IRC logs of #shogun for Thursday, 2013-05-02

« Wednesday, 2013-05-01 Index Friday, 2013-05-03 »
--- Log opened Thu May 02 00:00:29 2013
cameron__not sure my changes went in00:08
cameron__do people usually commit via the submodule - data or the master shogun project?00:08
-!- cameron__ [~quassel@global-2-1.nat.csx.cam.ac.uk] has quit [Remote host closed the connection]00:16
gsomixsonney2k, I sent the proposal.00:45
gsomixnite, guys00:50
@iglesiasgnight00:51
-!- codinninja [~satya@202.3.77.221] has joined #shogun01:20
-!- codinninja [~satya@202.3.77.221] has left #shogun []01:21
-!- syst3mw0rm [~quassel@ec2-54-235-165-148.compute-1.amazonaws.com] has left #shogun ["http://quassel-irc.org - Chat comfortably. Anywhere."]01:57
-!- iglesiasg [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving]02:16
shogun-buildbotbuild #326 of nightly_none is complete: Failure [failed compile]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_none/builds/32603:03
-!- sids_aquarius [~sidi_@li400-124.members.linode.com] has quit [Quit: Argh! Flaky internet]03:31
-!- votjakovr [~votjakovr@host-46-241-3-209.bbcustomer.zsttk.net] has joined #shogun04:46
-!- wencan [~wencan@c-71-61-182-121.hsd1.pa.comcast.net] has quit [Quit: wencan]06:29
-!- mizobe [~mizobe@189.35.244.197] has joined #shogun07:21
-!- abi_ [3d0c137b@gateway/web/freenode/ip.61.12.19.123] has joined #shogun07:36
-!- abi_ [3d0c137b@gateway/web/freenode/ip.61.12.19.123] has quit [Ping timeout: 245 seconds]07:44
-!- foulwall [~foulwall@li379-21.members.linode.com] has joined #shogun07:45
-!- abinayam [3d0c137b@gateway/web/freenode/ip.61.12.19.123] has joined #shogun08:17
abinayamHi. Do we need to submit different proposals for different topics we are interested in or a single proposal listing all the topics is enough?08:18
-!- abinayam [3d0c137b@gateway/web/freenode/ip.61.12.19.123] has quit [Client Quit]08:21
@sonney2kgsomix, seen it. You should get your hands dirty in that direction now. Write some function that reads a single line of an ascii file - like GNU's getline08:25
@sonney2kand maybe start a .csv reader08:25
@sonney2kabinayam, one per topic but no more than 2-3 otherwise we won't take you seriously :)08:27
@sonney2kHeikoS, did you merge the last commit? Somebody forgot the git push to data / merge there08:29
-!- shogun-notifier- [~irker@7nn.de] has joined #shogun08:31
shogun-notifier-shogun: Soeren Sonnenburg :develop * fdedc4c / src/shogun/regression/GaussianProcessRegression.h: https://github.com/shogun-toolbox/shogun/commit/fdedc4ce61f2921af61f9c4d5bfdcf68ecc8b46a08:31
shogun-notifier-shogun: fix compile error - GP regression only needs eigen3 now08:31
shogun-buildbotbuild #730 of cyg1 - libshogun is complete: Failure [failed configure]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/cyg1%20-%20libshogun/builds/730  blamelist: Soeren Sonnenburg <sonne@debian.org>08:34
gsomixsonney2k, ok!08:39
shogun-buildbotbuild #895 of bsd1 - libshogun is complete: Success [build successful]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/bsd1%20-%20libshogun/builds/89508:40
-!- abinayam [3d0c137b@gateway/web/freenode/ip.61.12.19.123] has joined #shogun08:43
abinayam@sonney2k : I am interested in 2 topics only. Thank you :)08:44
-!- sonne|work [~sonnenbu@sams-office-nat.tomtomgroup.com] has joined #shogun08:46
-!- abinayam [3d0c137b@gateway/web/freenode/ip.61.12.19.123] has quit [Ping timeout: 245 seconds]08:59
shogun-buildbotbuild #327 of nightly_none is complete: Success [build successful]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_none/builds/32709:00
-!- foulwall [~foulwall@li379-21.members.linode.com] has quit [Ping timeout: 268 seconds]09:10
-!- erlenda_ [~erlenda@fw-oslo.intelcom.no] has joined #shogun09:18
-!- erlenda [~erlenda@fw-oslo.intelcom.no] has joined #shogun09:18
-!- erlenda_ [~erlenda@fw-oslo.intelcom.no] has left #shogun ["Leaving"]09:23
-!- iglesiasg [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun09:25
-!- mode/#shogun [+o iglesiasg] by ChanServ09:25
-!- vgorbati [d4029f22@gateway/web/freenode/ip.212.2.159.34] has joined #shogun09:30
hushellHi Soeren, may I ask a few questions about SGVector?09:42
hushellIf I want to change the size of a sgvector in a function, and I want the change being passed back09:43
hushellCan I just use resize_vector inside that function?09:44
@iglesiasghey hushell09:44
@iglesiasghushell: yeah, that should work09:45
hushellWhat I am curious about is when passing a SGvector as a parameter, I don't need to use * or reference09:45
@iglesiasghushell: you should pass the reference09:46
@iglesiasgI think that at least that passing the reference makes more explicit what you will do09:47
sonne|workhushell: yes resize_vector09:47
hushellThanks! Are you Fernando?09:47
@iglesiasghowever, I think that even if you don't use reference, what you want to do will happen09:47
sonne|workwhat is important to know here is that sgvectors properties are always copied09:47
@iglesiasgI am not entirely sure about it though09:47
sonne|workso if you resize_vector09:47
@iglesiasghushell: yeah I am Fernando :)09:47
sonne|workit won't change the original size09:48
sonne|workif you jsut do09:48
sonne|worka=b09:48
hushellNice to meet you! I am Shell09:48
sonne|workb.resize_vector(...)09:48
@iglesiasghushell: nice to meet you to09:48
@iglesiasgtoo*09:48
hushellSo in my case rescale_outputs(const SGVector outputs, SGVector& new_outputs), Soeren you suggest inplace modify the outputs, and the function can be rescale_outputs(SGVector& outputs)?09:50
hushellThe case is I have a vector, maybe length of 6, inside the function, the length of vector may be changed to 409:52
-!- cameron__ [~quassel@global-2-1.nat.csx.cam.ac.uk] has joined #shogun09:53
@iglesiasghushell: I think the version rescale_outputs(SGVector& outputs) looks good09:55
@iglesiasghushell: why would you need outputs and new_outputs?09:55
hushellFernando, I tried a bit, if passing SGVector, the content changed inside function will bring out, but if you just copy from another vector, like this a = b, then it won't bring out the change09:56
@iglesiasghushell: aham I see09:57
hushellI think have another copy looks safer, I tried SGVector::resize doesn't work, but maybe the resize_vector(&) would work09:58
@iglesiasgI agree with sonne|work, use resize_vector looks the good option09:58
@iglesiasgbtw I did a small test09:59
@iglesiasghttp://pastebin.com/4GcHbjkD09:59
@iglesiasgwhat I said before was wrong, it *must* be done with SGVector&09:59
hushellThat's works for me! Thanks! So I can kill one parameter.10:03
@iglesiasghushell: btw, I read in your mail you have been using SO learning in CV, right?10:08
hushellYep, SO becomes a mainstream in CV, especially the DPM, very popular10:09
hushellWhat are you working on?10:10
@iglesiasghushell: nice, so you are using it for something similar to facial landmark detection or so?10:10
hushellYeah, everything can divide into parts can be learned by SO10:11
-!- cameron__ [~quassel@global-2-1.nat.csx.cam.ac.uk] has quit [Ping timeout: 256 seconds]10:12
@iglesiasghushell: I am writing my final degree project (undergrad thesis) on SSVMs applied to graphs10:13
hushellI used the facial landmark code recent, it's like a variant of the deformable part model, with different loss function and use supervision instead of use latent positives10:13
hushellThat's awesome! I was doing something silly for my undergrad thesis. Are you going to apply a PhD?10:16
@iglesiasghushell: for sure :)10:17
hushellMachine learning?10:17
@iglesiasghushell: I am into machine learning and autonomous robots10:17
@iglesiasgI will see what I can find around in those fields and choose what I like the most10:18
-!- cameron__ [~quassel@global-2-11.nat.csx.cam.ac.uk] has joined #shogun10:19
hushellIf you like more math, machine learning will be a better choice in my opinion10:19
@iglesiasgaham, I will take it into account10:20
hushellGood luck for your application!10:20
@iglesiasghushell: thanks!10:20
@iglesiasgI am actually trying to find something a little bit simple yet fancy to show during the project presentation10:20
@iglesiasgsomething from the CV domain10:21
hushellHow about image segmentation or denoising?10:21
@iglesiasghushell: those were the choices I was thinking of10:22
@iglesiasghushell: I tried fg/bg segmentation but I failed :S10:22
hushellI am going to include an example to Shogun, if I got accepted this year10:22
hushelltry to use this inference lib http://www.di.ens.fr/~mschmidt/Software/UGM.html10:23
hushellyou can switch the learning by SSVM10:23
@iglesiasghushell: I see, so you have previously done fg/bg segmentation with SSVMs?10:23
@iglesiasglet me see10:23
@iglesiasgso the idea would be to switch the parameter estimation part for the SSVM10:25
hushellI have some dirty code, maybe helpful for you https://code.google.com/p/avatol/source/browse/trunk/nematocyst/spinesproject/findMVC_UGM.m10:25
hushellThe SSVM I am using is actually calling the libqp10:25
hushellMaybe this is not the best way, if you can include the loss into the node potential in the inference, the results will be better10:28
@iglesiasgaham10:30
@iglesiasghushell: what sort of features do they use for this?10:30
@iglesiasgI have seen some HOG around10:30
hushellyep, I divide the image into blocks, each block extracted HOG10:31
hushellFor edge feature, just counts of y_i == y_j and y_i != y_j, so it's Potts model10:32
@iglesiasghushell: I am going to take a closer look at the code and will ask you again if you don't mind ;)10:34
hushellSure. I am glad if I can help10:35
hushellAre you going to apply the GSOC this year?10:35
@iglesiasgyeah10:36
@iglesiasghushell: I have applied for SO learning too and for LMNN10:36
@iglesiasgbut my main objective is LMNN10:37
-!- thoralf [~thoralf@enki.zib.de] has joined #shogun10:37
thoralfHey.10:37
@iglesiasghey thoralf10:39
hushellHaha, don't worry I will also be happy if you get accepted to SO learning10:40
