Open in new window / Try shogun cloud
--- Log opened Mon Jul 22 00:00:27 2013
--- Day changed Mon Jul 22 2013
@sonney2klisitsyn, already having the option not to init memory is very helpful00:00
lisitsynsonney2k: if you realloc too much it may be a wrong data structure00:00
@sonney2kso in a loop where you fill an array anyway00:00
lisitsynsonney2k: reserve doesn't init memory00:00
lisitsynresize does reserve doesn't00:01
@sonney2klisitsyn, sure but in cases you don't know how big the thing will be00:03
lisitsynsonney2k: that's for linked list right?00:03
@sonney2klisitsyn, for e.g a tree00:03
lisitsynwhy to store it in a vector?00:04
@sonney2kto save memory00:05
@sonney2kyou only have one big memory block and indices into the nodes00:05
lisitsynsonney2k: I wouldn't use vector for that then!00:05
@sonney2klisitsyn, what else?00:06
lisitsynsonney2k: it is like when you write your own data structure00:07
lisitsynso you can put some manually controlled memory00:07
thoralfsonney2k: examples from libshogun are working, but cmdline_static has 5 segfaulting tests...00:14
thoralf#0  shogun::CGUIPreprocessor::add_preproc (this=0x620620, preproc=0x62a2f0) at ui/GUIPreprocessor.cpp:12200:16
thoralf#1  0x00007ffff792c03f in shogun::CSGInterface::cmd_add_preproc (this=0x613e00) at ui/SGInterface.cpp:526300:16
thoralf#2  0x00007ffff792ecc2 in shogun::CSGInterface::handle (this=0x613e00) at ui/SGInterface.cpp:796500:16
thoralf#3  0x0000000000408af8 in main (argc=<value optimized out>, argv=0x7fffffffe0f8) at CmdLineInterface.cpp:75300:16
@iglesiasglisitsyn, sonney2k : PR alarm!00:16
thoralfin distance_hammingword.sg00:16
@sonney2kthoralf, that sounds like a change of Heiko00:19
@sonney2kI didn't do anything there00:19
thoralfsonney2k: I'm putting away my old laptop now and get some sleep.00:20
thoralfsonney2k: If you need more help, you need to wait until tomorrow after work.00:21
@sonney2kthoralf, no it is all good! thanks!00:21
@sonney2klisitsyn, whenever you see Heiko/wiking tell them we need tests back working or we will run into a nightmare mess00:22
* sonney2k is finally entering the nips reviews00:23
thoralfGuys, see you tomorrow.00:23
thoralfSorry, today. ;)00:23
@iglesiasgthoralf: good night!00:23
-!- thoralf [] has quit [Quit: Konversation terminated!]00:24
@sonney2kohh man why do I only get crappy papers for review01:05
lisitsynsonney2k: again?01:08
@iglesiasgsonney2k: what are they about?01:10
@sonney2kcan't say svm, mkl this kind of stuff01:19
@sonney2kall submitted just 9 days late01:36
@sonney2kalrighty sleep!01:37
@iglesiasggood night!01:38
-!- FSCV [] has quit [Ping timeout: 268 seconds]01:46
-!- FSCV [~FSCV@] has joined #shogun01:49
-!- nube [~rho@] has quit [Ping timeout: 264 seconds]01:51
-!- iglesiasg [~Fernando@] has quit [Quit: Leaving]01:54
-!- nube [~rho@] has joined #shogun02:04
-!- FSCV_ [~FSCV@] has joined #shogun02:14
-!- FSCV [~FSCV@] has quit [Ping timeout: 248 seconds]02:15
-!- FSCV_ [~FSCV@] has quit [Quit: Leaving]02:21
-!- shogun-notifier- [] has quit [Quit: transmission timeout]02:33
-!- nube [~rho@] has quit [Quit: Leaving.]02:47
shogun-buildbotbuild #400 of nightly_all is complete: Failure [failed compile]  Build details are at
shogun-buildbotbuild #465 of nightly_default is complete: Failure [failed test]  Build details are at
-!- gsomix [~gsomix@] has quit [Ping timeout: 245 seconds]06:30
-!- gsomix [~gsomix@] has joined #shogun06:37
-!- lisitsyn [] has quit [Quit: Leaving.]07:43
-!- nube [~rho@] has joined #shogun07:46
-!- nube [~rho@] has quit [Quit: Leaving.]08:02
-!- foulwall [~user@2001:da8:215:503:1d6f:dc69:2111:a45a] has joined #shogun08:12
-!- gsomix_ [~gsomix@] has joined #shogun08:44
-!- gsomix [~gsomix@] has quit [Ping timeout: 248 seconds]08:47
-!- foulwall [~user@2001:da8:215:503:1d6f:dc69:2111:a45a] has quit [Remote host closed the connection]09:14
-!- nube [~rho@] has joined #shogun09:18
-!- lambday [67157d4f@gateway/web/freenode/ip.] has joined #shogun09:29
-!- iglesiasg [] has joined #shogun10:21
-!- mode/#shogun [+o iglesiasg] by ChanServ10:21
@iglesiasggood morning10:22
-!- foulwall [~user@2001:da8:215:503:d2e:b5f1:ff20:a998] has joined #shogun10:32
-!- gsomix_ [~gsomix@] has quit [Remote host closed the connection]10:43
-!- lambday [67157d4f@gateway/web/freenode/ip.] has quit [Quit: Page closed]11:07
-!- Yanglittle [b74040fc@gateway/web/freenode/ip.] has joined #shogun11:16
Yanglittleexcuse me, is there any body?11:16
Yanglittleexcuse me, is there any body?11:16
-!- lisitsyn [] has joined #shogun11:41
lisitsynhey there11:45
-!- van51 [] has joined #shogun12:01
-!- nube [~rho@] has quit [Quit: Leaving.]12:08
-!- lambday [67157f36@gateway/web/freenode/ip.] has joined #shogun12:09
@sonney2kvan51, hey there!12:13
van51sonney2k: hello :)12:13
@sonney2kand hey foulwall !12:13
@sonney2kvan51, so what would you want to do now?12:14
-!- foulwall [~user@2001:da8:215:503:d2e:b5f1:ff20:a998] has quit [Ping timeout: 245 seconds]12:14
van51sonney2k: hmm I think I could add support for quadratic features in the hashed sparse and dense classes12:14
van51sonney2k: don't think it will take long12:15
@sonney2kvan51, it would be could if you could continue the dotfeatures hacking spree doing things like the random fourier features and lbppyrfeatures etc12:15
@sonney2kvan51, I finally finished nips reviews and so I can use my spare time for shogun stuff12:16
