50 |
PhilTillet | when running on GPU | 19:50 |
PhilTillet | each work group caches chunk of the matrix | 19:50 |
PhilTillet | i don't really know how to explain | 19:51 |
PhilTillet | well there are multiple problems, but this is the first one | 19:51 |
PhilTillet | the work groups on gpu will cache chunks of the feature matrix | 19:52 |
n4nd0 | sonney2k: it looks like it failed again | 19:53 |
-!- blackburn [~qdrgsm@83.234.54.186] has quit [Ping timeout: 246 seconds] | 19:53 | |
n4nd0 | sonney2k: same error, same place | 19:53 |
@sonney2k | PhilTillet, ok yes matrix1 * matrix2 is faster but neither matrix1 nor matrix2 fit in GPU memory | 19:55 |
PhilTillet | yes, this is why there has to be some complicated tricks to do | 19:56 |
PhilTillet | to do submatrix1*submatrix2 | 19:56 |
@sonney2k | you could even imagine test features to be streamed from disk | 19:56 |
PhilTillet | shouldn't it be possible to stream it into a buffer, and when the buffer has a certain size, transfer on gpu, compute, etc | 19:56 |
@sonney2k | yes exactly | 19:57 |
PhilTillet | I think if we compute example by example, performance would be even worse on GPU than on CPU | 19:57 |
PhilTillet | plus, for each GPU operation (memory transfer, opencl kernel initialization, etc...) | 19:57 |
PhilTillet | there is about a 50microsec time | 19:57 |
@sonney2k | but I would really focus on only doing submatrix1 stuff and use vectors on the right hand side | 19:57 |
PhilTillet | so if you get a 50microsec overhead for each example | 19:58 |
PhilTillet | but the main issues is the gigaflop one :p | 19:58 |
PhilTillet | hmmm | 19:58 |
PhilTillet | right hand side is the test examples? | 19:59 |
@sonney2k | if the overhead is really that huge then it is becoming tough | 19:59 |
@sonney2k | yes | 19:59 |
PhilTillet | I mean, why not accumulating the test examples into a sufficiently small buffer? | 20:00 |
@sonney2k | brb | 20:00 |
PhilTillet | ok :) | 20:00 |
n4nd0 | sonney2k: QDA.cpp:87, I used int instead of int32_t there, do you think it can be related? | 20:01 |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds] | 20:01 | |
-!- blackburn [~qdrgsm@83.234.54.186] has joined #shogun | 20:08 | |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun | 20:12 | |
-!- gsomix [~gsomix@188.168.5.227] has joined #shogun | 20:27 | |
@sonney2k | n4nd0, be patient ... the buildbot is still working | 20:29 |
blackburn | gsomix: here, wanted to catch me? ;) | 20:31 |
blackburn | sonney2k: what's up? | 20:31 |
blackburn | strange issue with 'i' | 20:32 |
@sonney2k | no | 20:32 |
@sonney2k | not strange and problem resolved | 20:33 |
blackburn | sonney2k: that mail from jacob - I can hardly come with any answer, can you? | 20:34 |
@sonney2k | blackburn, no - I am not a GP expert | 20:34 |
@sonney2k | olivier has to answer to this | 20:34 |
blackburn | yes but he listed oliver's ideas mostly :) | 20:35 |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Remote host closed the connection] | 20:35 | |
@sonney2k | no idea | 20:36 |
@sonney2k | blackburn, did the buildbot get any break recently? | 20:39 |
@sonney2k | seems to me it is hardly diling | 20:39 |
@sonney2k | idling | 20:39 |
blackburn | sonney2k: what kind of break? | 20:39 |
blackburn | ah | 20:39 |
blackburn | yeah strange | 20:40 |
blackburn | a few small commits probably | 20:40 |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun | 20:41 | |
CIA-64 | shogun: Sergey Lisitsyn master * rf1564b9 / (3 files in 3 dirs): Warnings removal - http://git.io/w639Ng | 20:44 |
blackburn | sonney2k: have you seen git network? | 20:44 |
blackburn | of shogun | 20:44 |
blackburn | https://github.com/shogun-toolbox/shogun/network kind of subway | 20:45 |