hushellThe LMNN is a research topic of my college, you may like to talk to him10:40
@wikingmorning10:55
-!- mizobe [~mizobe@189.35.244.197] has quit [Remote host closed the connection]10:55
@wikingHeikoS: ping10:56
hushell\quit10:57
@wikinghushell: /quit ;)10:57
hushell:p10:57
-!- hushell [43bd6474@gateway/web/freenode/ip.67.189.100.116] has quit []10:57
-!- vgorbati [d4029f22@gateway/web/freenode/ip.212.2.159.34] has quit [Ping timeout: 245 seconds]10:59
-!- vgorbati [d4029f22@gateway/web/freenode/ip.212.2.159.34] has joined #shogun11:11
@HeikoSwiking: pong11:12
@wikingHeikoS: get_mean_vector11:14
@HeikoS?11:14
@wikingah ok11:14
@wikingit's not rdn11:14
@wikingrnd11:14
@wikingit's based on the model11:15
@HeikoSwiking: no idea what you are talking about :D11:15
@HeikoSexplain11:15
@wikinganyways i was looking into RNG11:15
@HeikoSI see11:15
@wikingand now i'm trying to figure out where the changes are going to be11:15
@HeikoScool what did you find11:15
@wikingas i thought to use11:15
@wikingor create an RND11:15
@wikingRNG11:15
@HeikoSwhat about sergey's c++11 idea?11:16
@wikingthat can give you back like a SGVector with samples of a standard distrib11:16
@wikinglike the new c++1111:16
@wikingc++11 is great11:16
@wikingbut what if the user doesn't have a compiler yet11:16
@wikingthat supports c++1111:16
sonne|workwiking: changes would just be in CMath::rand* and init_random11:16
@wikingHeikoS: std::normal_distribution<double> normal_dist(mean, stddeviation);  //11:17
-!- van51 [~van51@athedsl-401908.home.otenet.gr] has joined #shogun11:17
@wikingHeikoS: it has a good api11:17
@wikingonly thing that we cannot say that we only support compilers that supports c++1111:17
@HeikoSwiking yes, we need the random normals and the uniform ones, I think we dont have more distributions right?11:17
@wikingsonne|work: i've checked a comparison between R's, c++11's and boost's RNG, and c++11 was the fastest11:18
@HeikoSwhat about something as libssl then?11:18
@wikingHeikoS: yeah i think that 2 distrib would be sufficient at the moment11:18
@wikingHeikoS: well it's an option, but then again sonne|work disagrees11:18
@HeikoSok11:18
@HeikoSbut boost?11:18
@HeikoSthats massive11:18
@wikingboost is bloated11:18
@wikingthis RNG is meant to be really fast: http://en.wikipedia.org/wiki/Ziggurat_algorithm11:19
@wikingthe code for it is like 100 lines in c11:20
@wikingwhat we could do is to have 2 standard engines for RNG11:20
@wikingZiggurat algorithm and Mersenne Twister11:20
@wikingboth of them are pretty standard for PRNG11:21
@wikingand then one can choose, but would say default to Ziggurat the PRNG as that's faster than mersenne twister11:21
@wikingand then we are pretty much done... :P11:22
@HeikoSokay11:22
@HeikoSquestion:11:22
@HeikoSwill we still have problems with the tests?11:22
@wikingboth of their implementation is available in various forms ...11:22
@wikingHeikoS: well if we use our own PRNG then no11:22
@HeikoSsince the problem is that the streams are not reproducable11:22
@HeikoSbut why have two then?11:22
@HeikoSalso, btw for us, speed of rng does not really matter that much11:23
@HeikoSso why not just go with one?11:23
@wikingwell then let's just decide which PRNG we would like to use11:23
@wikingand that's it11:23
@HeikoSwiking or actually11:23
@HeikoSwe can add more11:23
@HeikoSand then let user decide which one to use in his code11:24
@HeikoSbut std one is fixed11:24
@HeikoSso that the examples all work11:24
@wikingHeikoS: well as said11:24
@wikingwe have a default PRNG engine11:24
sonne|workwiking: did you measure speed also for opencv's RNG?11:24
@HeikoSthis is the way its done for example in matlab, you can choose your stream and seed it11:24
@wikingbut we provide more...11:24
@wikingsonne|work: they use Mersenne11:24
sonne|workdidn't they have multiple?11:24
@lisitsynI woke up! :D11:25
@wikingsonne|work: the code is imo too much bundled into opencv11:25
@lisitsynspeed of RNG - does that matter?11:25
@wikingsonne|work: so i'd rather not even touch opencv's RNG11:25
sonne|worklisitsyn: yes of course11:26
@HeikoSsonne|work: why does it matter?11:26
@HeikoSfor shogun I mean11:26
sonne|workdo you want to wait 1 hour to get a 1000x1000 rnd matrix?11:26
-!- vgorbati [d4029f22@gateway/web/freenode/ip.212.2.159.34] has quit [Ping timeout: 245 seconds]11:26
@lisitsynwell that's not possible to write that bad RNG :D11:26
sonne|workI used it in qsort for the pivot element11:26
@HeikoSI think the speed differences wrt to matrices that you can store in memory (small) is minor, but I might be wrong11:27
sonne|workbut sure one can always work around this11:27
-!- iglesiasg [~nando@s83-179-44-135.cust.tele2.se] has quit [Ping timeout: 264 seconds]11:27
sonne|worklisitsyn: just compare /dev/random and /dev/urandom11:27
@lisitsynsonne|work: yeah I know they are different in perf11:28
@lisitsynbut I mean it is not that usual thing to compute random values11:28
@lisitsynit usually happens when you either initialize random data (but that's just test) or initialize something in iterative algorithm11:28
-!- vgorbati [d4029f22@gateway/web/freenode/ip.212.2.159.34] has joined #shogun11:29
@wikinganyhow11:30
@wikingthe implementations you can find11:30
@wikingare like 100 lines11:30
@wikingsee http://www.bedaux.net/mtrand/mtrand.cpp.html for mersenne twister11:31
-!- shogun-notifier- [~irker@7nn.de] has quit [Quit: transmission timeout]11:31
@wikingafaik these are overoptimized already11:31
@wikingas the point of these various PRNG is that who can create/define a faster PRNG :P11:31
sonne|worklisitsyn: well I don't mind having a slow but very accurate prng and a fast one for the other case11:31
@wikingwhat's very accurate prng?11:32
@wikingthe things one usually measures of a prng is speed and the entropy of the output....11:33
-!- vgorbati [d4029f22@gateway/web/freenode/ip.212.2.159.34] has quit [Ping timeout: 245 seconds]11:33
sonne|workwiking: in limit hist(x) will match the wanted distribution p(x)11:33
sonne|worklike normal distribution, or equally distributed11:33
@wikingsonne|work: well PRNG is simply a random generator11:33
@wikingso the rest should be done by us11:33
@wiking;)11:33
sonne|workHeikoS: could you please click the 'I want to mentor knob'!11:34
@wikingPRNG will just assures us that what we get is actually something like a white noise11:34
sonne|workwiking: for uniform I guess?11:34
@HeikoSsonne|work: will do!11:34
@wikingsonne|work: white noise :)11:34
@wikingnothing else11:34
sonne|workHeikoS: done?11:35
@wikingand if we want to have different distribs11:35
@wikingthen we have to create our own methods for it11:35
@wikingbased on PRNG11:35
@wikingso yeah indeed will get a uniform distrib with a good PRNG11:35
@wikinguniform distrib that actually reproducible11:36
@wikingover different os/arch etc11:36
@wikingbased on the seed11:36
@HeikoSdonre11:36
@wikingso11:36
@wikinglet's have a CRNG class ;)11:36
@wikingor rather CPRNG11:37
@wikingshogun/lib/PRNG.h/cpp11:37
@wikingall agree? :)11:37
@wikinggood11:37
@lisitsynguyz guyz11:38
@lisitsynI kan haz a betta soluktion11:38
@lisitsynwe haz to criate a madule shogun-rendom11:38
@wikinglisitsyn: doz it!11:38
@lisitsyn& putz a few gigobytez of rendom veluez11:39
@wikinglol11:39
@wiking:>11:39
@wikinglisitsyn: had too much vodka yesterday when u celebrated the day of work? :)11:39
sonne|workwiking: shogun/math/Random.{h,cpp}11:39
sonne|workwiking: lisitsyn is always like this11:39
@lisitsynhaha11:40
sonne|workyou know Russian ;)11:40
@wikingok let's gitflow this one as well11:40
@wiking:)11:40
@lisitsynpeace! vodka! putin!11:40
@wikinglisitsyn: hasn't putin gave u the medal for the hero of work yesterday? :D11:41
@lisitsynwiking: no I am a foreign agent you know11:41
@lisitsynreceived money from the enemy two times11:41
@lisitsynno pasaran!11:41
@wikingsonne|work: you've ment shogun/mathematics/Random.{h,cpp} right?11:41
sonne|workwiking: yes11:41
-!- shogun-notifier- [~irker@7nn.de] has joined #shogun11:43
shogun-notifier-shogun: iglesias :develop * 100d655 / src/shogun/ (4 files): https://github.com/shogun-toolbox/shogun/commit/100d655bf89fdbd135dd8119a506f869e562674b11:43
shogun-notifier-shogun: Drop static SGVector::resize().11:43
shogun-notifier-shogun: This method can be misleading since it should not be used for SGVectors,11:43
shogun-notifier-shogun: use SGVector::resize_vector() instead.11:43
shogun-notifier-shogun: Soeren Sonnenburg :develop * 088dc29 / src/shogun/ (4 files): https://github.com/shogun-toolbox/shogun/commit/088dc299c9969da210290319afb28d23e4ec70bd11:43
shogun-notifier-shogun: Merge pull request #1040 from iglesias/feature/drop_resize11:43
shogun-notifier-shogun:11:43
shogun-notifier-shogun: Drop static SGVector::resize().11:43
@wikingdoh12:00
@wikingwhy i cannot add parameter to a class that's derived directly from SGObject? :)12:00
@wikinghttp://pastebin.com/ZVS2s7nR12:01
@lisitsyn21 proposals so far!12:03
shogun-buildbotbuild #731 of cyg1 - libshogun is complete: Success [build successful]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/cyg1%20-%20libshogun/builds/73112:05
@wikinglol12:15
@wikingi've just realised12:15
@wikinghttp://www.math.sci.hiroshima-u.ac.jp/~m-mat/MT/SFMT/12:15
@wikingso i think we should go with this12:15
@wikingas this is a standard mersenne twister12:15
@wikingbut if sse2 or alitvec is available12:15
@wikingit can use it in order to boost up the speed12:15
@wikingsonne|work: ^ ?12:18
-!- poojits [75d35d4a@gateway/web/freenode/ip.117.211.93.74] has joined #shogun12:42
poojitsI want to contribute to Shogun, so I looked at the entrance tasks and I want to give issue #1025 a shot. I don't know if it's a bad thing, but I feel like I need to understand the entire codebase, so I know exactly what I am doing. How do I start?12:44
-!- vgorbati [d4029f22@gateway/web/freenode/ip.212.2.159.34] has joined #shogun12:44
sonne|workwiking: ohh a maintained thing very good!12:57
sonne|workwebsite from 1995 but updated 2013 - so yes why not. give it a try...12:58