van51sonney2k: nice :)12:18
van51sonney2k: well I can also continue with those and wait if Benoir or Olivier have another suggestion for extension or something12:18
-!- thoralf [] has joined #shogun12:19
van51sonney2k: btw on another note, for the GSoC we are supposed to submit some answers about our mentor12:19
van51sonney2k: I have three mentors now I think. Who is supposed to be the official one?12:20
@sonney2kvan51, we are one entity12:21
@sonney2kvan51, and we are happy with what you are doing so no worries either12:22
van51sonney2k: ok then :)12:22
@sonney2kvan51, btw what n-gram length did you use in the webspam thing?12:22
van51sonney2k: size 812:23
@sonney2kahh ok that explains the good performance12:23
van51sonney2k: do you want me to try with size=3?12:23
@sonney2kvan51, and the experiment was with preprocessing>12:23
@sonney2kor with the dot features?12:24
van51sonney2k: no everything was on the fly12:24
van51sonney2k: with dot features12:24
@sonney2kso dot12:24
@sonney2kbut not streaming right?12:24
van51sonney2k: nope, because the current streaming classes read line by line12:24
van51sonney2k: instead of delimiting on \0 eg12:24
@sonney2kas in '\n' as separator12:25
@sonney2kyes o12:25
van51sonney2k: do you want me to try a different setting?12:25
van51sonney2k: I run it on another machine so I won't mind12:25
@sonney2kdid you compare with the results I had in my paper?12:25
@sonney2kvan51, btw how much memory do 100k examples take?12:25
van51sonney2k: I haven't checked that. I know that it crashes on 150k examples12:26
@sonney2ktable 4.412:27
@sonney2klook at the training time12:27
@sonney2k107661 seconds12:28
@sonney2kfor 100k12:28
@sonney2kauROC 99.64%12:28
van51sonney2k: well it's much faster and for a high number of bits in the hash seems to perform better12:29
@sonney2kvan51, what did you have?12:29
van51for eg 16 bits in hash12:30
@sonney2kvan51, I don't have my tomtom mail here so I cannot check12:30
van51sonney2k: it took 11044 seconds and had a auROC of 0.99818312:30
@sonney2kvan51, it is a weird that it is better12:31
@sonney2kshould be the same result12:31
@sonney2k10 times faster it seems - which is good12:31
@sonney2kvan51, would be interesting to see the result when you use the converter in a preprocessing step - with same sqrt normalization12:33
@sonney2kvan51, was that liblinear or svmocas btw?12:33
van51sonney2k: svmocas12:33
@sonney2kepsilon default?12:34
van51sonney2k: yeap and C=0.0112:34
van51sonney2k: I have the code here if you want to check it for anything I missed12:34
@sonney2kvan51, that explains it12:35
@sonney2kvan51, I had C=1 in the experiment12:35
van51sonney2k: and you had used  tri-grams?12:35
@sonney2kvan51, so C=0.01 seems to be a better choice12:36
@sonney2kno 8-grams12:36
van51sonney2k: ah12:36
van51sonney2k: do you want to compare in the same settings?12:36
@sonney2kvan51, you could pick e.g. 5k examples and compare12:36
@sonney2kbut I think it is alright12:37
van51sonney2k: ok I will, just to be thorough :D12:38
van51sonney2k: so about what's next, what do you suggest? quadratic or randomfourier/lbppyr ?12:39
@sonney2kvan51, quadratic is easy to do and very useful so it is good if you do that too. I don't know how to properly stack things though :/12:40
@sonney2kI mean consider that you have n-gram's and want to use quadratic on top12:40
@sonney2kno idea how this could work in a general framework12:41
@sonney2kso for now dense/sparse is OK12:41
van51sonney2k: and the user would specify the indices to be paired?12:41
@sonney2kI have no masterplan12:42
@sonney2kI would say for now do the comparison with C=1 on 5k examples and the quadratic features for dense and sparse12:43
@sonney2kmaybe sth comes to our mind in the meantime12:43
van51sonney2k: ok, just one question about the quadratic features12:44
@sonney2kvan51, btw for dense specifying which pairs should be quadratic is likely a good idea to do (sometimes)12:44
van51sonney2k: if I'm not mistaken in vw the user specifies the namespaces to be pired12:44
@sonney2kand for quadratic features it makes sense to keep also the linear terms (at least optional12:44
van51sonney2k: what's the alternative? to brute-force all possible pairs?12:44
@sonney2kso you have the x_1,...x_D and x_1*x_1,x_1*x_2...x_1*x_D, x_2*x_2...x_2*x_D...12:45
@sonney2kyes exactly12:46
@sonney2kfor dense that is simply 2 nested for loops12:46
@sonney2kfor sparse you do that only over the available dims12:46
van51sonney2k: ok got it12:47
van51sonney2k: will get right on it12:47
@sonney2kso for high sparsitye & sparse it is cheap for dense not so12:47
@sonney2kbut at least you don't need to store D^2 many features in memory12:47
@sonney2kvan51, btw you could compare it with the result of a polynomial kernel svm on some toy data if you use the factors 2*x_1*x_2 for cross terms12:48
@sonney2kwhich you in the final version can just leave out12:48
@sonney2kthey are not doing any good but are a necessary evil when using kernels12:49
@sonney2kso the polykernel cannot do w/o them12:49
@sonney2kvan51, questions so far?12:49
van51sonney2k: I don't think so12:50
@sonney2kif not I will swim a few km now12:50
van51sonney2k: if something comes up, I'll bug you again later12:50
van51sonney2k: nice, hopefully by the end of the week i'll be swimming as well :D12:50