blackburn | just like last year :) | 20:45 |
blackburn | sonney2k: btw have you got stats of shogun-toolbox.org? should be interesting | 20:46 |
@sonney2k | blackburn, would be more impressive if we both did pull requests too | 20:47 |
blackburn | sonney2k: I do for big things like edrt | 20:47 |
blackburn | but hey new branch to remove warnings.. | 20:47 |
blackburn | whoaa http://www.youtube.com/watch?feature=player_detailpage&v=K9EICyaEJq0 | 20:49 |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Remote host closed the connection] | 20:49 | |
harshit_ | blackburn :having a little error when i use lard dataset in octave : No matching function for overload | 20:49 |
blackburn | which function? | 20:50 |
harshit_ | and error points to Label() | 20:50 |
blackburn | Label? | 20:50 |
harshit_ | on line : labels=Labels(label_train_twoclass); | 20:50 |
harshit_ | where label_train_twoclass is a vector | 20:51 |
harshit_ | having labels | 20:51 |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun | 20:52 | |
@sonney2k | harshit_, what type is that? | 20:52 |
@sonney2k | harshit_, does Labels([1.0, 2.0, 3.0]) work? | 20:52 |
harshit_ | I dont declare any type as such ,labels is returned from libsvmread() function | 20:53 |
harshit_ | wait i'll check | 20:53 |
harshit_ | No Labels([1.0,2.0,3.0]) doesn't give any error | 20:54 |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds] | 20:58 | |
blackburn | http://cs5422.userapi.com/u464017/-14/x_ac4cb519.jpg skyscraper is on fire in moscow | 20:59 |
harshit_ | Done ! figured out the problem | 20:59 |
n4nd0 | blackburn: too warm around there? | 20:59 |
blackburn | n4nd0: I am fortunately 1000 km far away haha | 21:00 |
n4nd0 | blackburn: good | 21:00 |
n4nd0 | blackburn: I hope it is not a big disaster | 21:01 |
blackburn | today was a plane crash in other city | 21:01 |
blackburn | that was a disaster actualyl | 21:01 |
blackburn | 38 died | 21:01 |
blackburn | IIRC | 21:01 |
@sonney2k | harshit_, so what was the problem? | 21:01 |
harshit_ | sonney2k: In octave , newton SVM is taking 0.091 | 21:02 |
harshit_ | i just transposed labels and matrix and it worked ! | 21:02 |
harshit_ | 0.091 sec to train 60*1000 dataset | 21:03 |
harshit_ | dataset I used is called splice dataset | 21:04 |
@sonney2k | harshit_, can't you use a bigger data set? | 21:06 |
harshit_ | couldn't find any one which doesn't requires preprocessing | 21:06 |
@sonney2k | harshit_, http://mldata.org/repository/data/viewslug/covtypebinary_scale/ | 21:07 |
@sonney2k | you can even just download a .mat file | 21:07 |
@sonney2k | no work needed... | 21:07 |
harshit_ | great .. wait i'll run on it | 21:08 |
harshit_ | dam thats a really big one ! | 21:08 |
n4nd0 | sonney2k: do you have any other task in mind / something you would like to see done in shogun? | 21:10 |
@sonney2k | n4nd0, want to learn a bit more about swig? | 21:11 |
harshit_ | sonney2k : I saw S3VM somewhere in liblinear, so why in homepage of shogun semisupervised algos is crossed | 21:11 |
harshit_ | at place where all toolboxes are compared | 21:12 |
@sonney2k | n4nd0, or lets better ask which topic is of interest to you? | 21:12 |
@sonney2k | harshit_, liblinear has s3svm? | 21:12 |
n4nd0 | sonney2k: I am a bit open minded in that sense, I even prefer to learn some new stuff | 21:13 |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun | 21:13 | |
harshit_ | I dont remember exactly which svm library it was but somewhere i saw an enum where S3VM was an option for classifier type | 21:13 |
n4nd0 | sonney2k: I wondered before if you had something on shell scripting to be done | 21:14 |
harshit_ | sonney2k : Just wondering if you want to see co-clustering by blum in shogun ! | 21:14 |