@wikingsonne|work: the only question now of course13:00
@wikingwhether we want SFMT or dSFMT13:01
@wikingand that or is not necessarily exclusive or ;)13:01
@wikingas sfmt can create us fast int32/64 arrays13:02
@wikingas dsfmt can do the same with double13:02
sonne|workboth then :)13:02
vgorbatilisitsyn: here?13:02
@lisitsynvgorbati: yes13:02
@wikinghehehe yeah i guessed so13:02
@wikingsonne|work: there's only one stupid problem here with this lib13:03
sonne|work?13:03
@wikingThis function generates and returns 64-bit pseudorandom number. * init_gen_rand or init_by_array must be called before this function. * The function gen_rand64 should not be called after gen_rand32, * unless an initialization is again executed.13:03
@wikingso basically i'll have to maintain a state13:03
@wikingin the RNG13:03
@wikingand reinit the struct if genrand32 or the other has been called13:03
sonne|workwiking: you have to add locking13:04
-!- iglesiasg [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun13:04
-!- mode/#shogun [+o iglesiasg] by ChanServ13:04
sonne|workmaybe expose 2 functions then13:04
sonne|workone thread safe (where one has to input the state)13:04
sonne|workand one where the state is maintained internally13:05
@wikingah fuck yeah13:05
@wikingi havent' thought about being thread safe13:05
vgorbatilisitsyn: about multidimensional scaling: the method description looks quite understandable, while implementation does not:) Theoretically, the only method parameters should be num_neighbors and squishing_rate, am I right?13:06
@lisitsynvgorbati: multidimensional scaling? it doesn't have any parameters at all13:06
vgorbatilisitsyn: wait, what?:) At least K - the number of neighbors13:07
@iglesiasgpoojits: hey, I think that issue is already been taken care of by another person but you are free to do it if you want13:07
@lisitsynno, MDS has no deals with neighbors - that's isomap13:07
sonne|workwiking: woahh!13:07
sonne|workrand() is also not thread safe13:07
sonne|workI guess that could explain some issues we were having13:07
vgorbatilisitsyn: oh, my fault - manifold sculpting:D13:07
@lisitsynvgorbati: ahh13:07
-!- cameron__ [~quassel@global-2-11.nat.csx.cam.ac.uk] has quit [Ping timeout: 245 seconds]13:08
vgorbatilisitsyn: at least that's what the original paper says..13:09
@wikingsonne|work: heheh ok i'll make it thread safe13:09
@wikingjust first i want to import then both of the libraries13:09
vgorbatilisitsyn: but since it is iterational, I am thinking whether I should use a max_iteration parameter as well..13:10
@lisitsynvgorbati: yes why not13:10
@lisitsynit always makes sense to limit # of iterations13:10
@wikingoooh shit13:11
@wikingredefinitions :S13:11
@lisitsynvgorbati: I wish there was a matlab implementation13:12
vgorbatilisitsyn: ok, one question solved:) Now another one:since it is also gradient-descend-like, it uses learning rate - but in the implementation, Gashler sets it to average distance to neighbors and then during each iterations multiplies it by magic numbers:)13:13
@lisitsynvgorbati: well he knows better - he is the author  :D13:13
vgorbatilisitsyn: So, leave it as it is?)13:14
@lisitsynyeah for now13:14
vgorbatilisitsyn: the algorithm is actually full of these numbers:) about matlab - that would be awesome)13:16
vgorbatilisitsyn: but, well, I have what I have)13:16
@lisitsynvgorbati: I googled it a little but haven't found anything13:17
-!- mikhailBelous [~quassel@37.98.242.99] has joined #shogun13:19
vgorbatilisitsyn: now 3rd question: as I see it now, in embedManifoldSculpting method I should get the neighbors using find_neighbors method, then find 2 SparseWeightMatrix - one for distances and one for angles, and then actually call the method passing these matrices and parameters into. Does it make sense?13:19
@lisitsynvgorbati: let me check teh paper13:20
vgorbatihttp://www.cc.gatech.edu/~isbell/reading/papers/ManifoldSculpting.pdf13:20
mikhailBelousI sent my aplication yesterday somebody please watch it13:21
@lisitsynmikhailBelous: yeah we have seen13:22
mikhailBelousDid you like it?13:23
mikhailBelousAny comments,question suggestions?13:23
@lisitsynmikhailBelous: alright, first - it should be one proposal per idea so if you are applying to two ideas (with two different mentors) - please send two proposal13:26
@lisitsyns13:26
@lisitsynnext, please read it carefully13:27
mikhailBelousIdeas?13:27
-!- hoijui [~hoijui@dslb-088-074-121-064.pools.arcor-ip.net] has joined #shogun13:28
@wikingsonne|work: ok question: the CMath::random* function are static... while now i'm creating a CRandom class... hence we need to statically allocate that CRandom object somewhere... like in init_shogun() ?13:28
@lisitsynmikhailBelous: if you are applying to LMNN or dimension reduction - send two proposals - one for LMNN and one for DR13:28
sonne|workwiking: ok13:29
@wikingsonne|work: ah i see init_shogun creates a CMath class... then we can actually make CRandom part of CMath, or?13:29
mikhailBelousYear I got it.13:29
mikhailBelousWhat I have to read carefully?13:29
@wikingsonne|work: i mean to be a private variable of CMath13:29
@wikingalthough even then the reference on CRandom will be needed, i.e. like a global or something13:30
sonne|workwiking: maybe even make it a separate class13:31
@lisitsynmikhailBelous: your proposal, please check it for spelling and grammar13:31
@wikingsonne|work: it'll be a separate class...13:31
@wikingCRandome13:31
@wiking*CRandom13:32
@lisitsynwiking: french class13:32
@lisitsyn'CRandome'13:32
@lisitsyn:D13:32
mikhailBelousOkay.Sorry I'm enlgish is not my native.13:32
@lisitsynthat's ok it is not native for anyone here just read it carefully13:33
mikhailBelousMay I copypaste my proposal?13:36
@lisitsynwell except to some parts they can be similar13:36
-!- sids_aquarius [~sidi_@li400-124.members.linode.com] has joined #shogun13:39
@lisitsyn2313:43
@lisitsynsonne|work: I've told you soo!13:43
@lisitsyn2413:45
mikhailBelousAnd how about me am I good for this project13:45
mikhailBelous?13:45
votjakovrHi everyone! I have just submitted my proposal:)13:45
@lisitsynvotjakovr: yeah I have seen13:45
@lisitsynmikhailBelous: we are unable to say whether any student is going to be accepted or not - that's going to happen on 27 May13:46
@wikingsonne|work: any idea why i cannot do SG_ADD (see: http://pastebin.com/ZVS2s7nR13:48
mikhailBelousOki-doki. I'm just intrested did you like it or not13:49
@lisitsynmikhailBelous: you have got no contributions so far - that should be the top priority now13:50
-!- cameron__ [~quassel@global-2-11.nat.csx.cam.ac.uk] has joined #shogun13:53
-!- poojits [75d35d4a@gateway/web/freenode/ip.117.211.93.74] has quit [Ping timeout: 245 seconds]14:14
-!- shogun-notifier- [~irker@7nn.de] has quit [Quit: transmission timeout]14:43
-!- cameron__ [~quassel@global-2-11.nat.csx.cam.ac.uk] has quit [Ping timeout: 248 seconds]14:57
-!- votjakovr [~votjakovr@host-46-241-3-209.bbcustomer.zsttk.net] has quit [Quit: Leaving]15:03
-!- erlenda [~erlenda@fw-oslo.intelcom.no] has quit [Remote host closed the connection]15:03
-!- sids_aquarius [~sidi_@li400-124.members.linode.com] has quit [Ping timeout: 260 seconds]15:06
-!- wencan [~wencan@c-71-61-182-121.hsd1.pa.comcast.net] has joined #shogun15:27
-!- thoralf [~thoralf@enki.zib.de] has quit [Quit: Konversation terminated!]15:28
-!- vgorbati [d4029f22@gateway/web/freenode/ip.212.2.159.34] has quit [Ping timeout: 245 seconds]15:28
-!- erlenda [~erlenda@fw-oslo.intelcom.no] has joined #shogun15:30
-!- thoralf [~thoralf@enki.zib.de] has joined #shogun15:40
thoralfHey.15:40
@HeikoSthoralf: hi!15:40
@HeikoSjust about going into the bug15:40
@HeikoSsorry for the delay, gsoc eats a lot of time15:40
@wikingsonne|work: ping ping ping15:41
thoralfHeiko: Tell me it's a bug, so I know that the solution is more than not beeing stupid. ;)15:42
sonne|workwiking: you have to include lib/parameter.h IIRC15:43
@HeikoSthoralf: have to go into it :)15:43
@wikingsonne|work: we need a macro HAVE_SSE215:43
@wikingin case there's sse215:43
@wikingi've already made some changes15:43
sonne|workwiking: lets try first w/o sse15:43
sonne|workwe can speed things up later15:43
@wikingkk i'm just doing the configure now15:44
sonne|workHeikoS: ping ping ping - please assign yourself as mentor to students you intend to mentor15:45
@HeikoSsonne|work done15:46
@HeikoSis this possible mentoring or definite vote?15:46
sonne|workwe can change until may 22 IIRC15:47
sonne|workor 24 even15:47
@wikingsonne|work: ok it's ready15:48
@wikingsonne|work: i'll push the feature branch to the repo15:48
@wikingand comment it15:48
@HeikoSsonne|work: question is: do I have to say yes to multiple people?15:48
@wikingit's purely the import of SFMT and adding a new CRandom class15:49
sonne|workyou can say want to mentor to multiple people yes15:49
-!- cameron__ [~quassel@global-2-11.nat.csx.cam.ac.uk] has joined #shogun15:49
sonne|workI will assign you only 1 per task for sure :)15:49
-!- iglesiasg [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving]15:49
@wikingthe static allocation of the object in init_shogun is not done yet15:49
@HeikoSsonne|work: I am fine with that :D15:49
@wikingnor that CMath::random is using the new PRNG15:50
@HeikoSsonne|work: if people send PRs against shogun-data, they can only do it to "master", but thats fine right? The data version of the develop/master repository selects the correct one15:51
sonne|work11. What happens if two students are accepted to work on the same project, e.g. from an organization's Ideas list?15:52
sonne|workThat's fine, a little duplication is par for the course in open source.15:52
sonne|workohh look at this15:52
sonne|workone could even have two students working on the same task15:52
sonne|workI didn't know15:52
sonne|workHeikoS: yes that's ok15:52
@HeikoSlisitsyn: ^15:52
@HeikoSok15:52
sonne|workwiking: then do the static analysis15:53
@wikingssh: connect to host github.com port 22: Connection refused15:53