van51sonney2k: although I don't think I can last that much, more like 50m :p12:51
@sonney2kvan51, heh weather is nice today no wind no clouds just calm sea12:51
@sonney2kand water is cold which makes me swim fast12:51
@sonney2kwater here I come12:51
van51sonney2k: enjoy then!12:52
* sonney2k off12:52
@iglesiasgsonney2k, hey! I think there is something going wrong with c++11 in travis clang12:58
-!- HeikoS [~heiko@] has joined #shogun13:18
-!- mode/#shogun [+o HeikoS] by ChanServ13:18
thoralfHey HeikoS13:19
@HeikoSthoralfhi! :)13:19
@HeikoSlambday:  hi!13:19
@HeikoSthoralf: sorry did not yet have time to write the automated testing thing down but will soon! :)13:20
thoralfHeikoS: sonney2k told us, if you get to the channel, please try to fix travis. ;)13:20
@HeikoSthoralf: yes, will do today13:20
thoralftold us to tell you13:20
@HeikoSthoralf: thanks! :)13:21
lambdayHeikoS: hi13:21
@HeikoSlambday: how are things?13:21
lambdayHeikoS: a bit slow this weekend.. :(13:21
@HeikoSlambday: any particular issues?13:21
lambdayHeikoS: hopefully I'll finish COCG_M by tonight13:22
@HeikoSlambday: ok, good to hear. what about the triangular solver?13:22
lambdayHeikoS: no just got occupied with a few other things :(13:22
lambdayHeikoS: well, turns out that alglib actually uses lapack for that particular thing13:22
@HeikoSlambday: I see well we are pretty good in schedule13:22
@HeikoSlambday: I see, mmh13:23
@HeikoSlambday: what about eigen3 again?13:23
@HeikoSI dont like lapack so much13:23
lambdayHeikoS: eigen3 I guess don't have that13:23
@HeikoSlambday: it IS in there, but not exposed in a nice way13:23
@HeikoSwhich is annoying13:23
lambdayHeikoS: not exposed as in?13:24
@HeikoSlambday: cannot access it easily13:24
lambdaybut shall I try to study that instead of incorporating lapack routine?13:24
-!- iglesiasg [] has quit [Quit: Leaving]13:25
@HeikoSlambday: how for advanced are you with that?13:26
lambdaywith the lapack?13:26
@HeikoSJust as with dense matrices, the triangularView() function can be used to address a triangular part of the matrix, and perform triangular solves with a dense right hand side:13:26
@HeikoSisnt that helpful?13:27
lambdayHeikoS: can we set the subdiagonals in sparse easily?13:27
@HeikoSlambday: dont know ...13:27
lambdaywell, for lanczos, the system is not that huge13:28
lambdaywe can use dense things13:28
@HeikoSlambday: yes before I meant how much of the lapack/alglib integration have you done13:28
lambdayHeikoS: not started with that yet, was working on COCG_M13:28
@HeikoSlambday: okay good,13:28
@HeikoSso its worth thinking about it13:29
@HeikoSfor what do you need the triangular exactly again?13:29
lambdayfor lanczos, it sets the diagonal of a matrix from some values generated in its cg iterations13:29
lambdayso, its not that huge13:29
lambdayHeikoS: can we think about importing readymade Lanczos implementation from somewhere?13:30
lambdayHeikoS: well, that would create another dependency.. CG things are pretty much same13:31
@HeikoSlambday: would be good, but also sounds painful ;)13:31
@HeikoSlambday: so erlend spent quite some time on alglib design13:31
@HeikoSmaybe its a good idea just to make it the same13:31
lambdayHeikoS: well, krylstat simply uses alglib routine for solving the tridiagonal system.. we have to import that method in shogun if we wanna do the same13:32
@HeikoSlambday: we are already depending on lapack13:33
@HeikoSand the I ported quite a few alglib routines already, its a bit work but not hard to to and doesnt create dependencies13:33
lambdayHeikoS: I was taking a look.. they use quite a few internal data structures which I gotta check how they works... plus they say that they imported that particular routine from lapack13:34
@HeikoSlambday: oh ....13:34
lambdayI was then having a look at our lapack.h13:34
lambdayif that was already there..13:34
lambdaybut I didn't get how our lapack works13:35
@HeikoSlambday: ha, through chaos :D13:35
lambdayhehe :D13:35
@HeikoSok, we should check this out a bit more then, I wil lhave lunch now and then be back, will you be around a bit longer?13:35
lambdayHeikoS: yes I am here13:35
@HeikoSlambday: ok cool13:36
lambdayHeikoS: enjoy your meal :)13:36
@HeikoSlambday: ah btw13:36
lambdayHeikoS: this is another paper? :D13:36
@HeikoSlambday: yes thats the nips kernel sampler13:37
@HeikoSlambday: nice pictures in there ;)13:37
@HeikoSok going now, see you later13:37
lambdayHeikoS: I am checking.. ciao :)13:37
-!- HeikoS [~heiko@] has left #shogun []13:41
-!- iglesiasg [~iglesias@2001:6b0:1:1041:788e:31ff:81c6:5bc9] has joined #shogun14:07
-!- mode/#shogun [+o iglesiasg] by ChanServ14:07
-!- foulwall [~user@2001:da8:215:503:d2e:b5f1:ff20:a998] has joined #shogun14:08
-!- FSCV [~FSCV@] has joined #shogun14:08
-!- nube [~rho@] has joined #shogun14:33
-!- foulwall [~user@2001:da8:215:503:d2e:b5f1:ff20:a998] has quit [Read error: Connection reset by peer]14:34
-!- HeikoS [] has joined #shogun14:47
-!- mode/#shogun [+o HeikoS] by ChanServ14:47
lambdayHeikoS: hi14:48
@HeikoSloa hi14:49
@HeikoSlambday:  hi14:49
lambdayHeikoS: so in the Lanczos, we form this T matrix.. whose dimension is nxn, n=number of cg iteration...14:50
lambdayHeikoS: now, I can treat that as a dense matrix and run a self adjoint eigen solver14:50