n4nd0 | sonney2k: among the things I have done this far, I like the most QDA | 21:15 |
n4nd0 | sonney2k: other classifier (if there's something not around here yet!) could be good then | 21:15 |
@sonney2k | n4nd0, some 'easy' decision tree then? | 21:16 |
blackburn | I have my own id3 python prototype actually | 21:17 |
n4nd0 | sonney2k: sure, I don't know about decision trees that much | 21:17 |
n4nd0 | it will be good to learn about them | 21:17 |
blackburn | parzen window classifier | 21:17 |
@sonney2k | maybe even start with decision stumps | 21:17 |
@sonney2k | blackburn, true | 21:17 |
n4nd0 | haha you guys definetely have lot of ideas | 21:18 |
@sonney2k | e.g. https://en.wikipedia.org/wiki/ID3_algorithm | 21:19 |
blackburn | not a lot, only a few (about one thousand) | 21:19 |
@sonney2k | or rbf networks https://en.wikipedia.org/wiki/Radial_basis_function_network | 21:20 |
blackburn | ARMA model haha | 21:20 |
blackburn | ransac | 21:21 |
@sonney2k | blackburn, btw does KNN use covertree now? | 21:22 |
blackburn | no | 21:22 |
blackburn | I suggested to do that a few times :) | 21:22 |
n4nd0 | blackburn: oh yes, that's true, you told me about that | 21:24 |
blackburn | n4nd0: now you can do that I think | 21:24 |
n4nd0 | blackburn: ok, I will take a look to covertree in KNN then | 21:24 |
blackburn | (after struggles with spe) | 21:24 |
n4nd0 | sonney2k: is that ok? | 21:25 |
n4nd0 | blackburn: sure ;) | 21:25 |
@sonney2k | yup | 21:25 |
n4nd0 | ok | 21:25 |
blackburn | n4nd0: but will you finish spe? | 21:25 |
blackburn | :) | 21:25 |
n4nd0 | blackburn: yeah! | 21:25 |
blackburn | ok | 21:25 |
n4nd0 | blackburn: I said sure :P | 21:25 |
blackburn | I believe we need to add more cats to shogun | 21:27 |
@sonney2k | harshit_, so what are the timings? | 21:27 |
blackburn | we have no graphical examples with cats | 21:27 |
@sonney2k | blackburn, well we have one cat - your gf :) | 21:28 |
blackburn | sonney2k: http://www.persian-catsclub.com/images/persian-cat.jpg looks like shogun mascot | 21:28 |
@sonney2k | yeah | 21:28 |
blackburn | sonney2k: I pasted you msg to gf :D | 21:29 |
@sonney2k | kind of a yakuza mafia shogun cat | 21:29 |
blackburn | yeah | 21:29 |
blackburn | very dangerous | 21:29 |
harshit_ | sonney2k: strange error ! nothing comes after first iteration of newtonSVM | 21:30 |
harshit_ | But everything was working fine for other datasets | 21:31 |
@sonney2k | harshit_, even w/ matlab? | 21:32 |
harshit_ | I am working in octave for now | 21:32 |
blackburn | sonney2k: is scatter mc svm a good idea still? | 21:32 |
@sonney2k | not so much | 21:32 |
blackburn | ah, n4nd0 - nearest centroid classifier! | 21:33 |
n4nd0 | blackburn: haha another idea!? | 21:33 |
blackburn | and one with median | 21:33 |
blackburn | yes | 21:33 |
blackburn | two even | 21:33 |
blackburn | I forgot how do they call it | 21:33 |
harshit_ | Could that be bcoz of less memory available to octave? | 21:33 |
@sonney2k | harshit_, well you could use 1/2 of the data if the data set is too big | 21:34 |
harshit_ | sonney2k : its running now with full dataset, Problem was that I had a lot of things opened earlier | 21:37 |
PhilTillet | hi hi | 21:40 |
CIA-64 | shogun: Soeren Sonnenburg master * refd2da9 / testsuite/python_modular/tester.py : add try catch around len() - http://git.io/8yuhSA | 21:40 |
n4nd0 | sonney2k: what about the QDA.cs issue? I have checked and think that the error is still there | 21:43 |
n4nd0 | sonney2k: but maybe I didn't check it correctly | 21:43 |
@sonney2k | n4nd0, it compiled locally here... | 21:44 |
@sonney2k | PhilTillet, that's a linear method - hard to imagine how GPUs can speed up anything for this | 21:46 |
n4nd0 | I have no idea to check the buildbot then :S | 21:46 |