@wikingLOL15:53
@wikingwtf is this15:53
@wikingok now it went ok15:53
-!- shogun-notifier- [~irker@7nn.de] has joined #shogun15:53
shogun-notifier-shogun: Viktor Gal :feature/PRNG * 35b7254 / src/shogun/lib/external/SFMT/ (16 files): https://github.com/shogun-toolbox/shogun/commit/35b7254d4e861b78a24ae35323450a9071b55d9915:53
shogun-notifier-shogun: Import SFMT PRNG15:53
shogun-notifier-shogun: Viktor Gal :feature/PRNG * 31da88a / src/shogun/mathematics/Random.cpp,src/shogun/mathematics/Random.h: https://github.com/shogun-toolbox/shogun/commit/31da88aac065398f6aaf0ad0a0ff897fbb72c15115:53
shogun-notifier-shogun: Create class for PRNG15:53
shogun-notifier-shogun: wrapper class for SFMT15:53
shogun-notifier-shogun: Viktor Gal :feature/PRNG * 81599e1 / src/configure: https://github.com/shogun-toolbox/shogun/commit/81599e1c93ab83d54c1599697990dfd9c713d5fb15:53
shogun-notifier-shogun: Add --mexp and HAVE_SSE2 defines in ./configure15:53
@wikingthere you go comment it15:53
thoralfsonne|work: AFAIK it must be clear, which contribution is attributed to which student.  Beside of this,15:54
thoralfsonne|work: i.e. no "team work"15:54
sonne|workwiking: what is +SFMT_MEXP=1993715:54
sonne|workthoralf: yes15:54
@wikingsonne|work: mersenne exponent15:55
sonne|workbut ML is soo wide spread basically any idea we have on the list can be diveded into multple15:55
@wikingsonne|work: it's required by SFMT15:55
sonne|workwiking: so why is this in configure then?15:55
sonne|workor which one should one use15:55
@wikingsonne|work: because one can use other exponents15:56
@wikingsee ./configure --help15:56
sonne|workhow does it affect randomness15:56
sonne|workyeah seen it15:56
@wikingSFMT_MEXP This macro is required. This macro means Mersenne exponent and the period of generated code will be 2SFMT_MEXP-1. SFMT_MEXP must be one of 607, 1279, 2281, 4253, 11213, 19937, 44497, 86243, 132049, 216091.15:56
@wiking2^SFMT_MEXP-115:57
@wikingto be correct with the above copy paste15:57
@wikingit's a prime i know that but how that effects randomness15:57
@wikingno clue15:57
@wikingone should read their paper15:57
sonne|workthe period of way over 1e5000 is certainly good enough for my life time15:58
sonne|workyeah it is kind of weird. I guess it will affect speed or sth15:58
@wikinglet's stick to the default15:58
@wikingbut since it's a parameter15:59
@wikingi reckon one should have access to change it15:59
sonne|workWe have no clue what it does.15:59
sonne|workChanging it is dangerous aka might kill properties of the RN and I would rather not15:59
@wikingsonne|work: ok then you'll fork it16:00
sonne|workbut hey you did the work so whatever keep it16:00
@wikingi strongly disagree with you hence it's up to you what you cherry-pick from the commmits16:00
sonne|workwiking: ?16:00
@wikingsonne|work: ok then we agree ;)16:00
@wikingthat we disagree16:00
@wiking:>16:00
@wikingmmm16:01
sonne|workparse error16:01
@wikingwhere?16:01
sonne|workI don't understand what you are saying16:01
@wikingbtw how do you SG_ADD a pthread?16:01
sonne|workwhere do we agree / disagree16:01
@wikingpthread_mutex16:01
sonne|workdon't16:02
@wikingah ok16:02
@wikingany enum ?16:02
@wikingand (void *) ?16:02
sonne|workvoid* -> impossible16:03
sonne|workenum IIRC machine_int_t16:03
@wikingsfmt_t* m_sfmt;16:03
@wikingas i have this16:03
@wikingin the class16:03
sonne|workSG_ADD((machine_int_t*) &my_enum,...)16:04
sonne|worksfmt* is what?16:04
sonne|workwiking: please drop the sfmt_altivec file16:05
@wikingsonne|work: well that's an internal struct of the SFMT16:05
sonne|worklets just keep the generic and the sse variant16:05
@wikingsonne|work: eheh that will take some fair amount of digging :P16:05
@wikingsonne|work: to get out altivec from that code16:05
@wikingit's quite bundled16:05
@wikingbut if we never set the HAVE_ALTIVEC macro16:05
@wikingas we do right now16:05
@wikingthen we are on the safe side16:06
sonne|workwiking: no - it is just included once16:06
sonne|workso just remove the file16:06
sonne|work+#if defined(HAVE_ALTIVEC)16:06
sonne|work+  #include "SFMT-alti.h"16:06
sonne|work+#elif defined(HAVE_SSE2)16:06
sonne|work+  #include "SFMT-sse2.h"16:06
sonne|work+#endif16:06
sonne|workand remove it from there too16:06
@wikingsonne|work: you can commit to this branch as well ;)16:06
sonne|workno time16:06
@wikingok so how should i sg_add that sfmt*16:06
@wiking?16:06
sonne|workno that cannot work if not derived from SGObject16:07
@wikingmmm16:07
@wikingbut that'll cause troubles with python_modular16:07
@wikingor?16:07
@wikingor since this is not going to be directly used from modular interfaces16:07
@wikingwe are good?16:08
@HeikoSthoralf: so yes you found a bug16:08
sonne|workwell in principle we don't need any serialization for the RNG at all16:08
@HeikoSI was aware that something in the streaming framework is wrong but this clearly points it out16:08
sonne|worksince we set the seed anyway16:08
@HeikoSin fact, there already occur problems without loading any files16:08
sonne|workohh it would even hurt16:08
sonne|workwe would then need to set the seed after de-serializing16:09
sonne|workwiking: I am rather in favor of not making the object serializable16:09
thoralfHeikoS: Okay, thanks.16:09
@HeikoSthoralf: I will have a look now and put up an issue on git16:10
@HeikoSsomething I wanted to have fixed for a while, but did not find the time yet16:10
sonne|workHeikoS: what is the bug16:10
@wikingsonne|work: ok any special thing i should do to disable serialization?16:10
sonne|workjust remove SG_ADD's16:10
@wikingsonne|work: all of them? :) okey16:10
@HeikoSsonne|work: streaming features have memory problems: double free an other things16:10
@HeikoScurrently trying to find out more16:11
sonne|workHeikoS: yes I know - but is this the SGVector thing?16:11
sonne|workwiking: then please do the initialization in init_shogun16:11
sonne|workand then run generator.py again16:11
sonne|workand check if you manage to get the same results on *bsd and linux...16:12
@wikingok16:12
@HeikoSthoralf: does it currently prevent you from working?16:12
@HeikoSsonne|work: not sure what you mean16:12
shogun-notifier-shogun: Viktor Gal :feature/PRNG * 4736671 / src/shogun/mathematics/Random.cpp,src/shogun/mathematics/Random.h: https://github.com/shogun-toolbox/shogun/commit/473667152bc96e8bbbe2f1438b997ac918d7897f16:12
shogun-notifier-shogun: Add parameters of Random and add HAVE_PTHREAD macro check16:12
shogun-notifier-shogun: Viktor Gal :feature/PRNG * 7dd6b47 / src/shogun/mathematics/Random.cpp,src/shogun/mathematics/Random.h: https://github.com/shogun-toolbox/shogun/commit/7dd6b47b92a73cf282135a6a936a060f622e0b1216:12
shogun-notifier-shogun: Make Random unserializable16:12
thoralfHeikoS: No, as I've got plenty of work.  Yes, as it blocks my shogun experiments. :)16:13
@HeikoSthoralf: I mean, a quick hack for now is just to remove the SG_UNREF16:14
thoralfHeikoS: Yes, but then it does not free the memory - and my data files are quite big.16:15
@HeikoSthoralf: I see16:16
@HeikoSthoralf: I will explore and discuss with others, we are definitely interested in having this resolved soon16:16
@HeikoSthoralf:  just wondering what do you want to do?16:16
sonne|workHeikoS: recall that we switched to using SGVector at some point?16:16
@HeikoSsonne|work: yes16:16
sonne|workHeikoS: the streaming features were written when that stuff did not exist16:17
sonne|workso now we have the issue that SGVector auto-frees memory16:17
@HeikoSsonne|work: so this still has to be migrated, right, but I think last time I went into this, there were some more problems16:17
sonne|work(or ot)16:17
@HeikoSsonne|work: should be easy to fix then16:17
thoralfHeikoS: Ive got some external svmlight file and try to experiment with different hash kernels (MKL) on this data.16:17
sonne|workno it is not an easy fix :/16:17
thoralfHeikoS: And MKL lead me to shogun.16:17
thoralfsonne|work: Sounds bad.  If you need some assistance, let me know.16:18
thoralfsonne|work: I think the bug is easy to track down, since it already happens with a small one-vector-data set.  So it would be possible to track the reference counts "by foot". :)16:20
@HeikoSthoralf: would you be interested in working on this? :)16:22
-!- erlenda [~erlenda@fw-oslo.intelcom.no] has quit [Quit: Leaving]16:23
shogun-notifier-shogun: Viktor Gal :feature/PRNG * c204e89 / src/shogun/base/init.cpp,src/shogun/base/init.h: https://github.com/shogun-toolbox/shogun/commit/c204e89ad4828227c7c648bd0d3d5fc4903ebc1e16:23
shogun-notifier-shogun: Create a global Random object in init16:23
thoralfHeikoS: Damn. I was begging for this, right? ;)16:24
@wikingsonne|work: ^ should be like this right?16:24
thoralfHeikoS: I'll have a look.16:24
sonne|workwiking: yes16:24
@wikingok now it's really only about starting to use this PRNG in shogun ;P16:25
@HeikoSthoralf: hehe :)16:25
sonne|workwiking: yes just run generator.py16:25
@HeikoSthoralf: would be nice if you have time, we are postponing this one for a while now :) let me know if you need any help/questions/answers16:26
sonne|workand then rsync this stuff to some *bsd machine and try if it gives the same result16:26
sonne|workwiking: and tester.py of course too :)16:26
@wikingsonne|work: mmm how to i get reference on that sg_rand ? :)16:27
sonne|workwiking: did you add an #extern ?16:28
sonne|workextern actually16:28
@wikingwhere?16:28
sonne|workwiking: whereever you need it16:32
sonne|workI don't know why you need it actually16:32
sonne|workyou could just use the getter16:33
@wikingyep16:34
@wikingbut i wonder why i cannot SG_UNREF(CRandom*)16:34
@wiking/Users/wiking/shogun/src/shogun/mathematics/Math.h:476,4 - Error - member access into incomplete type 'shogun::CRandom'16:34
sonne|workI don't have time to look into this right now :/16:35
@wikingok done16:35
@wikingwill work now16:36
@wikingnow it's 'only' rewriting all the functions in CMath16:36
-!- travis-ci [~travis-ci@ec2-50-16-143-238.compute-1.amazonaws.com] has joined #shogun16:43