lambdaybecause, eigen3 has this tridiagonalization thing... but I don't know if I can use that for solving eigenvalue problems14:51
@HeikoSlambday: it *is* in fact dense richt?14:51
lambdayHeikoS: its tridiagonal...14:51
lambdayHeikoS: not so dense14:51
@HeikoSlambday: ah14:52
@HeikoSthat would only work for a few thousand iterations then14:52
@HeikoSso not so good14:52
lambdayyep :(14:52
lambdayI guess that's why erlend used particular method for tridiagonal specifically14:53
lambdaywhich is not aparently there in eigen314:53
@HeikoSlambday: I guess so14:53
@HeikoSlambday: lets maybe discuss with him?14:53
@HeikoSlambday: could you write an email, summarise the problem and possible solutions?14:53
@HeikoSmaybe he has some new insights14:53
@HeikoSlambday:  how long do you think until the rest is finished and we need this?14:54
lisitsyn I heard Lanczos14:54
lisitsynI have highlight on this word lol14:54
lambdayHeikoS: this is for the extremal eigenvalues... so even if our COCG_M works, we can't approximate logdet until we have this module working14:55
lambdayI planned to finish it within next week14:55
@HeikoSlambday: ok14:55
lambdayand then move on towards graph coloring things cause the rest of the stuffs will be almost ready14:55
@HeikoSlambday: : might be good to ask sergey :)14:55
@HeikoSlambday: the is always my joker ;)14:55
lambdaylisitsyn: hehe :D14:55
@HeikoSlisitsyn: how is it going btw?14:55
@HeikoSlong time no see14:55
lisitsynHeikoS: hey14:56
@HeikoSsettled in again?14:56
lisitsynHeikoS: yeah14:56
lambdayI saw you guys' pics :D14:56
lambdaylisitsyn: we need Lanczos for large sparse Hermitian matrix14:57
lambdayHeikoS: shall I cc to Dan too?14:58
-!- iglesiasg [~iglesias@2001:6b0:1:1041:788e:31ff:81c6:5bc9] has quit [Quit: Ex-Chat]15:00
-!- iglesiasg_ [] has joined #shogun15:00
@HeikoSlambday: yes15:02
@HeikoSlisitsyn: or do you have an idea?15:02
@HeikoSlisitsyn: since we a talked about this before15:02
@HeikoSlambday: also cc sergey15:02
lambdayHeikoS: alright :)15:02
@sonney2kvan51, no idea if it is good news but I survived ;)15:03
-!- iglesiasg_ [] has quit [Read error: Connection reset by peer]15:03
-!- iglesiasg__ [] has joined #shogun15:03
@sonney2khey HeikoS !15:03
@sonney2kHeikoS, any plans to get the tests back to greeeen?15:03
@HeikoSsonney2k: yes will do that later today15:03
@HeikoSsorry for having broken them for so long15:04
@HeikoSalready spent so much time on fixing that I had to do some other things I postponed ;)15:04
@HeikoSNIPS review is coming soon and we have to prepare stuff for that15:04
@HeikoSsonney2k: btw
@HeikoSlisitsyn:  ^15:05
@sonney2kHeikoS, well the alternative is to blacklist all failing tests15:08
@HeikoSsonney2k: nono, I am sure I can fix them soon15:08
@sonney2kHeikoS, thoralf yesterday figured out that we have some failing examples in static interfaces15:09
@sonney2kHeikoS, so these might be due to some fixes in the GUI* classes15:09
@HeikoSsonney2k: I hope that wasnt me touching the GUI15:09
@HeikoSsonney2k: ah damn15:09
@HeikoSsonney2k: ok will have a look into them15:09
@sonney2kHeikoS, could be :P15:09
@HeikoSwhich language?15:09
thoralfHeikoS: cmdline_static15:10
@HeikoSsonney2k: buildbot should tell that right?15:10
@sonney2kHeikoS, well get the tests back to work and we will see anyway15:10
@sonney2kHeikoS, the buildbot doesn't build the rest if already libshogun & tests fail15:10
@HeikoSsonney2k: ah15:10
@sonney2kHeikoS, btw I want to get rid of the get_feature_type / get_feature_class / get_kernel_type and all these ENUMS15:10
@HeikoSso that slipped through15:10
iglesiasg__sonney2k, HeikoS, thoralf: and what's with travis clang and c++11?15:10
@HeikoSsonney2k: ok I will work on it soon15:11
@sonney2kiglesiasg__, is clang not supporting c++11? What do you mean?15:11
iglesiasg__sonney2k, let me show you15:11
@sonney2kHeikoS, ahh and maybe important to I added a check to detect c++1115:11
@sonney2kHeikoS, if available we use atomic<int> for locking in SGReferencedData (this has 0 memory overhead!!)15:12
iglesiasg__sonney2k, huge pile of cc1plus: error: unrecognized command line option '-std=c++11'15:12
@sonney2kiglesiasg__, clang is a b0rken mess15:13
iglesiasg__sonney2k, it seems that started with the add cxx11 support:
-!- iglesiasg__ is now known as iglesiasg15:14
-!- mode/#shogun [+o iglesiasg] by ChanServ15:14
@sonney2khow can they not fail when they don't support an option15:14
@sonney2kiglesiasg, so it seems the build just went fine15:14
@iglesiasgsonney2k, mmm I don't think so15:16
-!- shogun-notifier- [] has joined #shogun15:17
shogun-notifier-shogun: Soeren Sonnenburg :develop * 99cfdb8 / src/configure:
shogun-notifier-shogun: use -Werror to workaround b0rken clang15:17
lambdayHeikoS: sent the mail15:19
shogun-notifier-shogun: Soeren Sonnenburg :develop * 79ad3c1 / src/shogun/features/ (2 files):
shogun-notifier-shogun: work in the shogun namespace even in the .cpp file for templated classes15:19
@sonney2kiglesiasg, indeed15:20
@HeikoSlambday: nice thanks, lets see what they say ,we still got some days until COCG is finished15:20
lambdayHeikoS: yes :)15:20
lambdayHeikoS: by the way, so the stuffs left as of now are 1)COCG_M 2) Lanczos and 3) probing vectors, right?15:21
lambdayand preconditioned CG15:21
@HeikoSlambday: yes15:21
@HeikoSpretty much15:21
@HeikoSprobing will be some work15:21
@HeikoStoo :)15:21