@sonney2k | n4nd0, it is overly busy... | 21:46 |
PhilTillet | sonney2k, what are you talking about? :p | 21:46 |
@sonney2k | heh | 21:46 |
blackburn | sonney2k: was it an answer for 'hi hi'? | 21:46 |
blackburn | I can hardly imagine what you would answer for ho ho | 21:47 |
PhilTillet | lol | 21:47 |
@sonney2k | n4nd0, lets see if everything is OK tomorrow | 21:47 |
blackburn | kernel laplace transformation can not map into non-euclidean space? | 21:47 |
blackburn | :D | 21:47 |
@sonney2k | blackburn, did I say that I love the shogun killer cat(tm)? | 21:47 |
@sonney2k | anyway bed time for me | 21:48 |
@sonney2k | cu all | 21:48 |
PhilTillet | cu | 21:48 |
n4nd0 | good night | 21:48 |
blackburn | sonney2k: yeah cats are cool :) nite | 21:48 |
-!- harshit_ [~harshit@182.68.67.61] has quit [Read error: Connection reset by peer] | 21:49 | |
-!- PSmitAalto [82e9b263@gateway/web/freenode/ip.130.233.178.99] has quit [Ping timeout: 245 seconds] | 21:52 | |
-!- harshit_ [~harshit@182.68.67.61] has joined #shogun | 21:57 | |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds] | 22:03 | |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun | 22:16 | |
-!- harshit_ [~harshit@182.68.67.61] has quit [Read error: Connection reset by peer] | 22:16 | |
shogun-buildbot | build #196 of nightly_all is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_all/builds/196 | 22:17 |
blackburn | n4nd0: gsomix vodka! ^ | 22:17 |
n4nd0 | awesome!! | 22:17 |
n4nd0 | C# should work soon then | 22:18 |
n4nd0 | btw | 22:19 |
n4nd0 | what does nightly stands for / mean? | 22:19 |
n4nd0 | I guess it must be a name used to refer to a special release or sth like that | 22:19 |
blackburn | n4nd0: it build on nights ;) | 22:22 |
blackburn | is builded* | 22:22 |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Remote host closed the connection] | 22:22 | |
blackburn | firefox also has nightly builds and a lot of other projects too | 22:22 |
n4nd0 | aham, I see | 22:23 |
n4nd0 | so it was the obvious answer :D | 22:23 |
-!- wiking_ [~wiking@huwico/staff/wiking] has joined #shogun | 22:23 | |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun | 22:26 | |
-!- naywhaya1e [~ryan@spoon.lugatgt.org] has joined #shogun | 22:27 | |
-!- wiking [~wiking@huwico/staff/wiking] has quit [Ping timeout: 260 seconds] | 22:30 | |
-!- naywhayare [~ryan@spoon.lugatgt.org] has quit [Ping timeout: 260 seconds] | 22:30 | |
-!- wiking_ is now known as wiking | 22:31 | |
shogun-buildbot | build #435 of csharp_modular is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/csharp_modular/builds/435 | 22:31 |
-!- harshit_ [~harshit@182.68.67.61] has joined #shogun | 22:36 | |
harshit_ | sonney2k: done ! | 22:37 |
harshit_ | it was not running because of the value of C I set was too low | 22:37 |
harshit_ | now it has took 7.892sec for whole dataset | 22:37 |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds] | 22:42 | |
blackburn | see you guys | 22:48 |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun | 22:54 | |
-!- blackburn [~qdrgsm@83.234.54.186] has quit [Ping timeout: 252 seconds] | 22:56 | |
-!- vikram360 [~vikram360@117.192.161.93] has quit [Read error: Connection reset by peer] | 23:05 | |
n4nd0 | good night | 23:13 |
gsomix | ta-dam | 23:20 |
gsomix | i finished repair of room. | 23:21 |
-!- flxb [~cronor@fb.ml.tu-berlin.de] has quit [Quit: flxb] | 23:27 | |
-!- PhilTillet [~Philippe@tillet-p42154.maisel.int-evry.fr] has quit [Remote host closed the connection] | 23:31 | |
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking] | 23:46 | |
-!- tibi_popa [tibi_popa@95.76.45.235] has joined #shogun | 23:54 | |
--- Log closed Tue Apr 03 00:00:19 2012 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!