travis-ci[travis-ci] it's Viktor Gal's turn to pay the next round of drinks for the massacre he caused in shogun-toolbox/shogun: http://travis-ci.org/shogun-toolbox/shogun/builds/682017616:43
-!- travis-ci [~travis-ci@ec2-50-16-143-238.compute-1.amazonaws.com] has left #shogun []16:43
-!- vgorbati [d4029f22@gateway/web/freenode/ip.212.2.159.34] has joined #shogun16:46
-!- gsomix_ [~gsomix@95.67.166.72] has joined #shogun16:47
-!- van51 [~van51@athedsl-401908.home.otenet.gr] has left #shogun ["QUIT :Leaving."]16:49
-!- gsomix [~gsomix@95.67.181.139] has quit [Ping timeout: 264 seconds]16:50
-!- KMcQuisten [d8338942@gateway/web/freenode/ip.216.51.137.66] has joined #shogun16:58
KMcQuistenMorning, all16:58
@HeikoSKMcQuisten:  hi!16:58
KMcQuistenHope all is well16:58
@wikinguff why :)16:58
@HeikoSHope for you too!16:58
KMcQuistenWhy not?16:58
@HeikoSabout your mail16:59
@wikingmmm i was commenting to travis-ci16:59
KMcQuistenOh, ok16:59
@HeikoSso this is an issue that the model selection framework has16:59
@HeikoSif you wanted to do a search over multiple combined kernels16:59
-!- Asix3 [~Asix3@d47-69-144-148.nap.wideopenwest.com] has joined #shogun16:59
@HeikoSyou would have to create all of them by hand16:59
@HeikoSbut thats kind of not what you want16:59
KMcQuistenExactly16:59
@HeikoSso one thing I dont get is:16:59
@HeikoSyou do MKL right? so the program selects the subkernel weights for you17:00
@HeikoSwhy do you want to optimise them seperately?17:00
@HeikoSI would just throw in all kernels that might be relevant and look at the weights17:00
@HeikoSmodelselection only necessary for the SVM parameters17:00
@HeikoSor am I wrong?17:00
KMcQuistenI don't want to optimize the subkernel weights separately.  I want to set up specific subkernels and optimize the parameters for those kernels, while still letting the MKL optimize the subkernel weights17:01
@HeikoSI see17:01
@HeikoSso one thing you could do:17:01
@HeikoSsay you want to select a Gaussian width17:01
@HeikoSjust add subkernels with all those widths to the combined kernel17:01
KMcQuistenThat won't work17:01
KMcQuistenLet me explain:17:02
KMcQuistenI have two kernels, one for features related to one aspect of my data, and another for different features on the same data17:03
KMcQuistenIf I just threw all the possible parameterizations for both types of kernels into an MKL, there is no way for me to force it to choose one from each aspect17:03
-!- _Asix3 [~Asix3@d47-69-144-148.nap.wideopenwest.com] has joined #shogun17:03
-!- Asix3 [~Asix3@d47-69-144-148.nap.wideopenwest.com] has quit [Ping timeout: 256 seconds]17:03
@HeikoSKMcQuisten: I see17:04
KMcQuistenI want to be able to do a grid search of subkernel parameters17:04
@HeikoSKMcQuisten: ok understoof17:04
@HeikoSd17:04
KMcQuistenwhile letting the MKL optimize the subkernel weights17:04
@HeikoSunfortunately, thats currently not possible17:04
-!- _Asix3 [~Asix3@d47-69-144-148.nap.wideopenwest.com] has quit [Client Quit]17:04
@HeikoSsince model selection does not handle subkernel stuff straight forwardly17:04
@HeikoSits not a very big deal to extend it though17:05
@HeikoSone more question:17:05
KMcQuistenIndeed.  I was thinking that there would need to be a CombinedKernel parameter for "subkernel parameters"17:05
KMcQuistenOk17:05
KMcQuistenShoot17:05
@HeikoSI mean you can add SG_Objects llike the subkernels as parameters17:05
@HeikoSso the other thing is17:05
@HeikoSyou choose only kernels on feature part A17:06
@HeikoSput in all the subkernels17:06
@HeikoSuse MKL to select parameters17:06
@HeikoSor something else17:06
@HeikoSthen do the same thing for feature part B17:06
@HeikoSthen use the best one for MKL17:06
@HeikoSin fact you can use the existing modsel framework then17:06
KMcQuistenDoing that tends to overfit the model and is poor at generalizing.  I've tried that.17:06
@HeikoSok17:07
@HeikoSwell then, we have another item in our issues list :)17:07
@HeikoSInterested in working on that btw?17:09
KMcQuistenI was thinking about how it could work, and if the CombinedKernel had a parameter that's just called "subkernel", then nodes for each subkernel could be built and added as children to the CombinedKernel node, and it would be rather intuitive.  I've not gotten into the guts of the modsel framework to know if that kind of thing would be simple to implement17:10
@HeikoSthats exactly how we would do it17:10
KMcQuistenI model genetic regulatory pathways, so I use string and numerical kernels together quite a bit.17:10
@HeikoSah I see so even different features17:11
KMcQuistenExactly17:11
@HeikoSthats the difficulty with this combined stuff17:11
@HeikoSand also the sukernels are stored in a list17:11
KMcQuistenI'm interested in looking at tree-based features as well, as there are some natural ways they could apply to some of my work.17:12
@HeikoSwhich is not really registered for modelselection currently17:12
@HeikoSI would love to have tree/graph based features in shogun17:12
@HeikoSsomeone would have to code it up :)17:13
@HeikoSthe modelselection framework currently works like this:17:13
KMcQuistenTaht's a project i'd be interested in helping with once my dissertation's done17:13
@HeikoSclasses register their parameter available17:14
@HeikoSthen a user builds these trees17:14
@HeikoSand shogun computes the tree "products" with all combinations of parameters (that make sense)17:14
@HeikoSso we would just have to fit the combined kernel into this framework17:14
@HeikoSKMcQuisten: feel free to get back to use thenm17:14
@HeikoSI will add an issue on github and think on how to implement this stuff17:15
KMcQuistenIndeed.  I've used it for single kernel parameter optimization and it's worked great.17:15
@HeikoSfeel free to add your  comments to it, Ill send the link17:15
@HeikoSKMcQuisten: yes, thats a thing it can do quite nicely17:15
KMcQuistenIt's areally well-designed framework17:15
@HeikoSthanks!17:16
KMcQuistenThanks so much for looking into this for me.17:16
@HeikoSwe are currently looking into ways of making the syntax easier17:16
@HeikoSin particular from python17:16
@HeikoSKMcQuisten: welcome, its always nice to get feedback on things17:16
@HeikoSKMcQuisten: btw have you used the nu-svr stuff that I added with that huge delay?17:20
-!- hoijui [~hoijui@dslb-088-074-121-064.pools.arcor-ip.net] has quit [Quit: Leaving]17:21
KMcQuistenI have.  It's been a nice way to look at SVR, and for some reason it's been a bit more intuitive when I've explained how it works to my biologist collegues17:21
KMcQuistenI always thought the epsilon tube was perfectly understandable, but they like proportions, it seems.17:22
KMcQuistenThank you17:22
@HeikoSI also like it more17:24
@HeikoSits a bit slower to optimise though17:25
shogun-notifier-shogun: cameron :develop * 8f6de9c / examples/undocumented/python_modular/regression_gaussian_process_modular.py: https://github.com/shogun-toolbox/shogun/commit/8f6de9c14fb4ef232b31d75b1c7e53ab8740781217:26
shogun-notifier-shogun: Changed typo in GP regression from RETURN_TYPE_MEAN to RETURN_TYPE_COV17:26
shogun-notifier-shogun: cameron :develop * c381460 / / (135 files): https://github.com/shogun-toolbox/shogun/commit/c3814609396742a2c3fdf19d567d4f89fa3b20cb17:26
shogun-notifier-shogun: Merge remote-tracking branch 'shogun/develop'17:26
shogun-notifier-shogun: cameron :develop * 7790dd1 / data: https://github.com/shogun-toolbox/shogun/commit/7790dd146bf5e315faba55ff926d9f7e080b540817:26
shogun-notifier-shogun: Modified submodule for gp regression test results17:26
shogun-notifier-shogun: cameron :develop * a33603b / data: https://github.com/shogun-toolbox/shogun/commit/a33603beb5fb35596f8d1309d797c7ee4804431c17:26
shogun-notifier-shogun: Updated submodule shogun-data for gp regression tests17:26
shogun-notifier-shogun: cameron :develop * 945a7ff / data: https://github.com/shogun-toolbox/shogun/commit/945a7ff7112becba39a32d5a12ad7345fb7734c317:26
shogun-notifier-shogun: Revert "Updated submodule shogun-data for gp regression tests"17:26
shogun-notifier-shogun:17:26
shogun-notifier-shogun: This reverts commit a33603beb5fb35596f8d1309d797c7ee4804431c.17:26
shogun-notifier-shogun: cameron :develop * e5c8dae / data: https://github.com/shogun-toolbox/shogun/commit/e5c8daed3d5b320929d7d1cff73404a1003bbece17:26
shogun-notifier-shogun: Revert "Modified submodule for gp regression test results"17:26
shogun-notifier-shogun:17:26
shogun-notifier-shogun: This reverts commit 7790dd146bf5e315faba55ff926d9f7e080b5408.17:26
shogun-notifier-shogun: cameron :develop * b4e23c6 / data: https://github.com/shogun-toolbox/shogun/commit/b4e23c6f30a1680974688cbd85e89cb53284a47317:26
shogun-notifier-shogun: Update submodule data for gp regression17:26
shogun-buildbotbuild #1083 of deb1 - libshogun is complete: Failure [failed git]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb1%20-%20libshogun/builds/1083  blamelist: cameron <cameronjml2@gmail.com>17:28
shogun-buildbotbuild #1084 of deb1 - libshogun is complete: Success [build successful]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb1%20-%20libshogun/builds/108417:29
shogun-buildbotbuild #733 of cyg1 - libshogun is complete: Failure [failed configure]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/cyg1%20-%20libshogun/builds/733  blamelist: Heiko Strathmann <heiko.strathmann@gmail.com>, cameron <cameronjml2@gmail.com>17:30
@HeikoSargh17:32
@HeikoSI dont get this17:32
-!- travis-ci [~travis-ci@ec2-50-16-143-238.compute-1.amazonaws.com] has joined #shogun17:33
travis-ci[travis-ci] it's Viktor Gal's turn to pay the next round of drinks for the massacre he caused in shogun-toolbox/shogun: http://travis-ci.org/shogun-toolbox/shogun/builds/682069117:33
-!- travis-ci [~travis-ci@ec2-50-16-143-238.compute-1.amazonaws.com] has left #shogun []17:33
@HeikoScameron__:  but the version is correct17:34
@HeikoS17627e76436866ae52f35875c8cf41df31571bc217:34
@HeikoSboth in data and shogun/data17:34
pickle27hey guys17:34