@HeikoSso thats the hardest parts of the project15:21
@HeikoSafter that, we will spend some time on testing and documenting15:21
@HeikoSmaybe even write some static interfaces as long as the R-bindings do not properly work15:22
lambdayalright :)15:22
lambdayHeikoS: well one thing... how much accuracy can we expect using Gaussian samples by the way?15:22
lambdayHeikoS: have you noticed anything in particular?15:23
@sonney2kiglesiasg, hopefully fixed...15:23
@iglesiasgsonney2k, yes, let's see15:23
@HeikoSlambday: well in the limit, 100%15:23
@HeikoSand with finite samples, I have no idea ;)15:24
@HeikoSlambday: why?15:24
lambdayHeikoS: just trying to get an idea whether I should get depressed when I write the unit-tests next week and it turns out to be not so accurate :(15:24
@HeikoSlambday: ah, no dont worry15:24
@HeikoSlambday: in fact, we should not unit test the accuracy too much15:24
@HeikoSrather use all basis vectors15:25
@HeikoSto get the exact trace15:25
@HeikoSlambday: the tests should not take too long15:25
@sonney2klambday, ohh you have seen the picture of us? Are you as shocked as I am how tall HeikoS is?15:26
@sonney2kGunnar must have felt puny!15:26
@HeikoSeveryone should be shocked! ;)15:26
* sonney2k already did15:26
lambdaysonney2k: lol you are tall too :-o15:26
lambdayHeikoS: you're 6'4"??15:26
@HeikoSlambday: 2meters15:26
@sonney2k10cm less15:27
* sonney2k ^15:27
lambdayerrr... :D15:27
@sonney2klambday, competition is not good sometimes ;)15:27
@HeikoStall people are awesome, as we always say in the boxing gym >:o15:27
lambdayhaha :D15:28
@iglesiasgI felt I was a dwarf next to you guys in Berlin :(15:28
lambdayiglesiasg: I'd have too :(15:28
lambdayiglesiasg: average Indian guy :(15:28
@iglesiasglambday, hehe how long is that? I am 1.7315:29
@HeikoSiglesiasg: more wheatbeer is needed then15:29
-!- gsomix [~gsomix@] has joined #shogun15:29
thoralflambday: What does average mean?15:29
@iglesiasgHeikoS, I was missing that! :)15:29
thoralfiglesiasg: That's good european average :)15:29
lambdayiglesiasg: well, 5'7"15:29
lambdayso, that too is around 1.72 I guess15:30
@iglesiasglambday, pretty much the same as me then :)15:30
lambdayiglesiasg: brother :D15:30
@iglesiasgthoralf, well living in Sweden and after my visit to Berlin, I think I am below the average :P15:30
@iglesiasgbut when I get to the subway in Madrid, that feeling changes15:31
@sonney2klambday, if we organize sth next year maybe you can come :)15:31
lambdaysonney2k: I so wish to.. let's see how things are :)15:32
shogun-buildbotbuild #1492 of deb1 - libshogun is complete: Failure [failed test]  Build details are at  blamelist: Soeren Sonnenburg <>15:32
@iglesiasgsonney2k, the flag work fine for clang travis indeed15:33
@sonney2kHeikoS, regarding the failing static interfaces - if one fails all fail - that is the good thing :D15:34
thoralfiglesiasg: Wikipedia says for men its 177cm as of 2003.15:35
shogun-notifier-shogun: Soeren Sonnenburg :develop * b199658 / src/configure:
shogun-notifier-shogun: fix c++11 error15:36
shogun-buildbotbuild #1491 of deb1 - libshogun is complete: Failure [failed test]  Build details are at  blamelist: Soeren Sonnenburg <>15:37
lambdaythoralf: is it global avg?15:37
@iglesiasgthoralf, almost there then!15:37
thoralflambday: No, Germany.15:37
thoralflambday: Sorry, Europe.15:38
thoralflambday: Again, sorry.  The statistics for 2003 are for Germany only.15:38
lambdaythoralf: hehe15:39
lambday1.647 m for India according to wiki15:39
shogun-buildbotbuild #1493 of deb1 - libshogun is complete: Failure [failed test]  Build details are at  blamelist: Soeren Sonnenburg <>15:43
lisitsyniglesiasg: well I was more like dwarf15:46
-!- travis-ci [] has joined #shogun15:46
travis-ci[travis-ci] it's Soeren Sonnenburg's turn to pay the next round of drinks for the massacre he caused in shogun-toolbox/shogun:
-!- travis-ci [] has left #shogun []15:46
gsomixyou need a beard15:50
lisitsyniglesiasg: I don't remember, am I taller?15:52
lambdayHeikoS: regarding COCG_M, I have a problem and a solution16:01
lambdayHeikoS: its tough to get all the solution vectors to be returned by our solve method, because the linear operator is real, the vector is real but just because of the complex shifts the solution is complex16:02
@iglesiasglisitsyn, I think you are taller than me yeah16:02
@iglesiasglisitsyn, I am about 173cm16:02
lambdayHeikoS: so I can't extend the base properly and its hard to fit in the system.. but I want to keep the base same (CLinearSolver)16:02
lisitsyniglesiasg: I am around that too16:02
lisitsynbut 2x of you anyway16:02
lambdayHeikoS: so, I am planning to add another level of abstraction - extending CIterativeLinearSolver (subclass of CLinearSolver) to CLinearShiftedFamilySolver... where shifts can be real/complex (templated)16:03
lisitsyniglesiasg: in other dimension16:03
lambdayHeikoS: then, in COCG_M implementation of that, solve just returns the basic solution (similar to what Krylstat does)16:04
lambdayHeikoS: and there will be two abstract methods for specially handling the shifts..16:04
lambdayso, in the COCG_M job, we'll first solve (which computes all of these and stores the solution matrix inside it).. and a subsequent get call will return the actual solution for the whole family16:05
@iglesiasgsonney2k, did you have time to have a look at the new LMNN PR?
@iglesiasgsonney2k, I addressed the issues you pointed out a some days ago. I think it should be ready to merge16:06
lambdayHeikoS: we can use this for CG_M/COCG_M as needed (although not for our project but may be later) - will be flexible16:06
lambdayHeikoS: the only downside is that, we'll have to remember calling "solve" before we call the "get" each time - but that's not a problem I guess16:07
lambdayor, we'll move solve inside that "get", call it "solve_shifted" or something, and in COCG_M jobs, call that instead16:07
lambdayHeikoS: I can finish this implementaion by tonight and send a PR and then may be you can have a look (not much work so I won't get sad if I gotta change it later :D)16:13
-!- travis-ci [] has joined #shogun16:13
travis-ci[travis-ci] it's Soeren Sonnenburg's turn to pay the next round of drinks for the massacre he caused in shogun-toolbox/shogun:
-!- travis-ci [] has left #shogun []16:13
lambdayHeikoS: brb16:18
-!- pickle27 [] has joined #shogun16:21
-!- travis-ci [] has joined #shogun16:39
travis-ci[travis-ci] it's Soeren Sonnenburg's turn to pay the next round of drinks for the massacre he caused in shogun-toolbox/shogun:
-!- travis-ci [] has left #shogun []16:39
-!- iglesiasg [] has quit [Ping timeout: 245 seconds]16:51
-!- iglesiasg [~iglesias@2001:6b0:1:1041:788e:31ff:81c6:5bc9] has joined #shogun17:04
-!- mode/#shogun [+o iglesiasg] by ChanServ17:06
thoralfgsomix: Doxygen gives a lot of warnings for your classes.  Mostly because the variables in @param are not matching the signatures.17:08
thoralfgsomix: Reproducable when building i.e. python_static17:09
gsomixthoralf, woops.17:10
gsomixI'll see, thanks.17:10
thoralfgsomix: Just check it when you're finished - this always happens with heavily chaning classes :)17:11
-!- foulwall [] has joined #shogun17:14
-!- lambday [67157f36@gateway/web/freenode/ip.] has quit []17:16
gsomixsonney2k, good evening, mr. Soeren.18:11
gsomixsonney2k, I have prepared pack of changes for shogun. have you a minute to look at it?18:12
pickle27hey guys18:13
pickle27having the weirdest problem here18:13
pickle27eigen3 will not transpose a matrix18:13
pickle27like the code runs but it doesn't get transposed18:13
pickle27anyone else ever seen this?18:13
@iglesiasgpickle27, what are you doing exactly?18:14
lisitsynpickle27: haha nice18:14
pickle27Im computing eigenvectors and then I need to transpose them18:14
@iglesiasgpickle27, A.transpose() makes the transpose of A by doesn't assign to it, I believe18:14
pickle27yeah you need to use transposeInPlace()18:15
lisitsynyes A.transpose() is expression18:15
pickle27but that doesn't work either18:15
gsomixvan51, hey. there are some changes with line reader in my PR.18:15
@iglesiasgso transposeInPlace() is like A=A.tranpose()18:15
pickle27yeah it explicitly says in the docs not to use A = A.transpose()18:15
lisitsynpickle27: too weird to be true!18:15
gsomixvan51, I replaced it with LineReader. more useful class.18:15
lisitsynI use transposeInPlace in my code18:16
lisitsynit works18:16
pickle27it usually works for me too18:16
van51gsomix: you mean FileReader?18:16
pickle27but it isn't in this case18:16
pickle27and I had another time the other day when it wouldn't either18:16
@iglesiasgpickle27, x-files!18:16
pickle27maybe it is transposing ...18:22
@iglesiasgpickle27, are you doing array() or anything else on the way?18:22
pickle27its confusing because of the values its a very symmetric thing18:22
pickle27i think I need to sort actually instead18:22
gsomixvan51, oops, exactly. :)18:23
van51gsomix: ok! thanks for letting me know :018:23
-!- shogun-notifier- [] has quit [Quit: transmission timeout]18:36
-!- foulwall [] has quit [Ping timeout: 248 seconds]18:56
lisitsynpickle27: iglesiasg: gsomix: van51: guess what19:12
@iglesiasglisitsyn, almost written! :)19:12
pickle27its coming!19:12
@iglesiasgpickle27, race! :P19:12
pickle27iglesiasg: I haven't started writting it yet, trying to fix up some code first19:13
@iglesiasgpickle27, aah ok. Next week then ;)19:14
gsomixlisitsyn, reports time!