@HeikoSlisitsyn, sonne|work could you help here?17:34
@HeikoSpickle27: hi17:34
pickle27quick question about my GSoC application17:34
@HeikoSyes?17:34
cameron__HeikoS: I am quite sure everything was right17:35
cameron__HeikoS: not sure what went wrong this time17:35
sonne|workHeikoS: did you git add data?17:35
sonne|workin the src17:35
cameron__sonne|work: i did git submodule update17:35
pickle27when I am actually submitting it on the GSoC site should I put the description from the shogun ideas list in the project description?17:35
shogun-buildbotbuild #899 of bsd1 - libshogun is complete: Failure [failed configure]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/bsd1%20-%20libshogun/builds/899  blamelist: cameron <cameronjml2@gmail.com>17:36
cameron__sonne|work: then git commit -a -m "comment"17:36
@HeikoSsonne|work: the data in src/ is set to the same hash as the head of shogun-data17:36
shogun-buildbotbuild #734 of cyg1 - libshogun is complete: Failure [failed configure]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/cyg1%20-%20libshogun/builds/734  blamelist: cameron <cameronjml2@gmail.com>17:36
pickle27nvm sort of answered my own question by comparing the character limit17:37
@HeikoSpickle27: no dont copy it17:38
@HeikoSjust your application which mentions which project you want to do17:38
sonne|workHeikoS: and you did push data?17:40
@HeikoShttps://github.com/shogun-toolbox/shogun/commit/abab8b64883c5479ddd3d9a533e38f4f50fc6f1617:41
@HeikoShttps://github.com/shogun-toolbox/shogun-data/commit/17627e76436866ae52f35875c8cf41df31571bc217:41
cameron__sonne|work: Yes, data was pushed17:41
sonne|workHeikoS: maybe you pushed data *after* pushing src17:41
sonne|workhttp://shogun-toolbox.org/buildbot/builders/deb1%20-%20libshogun17:41
@HeikoSno17:41
sonne|workit looks like it is happy17:41
@HeikoSah17:42
@HeikoScameron__: you had many intermediate commits right?17:42
cameron__HeikoS: yes I reverted the wrong commits I made beore17:42
sonne|workcameron__: ahh revert is a new commit17:43
sonne|worknext time use git reset17:43
sonne|workor git commit --amend17:43
-!- KMcQuisten [d8338942@gateway/web/freenode/ip.216.51.137.66] has quit [Quit: Page closed]17:43
cameron__sonne|work: okok, sorry for all the mess17:44
-!- votjakovr [~votjakovr@host-46-241-3-209.bbcustomer.zsttk.net] has joined #shogun17:46
@HeikoScameron__: no worries, works now, thanks for the patch!17:47
@HeikoSnow that you know how this works, you can help people wanting to do this :)17:48
cameron__HeikoS: Thanks for all the help! I can now do the unfinished gp regression interactive17:48
@HeikoSnice!17:48
@HeikoSthank roman for fixing the bug :)17:48
@HeikoSvotjakovr17:48
cameron__thanks votjakovr17:49
@HeikoSsonne|work: any thoughts on modelselection for subkernels?17:49
votjakovrIt's my pleasure :)17:49
@HeikoSsonne|work: maybe this can be added to the model-selection syntax project17:54
@HeikoSsince people have asked about this a few times17:54
@wikingsonne|work: make[1]: *** No rule to make target `../../shogun/SFMT-params.h', needed by `modshogun_wrap.cxx'.  Stop.17:57
@wikingwtf?17:57
@wiking :)17:57
@wikingah shit17:57
@wikinglisitsyn: ping17:59
-!- ErlendA [~ErlendA@cm-84.215.138.251.getinternet.no] has joined #shogun18:03
ErlendA@HeikoS: Did Dan talk to you about exactly how to precondition the methods?18:03
@HeikoSErlendA: hi18:04
@HeikoSa little bit18:04
ErlendAGood18:04
@HeikoSlog_det=det_Krylov_preconditioned(Q, P, precomputed, tol, maxiter);18:04
@HeikoSlog_det=log_det+2*sum(log(diag(P)));18:04
@HeikoSto speak in matlab18:04
@HeikoSuse P to create a better conditioned problem and then correct in the logdet result18:05
ErlendAYes. Exactly, or similarly with a spectral approach18:05
@HeikoSErlendA: so I tried this on a large problem18:05
@HeikoSand then discarded it18:05
@HeikoSsince its slower18:05
ErlendAWhat did you try?18:05
@HeikoSapplying P all the time causes so much overhead that the gain in the condition number is lost18:05
@HeikoSpreconditioned vs not preconditioned18:06
ErlendAAhh. But this depends on how the preconditioner is constructed.18:06
ErlendAIncomplete cholesky?18:06
sonne|workwiking: you need the include path to be relative to /src/18:06
@HeikoSErlendA: true, so I (we) tried incomplete cholesky18:06
@HeikoSdid not really do well18:06
sonne|work<shogun/foo/bar>18:06
@HeikoSjacobi18:06
thoralfHeikoS: Please have a look at shogun/lib/SGSparseVector.cpp, method ::copy_data, line "features = ((SGSparseVector*)(&orig))->features;"18:06
@HeikoSand block wise jacobi18:06
thoralfHeikoS: The reference gets copied to the new vector, so it gets double-freed?18:07
@wikingsonne|work: where...? i mean i have no idea where is this coming from18:07
@HeikoSErlendA: but thats on a very specific problem18:07
@wikingsonne|work: ok i might have an idea...18:07
@wiking:(18:07
ErlendA@HeikoS: Yes. But it is important to have the posibility of using it.18:08
@HeikoSErlendA: aggreed!18:08
@HeikoSthoralf: which line is this?18:08
@HeikoSnumber?18:08
sonne|workwiking: from the includes in the sfmt18:08
sonne|workyou have to make them all like18:08
ErlendA@HeikoS: I tried incomplete cholesky with wavelet sparsification. Then it works quite well for SPDE problems18:08
shogun-buildbotbuild #1032 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/1032  blamelist: cameron <cameronjml2@gmail.com>18:08
sonne|work#inlcude <shogun/foo/bar.h>18:08
@HeikoSErlendA: we should talka bout everything soon, we got a very promising guy here18:08
@HeikoSErlendA: nice, so then we should add it18:08
thoralfHeikoS: line 11418:09
@wikingsonne|work: yep i got that in the meanwhile :P18:09
ErlendA@HeikoS: It's difficult to program it in a way that is general and useful. The gist should be (IMO) that you can use your customized preconditioner, if you happen to have a good one18:09
@HeikoSErlendA: ok, sounds good18:10
@HeikoSyes, then we can make it optimal to code the actual preconditioner18:10
@HeikoSthoralf: I dont think this is a problem18:10
ErlendA@HeikoS: Yes. I think that in the end, that's more useful18:11
@HeikoSthoralf: I suspect if you run the code with SGVector instead of sparse vectors, then it will also crash18:11
@HeikoSI think it has to do with the way the streaming framework handles vector data18:11
@HeikoSErlendA: looking forward, will be a cool project18:11
sonne|workHeikoS: I didn't follow what the issue with subkernels & MS is...18:11
@HeikoSsonne|work: imageine you have two kernels that you combine in a combined one18:12
@HeikoSnow you want to select the parameters of the subkernels18:12
sonne|workok18:12
thoralfHeikoS: But since the old vector (including the "features" array) is already freed, this is my suspect #1.18:12
sonne|workHeikoS: and?18:13
@HeikoSthoralf: ok, maybe then try changing this and see where it brings you?18:13
@HeikoSsonne|work: so this is currently not possible with shogun18:13
@HeikoSand the modsel framework18:13
-!- iglesiasg [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun18:13
-!- mode/#shogun [+o iglesiasg] by ChanServ18:13
sonne|workHeikoS: because you need to combine each parameter setting of one kernel with the settings of the other right?18:14
shogun-buildbotbuild #1033 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/1033  blamelist: Heiko Strathmann <heiko.strathmann@gmail.com>, cameron <cameronjml2@gmail.com>18:14
@HeikoSsonne|work yes18:14
@HeikoSalso, the subkernels are not really seperately registered18:15
@HeikoSso some work to do there18:15
@HeikoSmight that fit within the modsel syntax task?18:15
sonne|workHeikoS: so best you could do is do various combined kernels ?18:15
@HeikoSsonne|work: yes if one creates all combined kernels by hand, one can let shogun do the rest18:15
@HeikoSbut thats the point, we want to create those combinations automatically18:15
sonne|workHeikoS: if we had just 2 fixed kernels it would work right?18:16
sonne|workI mean the machine had 2 kernels18:17
sonne|workbut not a combined one18:17
@HeikoShow do you mean that?18:17
@HeikoSah I see18:17
@HeikoSyes18:17
@HeikoStwo combined kernels18:17
sonne|worksay CKernel a; CKernel b18:17
@HeikoSyes or whatever kernel18:17
sonne|workyeah18:17
@HeikoSthis we can do18:17
sonne|workI don't see how we could do this in general18:18
@HeikoSwell18:18
sonne|workwe only ever thought about Machines for that18:18
@HeikoSthe combined kernel gets subparameters18:18
@HeikoSwhich are then applied to the subkernels one by one18:19
sonne|workah so each kernel would be a sub-parameter18:19
@HeikoSyes something like this18:19
@HeikoSproblem is its just a list18:19
@HeikoSso every element of the list would need a name18:19
sonne|workkind of !?!? considering that we can have a dynamic number of sub kernels18:19
@HeikoSand then we can give them parameters18:19
@HeikoSgets even more complicated when you have different types with different types of subparameters18:20
sonne|workHeikoS: yeah but we assumed the parameters to be fixed18:20
@HeikoSyes18:20
sonne|workthis here will add parameters dynamically18:20
sonne|workand you might want sometimes 2 kernels sometimes 418:20
@HeikoSI mean, this would be a kind of hack for combined kernels18:20
sonne|workor so18:20
sonne|workhorrible18:20
@HeikoSmmh18:21
@HeikoSmaybe then instead, we should offer a method that creates a bunch of combined kernels for a user18:21
sonne|workI guess we are better of supplying some function to create the cross-product for kernels18:21
@HeikoSand then these are passed to the machine one after another18:21
@HeikoSexactly18:21
sonne|workhaha18:21
@HeikoSsame thought :)18:21
sonne|workthen it must be the better solution18:21
sonne|workalso not so difficult18:22
@HeikoSso given a list of kernel lists, this one create combined kernels with all the combinations18:22
sonne|workyes18:22
@HeikoSthat sounds reasonable18:22
@HeikoSkernel type has to be fixed though since the features must not change18:22
sonne|workmaybe you could even tell him how to do it18:22