@iglesiasggsomix, dafuq O_O19:17
-!- lisitsyn [] has quit [Quit: Leaving.]19:21
gsomixiglesiasg, cool cartoon. I'm highly recommend.19:21
gsomixcartoons is for strong man :319:22
gsomixI need a book on English grammar =___=19:23
pickle27hey guys19:25
pickle27chasing down a slight bug in my program19:25
pickle27is this a rotation?19:26
@iglesiasgpickle27, mmmm (thinking)19:28
@iglesiasgpickle27, I think that if you regard the first matrix as a rotation matrix of angle theta, then the second matrix is the rotation matrix of angle theta plus or minus (not sure about the sign) pi/219:36
pickle27thats what I was thinking19:36
pickle27I'll try that out19:37
pickle27narrowing this down further20:26
pickle27I have sign differences in my eigenvectors between matlab and eigen320:26
pickle27any idea how to fix this?20:27
pickle27looks like the eigen3 equivalent of matlabs eig might be shcur20:28
pickle27still has sign mismatch20:33
-!- zxtx [] has quit [Ping timeout: 256 seconds]20:36
@HeikoSsonney2k: sorry have to postpont builds to tomorrow, but then for sure20:38
gsomixvan51, hey20:43
gsomixI have question20:44
van51gsomix: go ahead20:44
gsomixwebspam_example is your code? it's now at develop branch&20:44
gsomixsorry, no more questions20:45
van51gsomix: yes it's my code20:46
van51gsomix: but it's not submitted20:46
van51gsomix: I lost you somewhere :p20:46
gsomixit's just in my local rep, ok.20:46
van51gsomix: ah ok20:47
gsomixall ok now20:48
van51gsomix: I also have to adjust it to your new class20:48
van51gsomix: why are you doing this to me20:48
-!- lisitsyn [~lisitsyn@] has joined #shogun20:48
pickle27lisitsyn: narrowed down my problem20:49
pickle27sign differences on the eigenvectors found using eigen320:49
pickle27isn't the same result as octave's eig() function20:49
lisitsynpickle27: well different signs are kind of ok20:50
pickle27usually yeah but not for Jade apparently20:50
gsomixvan51, huh. life is pain. >:320:50
lisitsynpickle27: why?20:50
gsomixvan51, btw free testing.20:51
pickle27messes the cumulant matrices20:51
lisitsynpickle27: do you need positive eigenvalues?20:51
pickle27I don't neccessarily need positive ones I just need the same signs as matlab20:51
lisitsynpickle27: are they all positive with matlab?20:52
lisitsynor negative20:52
lisitsynpickle27: what about eigenvalues?20:54
pickle270.259276 and 2.15752 in both eigen3 and octave20:55
lisitsynpickle27: but different eigenvectors? weird!20:56
pickle27Im just going to flip the sign if they are all neg to be pos20:57
@iglesiasglisitsyn, is it so weird that there are several eigenvectors for the same eigenvalue?21:12
lisitsyniglesiasg: I think so21:16
@iglesiasglisitsyn, I just looked it up. Let me write it21:17
@iglesiasglisitsyn, Given that v is an eigenvector of a matrix A, then any scaled eigenvector c*v is also an eigenvector for the same eigenvalue.21:17
@iglesiasgI thinks it makes pretty much sense21:18
lisitsyniglesiasg: true21:18
lisitsynyeah my bad21:18
@iglesiasglisitsyn, it was good to recall the property :)21:18
lisitsynso signs are whatever :)21:18
pickle27yeah except it is affecting the algorithm ahhhhh21:19
@iglesiasgtrue, the constant can be negative21:19
@iglesiasgpickle27, maybe one of the implementations normalize the eigenvectors so they are unit21:19
lisitsynpickle27: then something like positivity is forced21:19
pickle27hmm yeah could be that21:19
pickle27it doesn't make any sense21:20
pickle27I have 2 examples one ported from scikit learn and one that andreas gave me21:20
pickle27in the scikit one eigen gives mainly neg eigenvectors and the orig alg gives positive21:21
pickle27then its the opposite on andreas' example21:21
lisitsynpickle27: I wonder why sign matters here21:24
lisitsynif it operates eigenvectors it should be ready to sign changes21:25
pickle27Im not sure21:28
lisitsynprince is born21:34
@iglesiasglisitsyn, what?21:35
@iglesiasgaaah UK?21:35
lisitsyniglesiasg: yeah21:36
-!- lisitsyn [~lisitsyn@] has quit [Ping timeout: 240 seconds]21:42
gsomixgood night guys21:44
@sonney2kvan51, please fix the warnings travis reports
-!- lisitsyn [~lisitsyn@] has joined #shogun21:44
-!- lisitsyn [~lisitsyn@] has quit [Client Quit]21:44
@sonney2kvan51, btw did C=1 give a more similar result?21:44
van51sonney2k: yea the speedup isn't that bigger with C=121:46
van51sonney2k: I think it's about the half in some cases21:47
@sonney2kvan51, I meant the auROC value...21:47
van51sonney2k: yeah the same pretty much21:47
van51sonney2k: I will make a table so that they can be compared more easily21:47
@sonney2kvan51, in that case use a more relaxed epsilon please21:47
@sonney2kocas is using a rather tough criterion, svmlight certainly is not as accurate21:48
-!- nube [~rho@] has quit [Ping timeout: 246 seconds]21:48
van51sonney2k: ok I will do that as well and I'll give a summary by tomorrow21:48
van51sonney2k: I have to go for now, but I'll make a PR for that warning when I get back21:49
van51sonney2k:  I 'll also have one for quadratic on hashed dense features21:50
@sonney2kvan51, cool!21:51
@sonney2kcu then21:51
@sonney2kthoralf, do you have clang installed?21:52
-!- van51 [] has quit [Quit: Leaving.]21:55
@sonney2kiglesiasg, do you have clang?21:55
@iglesiasgsonney2k, of course!21:56
@iglesiasgbest compiler ever :)21:56
@iglesiasghave you seen the error/warning messages it gives? I wish gcc worked like that21:56