@HeikoSbut thats fine18:22
sonne|workyeah well18:22
sonne|workthere will always be too complex ways to automate things18:23
@HeikoSindeed18:23
@HeikoSthis one would be good to have though, there was another guy that asked about it a while ago18:23
sonne|workso create an issue about it and mention the solution18:23
@HeikoSwill do18:23
sonne|workmaybe someone has time...18:24
-!- vgorbati [d4029f22@gateway/web/freenode/ip.212.2.159.34] has quit [Ping timeout: 245 seconds]18:28
-!- abinayam [3d0c137b@gateway/web/freenode/ip.61.12.19.123] has joined #shogun18:29
-!- abinayam [3d0c137b@gateway/web/freenode/ip.61.12.19.123] has quit [Client Quit]18:32
shogun-notifier-shogun: Roman Votyakov :develop * 74ac381 / src/NEWS: https://github.com/shogun-toolbox/shogun/commit/74ac38185707411786f53a7f2fdd159a044532ed18:32
shogun-notifier-shogun: Updated news18:32
shogun-notifier-shogun: Heiko Strathmann :develop * 22eb0a5 / src/NEWS: https://github.com/shogun-toolbox/shogun/commit/22eb0a5e85caae50f861c0cdd12bceef697a6af718:32
shogun-notifier-shogun: Merge pull request #1041 from votjakovr/develop18:33
shogun-notifier-shogun:18:33
shogun-notifier-shogun: Updated news18:33
@HeikoSsonne|work: whats this proa in the news? :)18:35
votjakovrI'd like to ask same question :)18:36
@wikingHeikoS: i'm getting [ERROR] assertion CMath::abs(mean1-mean2)<10E-8 failed in file statistics_hsic.cpp line 159 with the new PRNG :)18:40
@HeikoSwiking: no surprise :(18:40
@HeikoSrelax the bound a bit18:40
@wikingE-6?18:41
@HeikoSI have to change this at some point18:41
@HeikoSjust try, seed should be fixed18:41
thoralfHeikoS: The second reference to the sparse features comes from StreamingSparseFeatures.cpp, line 426.  The reference is just copied.  So later on, the same "features" pointer can be found in the vector returned by get_vector() as well as the (distinct) current_sgvector.18:41
@wikingHeikoS: well seed is fixed  ;)18:41
thoralfHeikoS: Can you verify this?18:41
@HeikoSyes18:42
@HeikoS!18:42
@HeikoSok I see18:43
@HeikoSso the current vector is freed18:43
thoralfHeikoS: I just SG_MALLOC'ed the array and returned this, so the double-free disappeared.  But where to hook for proper deallocation?18:43
@HeikoSwhich effectively deletes the data18:43
shogun-buildbotbuild #900 of bsd1 - libshogun is complete: Success [build successful]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/bsd1%20-%20libshogun/builds/90018:44
thoralfHeikoS: Maybe the semantic of get_vector() is just broken by design: What about returning a pointer to the vector?  Should fix the problem without need to copy everything all over again.18:44
@HeikoSget_vector should just add a reference count to the vector18:44
thoralfHeikoS: SG_REF?18:45
@HeikoSwe have automatic reference counts for copy by value on vectors18:45
@HeikoSso as long as you only do copy by value (=), vectors are fine18:45
@HeikoSwhen the counter reaches zero, memory is freed18:45
@HeikoSthis should be used here18:45
@HeikoSthen the framework internally deals with the vectors18:45
@HeikoSthe user also might18:45
@HeikoSand once everyone is done it gets deleted18:46
@HeikoScounter is decremented once a vector structure is removed from stack18:46
thoralfHeikoS: SG_ADD(current_sgvector.features)?18:46
thoralfHeikoS: SG_REF(current_sgvector.features)?18:46
@HeikoSnono18:47
@HeikoSsee if you have18:47
@HeikoSSGVector a18:47
@HeikoSand you do SGVector b=a18:47
@HeikoSthen it is incremented18:47
@HeikoSa and b share the same data18:47
thoralfYes.18:47
@HeikoSyou if you return a vector18:47
@HeikoScounter is incremented18:47
thoralfHeikoS: But which do I (we) have to take care about?18:49
thoralfHeikoS: Actually there is no explicit reference counting on .features18:50
@iglesiasgHeikoS: too many PRs lately18:51
@iglesiasgHeikoS: one forgets what's merged and what not :)18:51
@HeikoSiglesiasg: indeed :D18:51
@HeikoSthoralf: so thats what sonne|work meant when he said the issues are related to the SGVector stuff18:51
@HeikoSthe framework is from the days before we had the automatic stuff18:51
@HeikoSso it takes care of the SGVectors18:51
@HeikoShowever, it should not18:51
shogun-notifier-shogun: Viktor Gal :feature/PRNG * 972c7ea / src/shogun/lib/external/SFMT/ (4 files): https://github.com/shogun-toolbox/shogun/commit/972c7eaaf01491cbb768b9be246caa58a8927b7218:51
shogun-notifier-shogun: SFMT: add corrent shogun path for includes18:52
shogun-notifier-shogun: Viktor Gal :feature/PRNG * 9b9866d / src/shogun/mathematics/Random.cpp,src/shogun/mathematics/Random.h: https://github.com/shogun-toolbox/shogun/commit/9b9866d33d17fd74e88473084356231f8793b02c18:52
shogun-notifier-shogun: Several changes in Random PRNG API18:52
shogun-notifier-shogun: move state check in the lock part to avoid invalid state checks18:52
shogun-notifier-shogun: inline reinit() for faster running time18:52
@HeikoSno free'ing should happen, since its all done automatically18:52
shogun-notifier-shogun: add set/get function for seed18:52
shogun-notifier-shogun: add max macros for RAND_32 and RAND_6418:52
shogun-notifier-shogun: Viktor Gal :feature/PRNG * aa75124 / src/shogun/mathematics/Math.h: https://github.com/shogun-toolbox/shogun/commit/aa7512414bce51ba40ef026b832ec893114e5fba18:52
shogun-notifier-shogun: Change CMath's PRNG generator18:52
thoralfHeikoS: Okay, and which part belongs to "the old stuff"?  I guess sg_ref and sg_unref are new?18:52
shogun-buildbotbuild #735 of cyg1 - libshogun is complete: Success [build successful]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/cyg1%20-%20libshogun/builds/73518:53
@HeikoSno SG_REF is for SGObject, so class instances18:53
@HeikoSthoralf: the old stuff is18:53
@HeikoShaving a c-array float64_t* array18:53
@HeikoSand a length variable int32_t array_len18:53
@HeikoSand then allocate/deallocate array by hand18:53
@HeikoSall those pairs can however be replaced by SGVector which does everything18:53
shogun-buildbotbuild #1034 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/1034  blamelist: cameron <cameronjml2@gmail.com>18:54
@HeikoSlet me find an example ..18:54
@HeikoSvirtual float32_t dense_dot(const float32_t* vec2, int32_t vec2_len)=0;18:56
@HeikoSin StreamingDotFeatures18:56
@HeikoSthis is how we handled vectors before18:56
@HeikoSthoralf: it might be a bit of an overkill for you to solve this18:58
@HeikoSits a lot of stuff18:58
thoralfHeikoS: But I think we are close.  We saw, that we store two references to the same .features *and* we know, that the first instance is created by the CStreamingSparseFeatures constructor and the second by get_vector().19:02
thoralfHeikoS: If I just create a copy of .features just before returning current_sgvector.  Will this leak or cleaned up automatically?19:03
@HeikoSthoralf:  have you valgrinded the program?19:03
@HeikoSthen you get the line where the double free happens inside the framework19:04
-!- iglesiasg [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving]19:04
-!- vgorbati [d4029f22@gateway/web/freenode/ip.212.2.159.34] has joined #shogun19:04
@lisitsynHeikoS: hey19:05
@lisitsynyou asked for some help, is it still relevant?19:05
@HeikoSlisitsyn: hi!19:07
@HeikoSehm, on what?19:07
@lisitsynI am lazy to read all these things :D19:07
@HeikoSthoralf:  I would rather change the whole handling of vectors19:07
@HeikoSlisitsyn: so streaming features memory errors19:07
@lisitsynbut you asked a few hours ago or so19:07
@HeikoSis the most recent19:07
@HeikoSlisitsyn: ah that was the build, cannot remember19:08
@HeikoSSGSparseVector<T> current_sgvector;19:08
@HeikoS/// The current example's feature vector as an SGSparseVectorEntry<T>*.19:08
@HeikoSSGSparseVectorEntry<T>* current_vector;19:08
@lisitsynalright19:08
@HeikoSthoralf: these guys are in SparseStreaming19:08
shogun-buildbotbuild #1035 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/1035  blamelist: Roman Votyakov <votjakovr@gmail.com>19:08
@HeikoSand they get freed automatically on destructor call19:08
thoralfHeikoS: But thats alright - as long get_vector() is not called. ;)19:08
@HeikoScameron__: btw the python test fails ...19:09
@HeikoSthoralf: exactly19:09
@HeikoSso get_vector should has to either incremed counter or create a copy19:09
@HeikoSmaybe we can try to copy?19:10
@HeikoSmmmh19:10
@HeikoSdamn, all the examples/tests are not working, so no way of checking whether this makes things lead19:10
@HeikoSleak19:10
@HeikoSthoralf: can you check what happens with dense vectors?19:11
@HeikoSsame issue there?19:11
@HeikoSI use this for some kernel hypothesis testing stuff and fixed some things in there19:11
@HeikoSmight be the same problem19:11
cameron__HeikoS: but they do not fail on my machine19:12
@HeikoScameron__: weird, lets wait a bit, maybe they will start working at some point19:13
thoralfHeikoS: CStreamingDenseFeatures seems to work.19:13
@HeikoSthoralf: actually now that I think of it, the issues with the dense ones were pretty similar19:13
@HeikoSand I somehow managed to solve parts of it (its already a while ago) - so maybe thats a good point to look at19:14
cameron__HeikoS: I got to go , I'll hav a look later.19:14
@HeikoScameron__: ok see you19:14
thoralfHeikoS: Found a comment in dense features: /* needed to prevent double free memory errors */19:16
@HeikoSthoralf: ha! these are mine, what did I do there? which line19:17
thoralfHeikoS: Related to the get_vector() and current_vector handling :)19:17
thoralfStreamingDenseFeatures.cpp, 5919:17
thoralfYou just NULL'ed the vector, so it's not getting freed.19:17
thoralfPragmatic solution ;)19:18
@HeikoSthoralf: ehm, yes ..... hacks :)19:18
@HeikoSnot really the way to do it19:19
@HeikoSbut maybe for now, add a comment that makes sure our thoughts can be reconstructed19:20
vgorbatilisitsyn: here?19:20
@sonney2kHeikoS, the isssue with this is that it needs a big overhaul - and so quite some time19:23
@sonney2kthat is why we keep postponing it19:23
@HeikoSsonney2k: yes, maybe go with the hack for now and then fix at some point19:24