@iglesiasgactually, I just like it because of that, no idea if it is better/worse thatn gcc in other aspects21:57
-!- HeikoS [] has quit [Quit: Leaving.]22:00
-!- nube [~rho@] has joined #shogun22:02
@sonney2kiglesiasg, could you please run configure with latest shogun and --cxx=clang++ ?22:04
@sonney2kiglesiasg, and I agree about the quality of warning/error messages22:04
@iglesiasgsonney2k, sure, I thought you were never going to ask :)22:04
@iglesiasgsonney2k, should I enable any part in particular?22:05
@sonney2kiglesiasg, interfaces= is sufficient22:05
@iglesiasgall right22:05
@sonney2kiglesiasg, well you know I am on vacations so I am distracted even more22:06
@iglesiasgsonney2k, how are your holidays going btw? Lot of sun-bathing and swimming?22:06
@sonney2kI think i was only doing a ~2km swim today but the crazy thing is that my feet hurt from the walking back22:07
@iglesiasgsonney2k, I noticed Checking for C++11 Atomic Integer ... no, should it be yes?22:08
@sonney2kiglesiasg, which version is it?22:09
@iglesiasgsonney2k, clang?22:10
@iglesiasgUbuntu clang version 3.0-6ubuntu322:10
@sonney2kahh ok too old22:11
@sonney2k3.1 introduce atomic22:11
@iglesiasgaham I see22:12
@iglesiasgsonney2k, did you want to check that thing or it doesn't matter?22:12
@sonney2kand it seems one needs clang3.2 for anything reasonable22:14
@sonney2kwell would be nice22:14
@sonney2kmaybe wiking has a newer clang?22:14
@iglesiasgaham I see22:14
@iglesiasgsonney2k, I can check in my laptop too. I tried in the office computer so far22:15
@sonney2kiglesiasg, check clang++ -v first22:15
@sonney2kif the version is recent enough22:15
@iglesiasgsonney2k, 3.2.1 in my laptop22:16
@sonney2kiglesiasg, good enough22:17
@sonney2kiglesiasg, so please try22:17
@sonney2kit should detect c++1122:17
@iglesiasgsonney2k, it did indeed22:18
-!- zxtx [] has joined #shogun22:18
@iglesiasgsonney2k, I see c++ Atomic integer yes, but no c++11 flag in the list of c++ compiler flags22:19
@iglesiasgsonney2k, I am not sure if it should appear there too22:19
@sonney2kiglesiasg, well type make :)22:20
@iglesiasgsonney2k, compiling...22:21
@sonney2kiglesiasg, done?22:29
@iglesiasgsonney2k, nope, still compiling22:29
@iglesiasgsonney2k, done22:33
@iglesiasgsonney2k, what now?22:34
-!- FSCV [~FSCV@] has quit [Ping timeout: 245 seconds]22:41
@sonney2kiglesiasg, all good!22:42
@iglesiasgsonney2k, all right!22:42
@iglesiasgsonney2k, btw, did you have a moment to check the LMNN PR?22:42
@sonney2kiglesiasg, no :/22:44
@sonney2kgsomix, has sent a >1kLoC monster PR22:45
@iglesiasgsonney2k, hehe I read your comment22:45
@sonney2kand he has prio22:45
@iglesiasgsonney2k, then for another day maybe, mine is not precisely short actually. I think it is about 1k (including two class headers and two files with unit tests though)22:46
@sonney2kiglesiasg, maybe lisitsyn or so has time?22:48
@sonney2kI ahve to sleep now due to too nice weather I am occupied swimming at lunch break22:49
@iglesiasgsonney2k, he already took a look and said it was fine22:49
@iglesiasgsonney2k, bye!22:50
-!- shogun-notifier- [] has joined #shogun23:00
shogun-notifier-shogun: Fernando Iglesias :develop * cefe77a / src/shogun/ (6 files):
shogun-notifier-shogun: LMNN implementation based on auxiliary methods hidden in an implementation class.23:00
shogun-notifier-shogun: Fernando Iglesias :develop * 2913219 / tests/unit/metric/,tests/unit/metric/
shogun-notifier-shogun: A few unit tests for LMNN23:00
shogun-notifier-shogun: Fernando Iglesias :develop * 263c8d7 / / (8 files):
shogun-notifier-shogun: Merge pull request #1270 from iglesias/feature/lmnn23:00
shogun-notifier-shogun: LMNN implementation and a couple of unit tests23:00
shogun-buildbotbuild #1494 of deb1 - libshogun is complete: Failure [failed test]  Build details are at  blamelist: Fernando Iglesias <>23:03
shogun-buildbotbuild #1495 of deb1 - libshogun is complete: Failure [failed test]  Build details are at  blamelist: Fernando Iglesias <>23:17
-!- hushell [] has joined #shogun23:22
@iglesiasghey hushell23:22
hushellhey iglesiasg23:22
hushellI just got connected23:22
@iglesiasghushell, sure, I think I will be able to have a look at it by tomorrow morning23:22
hushellI have a meeting with Patrick tonight23:23
hushellmaybe he has more comments on it23:23
@iglesiasghushell, hehe tonight when?23:23
@iglesiasgI get confused when talking to people in different timezones :)23:23
@iglesiasghow many hours from now you mean with tonight?23:23
hushell6 or 7 hours later23:23
@iglesiasghushell,  so with tomorrow morning you mean about the time when the meeting for the mid-term evaluation takes place, right?23:24
hushellbut I don't think there are will be many things need to revise for him23:24
hushellWhat time for you right now?23:25
@iglesiasghushell, 23h2523:25
@iglesiasgor 11:25PM23:25
hushellyeah, I guess right :) so I say tomorrow morning your time23:25
@iglesiasghushell, ok got it :)23:26
hushelliglesiasg: okay, talk you later23:26
@iglesiasghushell, all right, bye!23:27
@iglesiasgsee you later, people23:35
-!- iglesiasg [~iglesias@2001:6b0:1:1041:788e:31ff:81c6:5bc9] has quit [Quit: Ex-Chat]23:35
-!- travis-ci [] has joined #shogun23:47
travis-ci[travis-ci] it's Fernando Iglesias's turn to pay the next round of drinks for the massacre he caused in shogun-toolbox/shogun:
-!- travis-ci [] has left #shogun []23:47
--- Log closed Tue Jul 23 00:00:43 2013