@HeikoScurrently is not the time for this19:24
@HeikoStoo many things already happening for gsoc19:24
@sonney2kwiking, how does it look?19:24
@sonney2kdo tests fail completely now?19:24
@sonney2kHeikoS, yeah overload mode19:25
@HeikoSyes ...19:25
-!- gsomix_ is now known as gsomix19:26
@sonney2klisitsyn, hmmhh how many slots do we minimally need19:26
@sonney2kit seems like 9 now...19:26
@sonney2kwhich might be more than we can get19:27
@sonney2klisitsyn, https://help.github.com/articles/duplicating-a-repository19:27
@sonney2kis this what we should do for the clone?19:27
@sonney2klisitsyn, any thoughts? should I call this say once per day?19:28
@sonney2kcameron__, things fail on the buildbot because it is testing each commit19:30
@sonney2kif you last commit is fine it will be happy then19:31
shogun-notifier-shogun: Heiko Strathmann :develop * 8bf9583 / src/NEWS: https://github.com/shogun-toolbox/shogun/commit/8bf9583a964ad1c4a8834b8c9a59bf8499c06e7319:33
shogun-notifier-shogun: news update19:33
shogun-notifier-shogun: Heiko Strathmann :develop * cec2f66 / src/NEWS: https://github.com/shogun-toolbox/shogun/commit/cec2f66beb322eba6ad5977a6aeb28199bc5795119:33
shogun-notifier-shogun: Merge pull request #1043 from karlnapf/develop19:33
shogun-notifier-shogun:19:33
shogun-notifier-shogun: news update19:33
@wikingsonney2k: ok this is crazy.... when i run generator.py on linux it modifies like 10 files in data19:33
@wikingsonney2k: the same on bsd it modifies 20+ files19:33
@sonney2kwiking, only? cool!19:33
@sonney2khmhh19:34
@sonney2kand there is no more call to rand etc in Math?19:34
@wikingmmm ok question19:36
@wikingin static inline uint64_t random(uint64_t min_value, uint64_t max_value)19:36
@wikingrandom19:36
@wikingcall19:36
-!- vikalp [~chatzilla@115.245.61.191] has joined #shogun19:36
@wikingbut i think that actually calls system's random()19:36
@wikingnot CMath::random19:36
thoralfHeikoS: My ideas won't work out.  Setting current_sgvector.features to NULL seems to be the best way.19:36
@HeikoSthoralf: go ahead then :)19:36
@HeikoScan you add a comment in the streaming features issues on github?19:37
@sonney2kwiking, haha there you have it!19:37
thoralfHeikoS: I can't easily fix the two references to to same data without introducing big leaks.19:37
@HeikoSok19:37
@HeikoSits ok for now19:37
@HeikoSsonney2k, lisitsyn I will be in Germany over the weekend, so I won't be around in IRC much, let me know if I should do anything via email, Ill try to check at least once a day19:38
-!- ErlendA [~ErlendA@cm-84.215.138.251.getinternet.no] has quit [Quit: Leaving]19:38
shogun-buildbotbuild #737 of cyg1 - libshogun is complete: Failure [failed configure]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/cyg1%20-%20libshogun/builds/737  blamelist: Heiko Strathmann <heiko.strathmann@gmail.com>19:39
@sonney2kHeikoS, most important currently is to keep in touch with students somehow and to come up with some close to final candidate list19:39
@sonney2kthat's it19:39
@HeikoS././configure-1095-2324.o: Permission denied19:39
@sonney2kHeikoS, yeah cygwin19:39
@HeikoSsonney2k: already doing that19:39
@sonney2kI have no clue what's going wrong there19:39
@HeikoSok cool, will do tomorrow19:40
shogun-buildbotbuild #738 of cyg1 - libshogun is complete: Failure [failed configure]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/cyg1%20-%20libshogun/builds/738  blamelist: Heiko Strathmann <heiko.strathmann@gmail.com>19:40
@HeikoSok then, going home now, see you later!19:40
@sonney2kI am not sure I can handle all students applying for the ideas I had19:40
@HeikoSsonney2k: yeah seems a lot of work19:40
@lisitsynsonney2k: well you are going to have two right?19:40
@HeikoSbye guys!19:41
@lisitsynHeikoS: bye19:41
thoralfBye HeikoS, thanks.19:41
@sonney2klisitsyn, currently it is 3 on the list19:41
@lisitsynsonney2k: I'll take one19:41
@sonney2kthoralf, thanks for `fixing' this!19:41
vikalpheya..19:41
@sonney2klisitsyn, ah ok19:41
@sonney2kvikalp, hey19:42
@lisitsynsonney2k: just to low your load I mean19:42
vikalphii sonney2k19:42
-!- HeikoS [~heiko@nat-171-79.internal.eduroam.ucl.ac.uk] has left #shogun []19:42
@sonney2klisitsyn, maybe we can get cheng involved too19:42
@lisitsynsonney2k: I hope so19:42
vikalp@sonney2k.. i have completed the task of writing general csv writer in C++19:42
vikalpi am about to write my proposal regarding Fast Reading and Writing Project..19:43
@sonney2kvikalp, ahh nice!19:43
vikalpi mean..completing it..19:43
vikalp:)19:44
-!- 45PAAIYF0 [~travis-ci@ec2-50-16-143-238.compute-1.amazonaws.com] has joined #shogun19:44
45PAAIYF0[travis-ci] it's Viktor Gal's turn to pay the next round of drinks for the massacre he caused in shogun-toolbox/shogun: http://travis-ci.org/shogun-toolbox/shogun/builds/682575419:44
-!- 45PAAIYF0 [~travis-ci@ec2-50-16-143-238.compute-1.amazonaws.com] has left #shogun []19:44
@lisitsynhaha nice nickname19:44
@lisitsyn45PAAIYF019:44
@sonney2kI thought we were being attacked19:44
@lisitsynhahah19:44
@sonney2klisitsyn, recall the bot last year?19:44
@lisitsynsonney2k: yes I regret you kicked it19:45
-!- monalisag [6eeaec2c@gateway/web/freenode/ip.110.234.236.44] has joined #shogun19:45
@sonney2kc'mon it was getting nasty ;)19:45
shogun-buildbotbuild #1036 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/1036  blamelist: Heiko Strathmann <heiko.strathmann@gmail.com>19:45
@lisitsynit was quite intelligent!19:45
shogun-notifier-shogun: Viktor Gal :feature/PRNG * 28311b0 / src/shogun/mathematics/Math.h: https://github.com/shogun-toolbox/shogun/commit/28311b0ba43d93f9721de4e68efb37e608bd512b19:45
shogun-notifier-shogun: Use CMath::random() in all other CMath::random functions19:45
shogun-notifier-shogun: Viktor Gal :feature/PRNG * 7f737fa / src/configure: https://github.com/shogun-toolbox/shogun/commit/7f737fa6f3a84c6733e48b81a6d797aaa2543de319:45
shogun-notifier-shogun: Add -I. into test_mexp() function in configure19:45
@sonney2klisitsyn, what do you think about mirroring?19:45
@lisitsynsonney2k: what? which mirroring?19:45
@sonney2klisitsyn, git!19:45
@sonney2kgithub -> google code19:46
@lisitsynahh19:46
@sonney2k& shogun toolbox19:46
@lisitsynwhy do we need it?19:46
@lisitsynI never knew19:46
@lisitsynI mean well if github was unreliable19:46
@sonney2klisitsyn, don't rely on one source19:46
@lisitsynbut actually git is distributed so even if chinese attack github19:46
vikalp@sonney2k.. As I have gone through the task list of project..what all are the necessaries that I need to complete under the timeline?19:46
@sonney2kservices come and go19:46
@lisitsynwe have the same thing on your machine19:46
@lisitsynand on my machine19:46
@sonney2kvikalp, well depends on you :)19:47
@sonney2kthe project can be stretched quite a bit19:47
@lisitsynsonney2k: I'd put a cron on the website server that git pulls and git pushs19:47
@sonney2kfrom lazy programming till super cow powers ;)19:47
@sonney2klisitsyn, yes that is what I was suggesting - I would do it once / day19:48
@lisitsynthis is definitely not the manual thing19:48
vikalpthats not a problem.. I would love to work on machine learning19:48
@lisitsynwell I have an alias19:48
@lisitsyngposg :D19:48
@lisitsyngit push origin && git push shogun && git push google19:48
@wikingsonney2k: do you know what library_fisher2x3_modular0.txt does19:48
@wikingor shoudl contain?19:48
@wikingsonney2k: because i dont really think that it depends on any random generator :)19:49
@sonney2kwiking, ohh it is some fisher test for some snp stuff19:50
-!- votjakovr [~votjakovr@host-46-241-3-209.bbcustomer.zsttk.net] has quit [Quit: Leaving]19:50
vikalpSo regarding timeline what all should I include in there?.. Can you please elaborate on this..:)19:50
@wikingsonney2k: because the diff between the generated file and the one in the data repo is really just this but still of course it fails...19:51
@wiking-(F0.0992749338401746219:51
@wiking+(F0.0992749338401802719:51
@wiking-S"\x11\xf7'\x02\x15j\xb9?\x06\xdd\xe3\x95{\xd2\xe7?:?\xbc%Qb\xe2>"19:51
@wiking+S"\xa8\xf8'\x02\x15j\xb9?\x06\xdd\xe3\x95{\xd2\xe7?:?\xbc%Qb\xe2>"19:51
@sonney2kwiking, so I guess exp() and whatever are different19:52
* sonney2k is scared19:52
@wikingi dont know19:52
@wikingbut this has of course nothing to do with19:52
@wikingrng :(19:52
@sonney2ktrue19:52
@wikingand i'm afraid same goes for kernel_log_modular0.txt19:53
@wikingand the others19:53
@sonney2kwiking, ohh do you have long double?19:53
@sonney2kI mean long double is 128bit on some archs19:54
@sonney2kand on some 9619:54
@wikingwhere should i check what19:54
@sonney2ksizeof(long double)19:54
@sonney2kanyway it seems we have to rething the comparison strategy19:55
@sonney2krethink19:55
@wikingsonney2k: 1619:55
vikalpsonney2k, So regarding timeline what all should I include in there?.. Can you please elaborate on this..:)19:56
@sonney2kvikalp, what you think you can do19:56
@sonney2k*reasonably*19:56
@sonney2kthere is always room for doing a little less /m ore19:56
@sonney2kor maybe even continue in the future19:56
vikalpsonney2k, well m planning to continue it19:57
-!- vgorbati [d4029f22@gateway/web/freenode/ip.212.2.159.34] has quit [Ping timeout: 245 seconds]19:57
-!- monalisag [6eeaec2c@gateway/web/freenode/ip.110.234.236.44] has quit [Ping timeout: 245 seconds]19:58
vikalpsonney2k, so I will be sending you my proposal for  your reviews in couple of hours..19:58
mikhailBelousI fixed my proposal.19:59
-!- vikalp [~chatzilla@115.245.61.191] has quit [Quit: ChatZilla 0.9.90 [Firefox 11.0/20120312181643]]20:00
-!- monalisag [6eeaec2c@gateway/web/freenode/ip.110.234.236.44] has joined #shogun20:01
thoralfBye guys.  Happy hacking.20:03
-!- thoralf [~thoralf@enki.zib.de] has left #shogun ["Konversation terminated!"]20:03
-!- mikhailBelous [~quassel@37.98.242.99] has quit [Remote host closed the connection]20:08
shogun-buildbotbuild #1037 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/1037  blamelist: Heiko Strathmann <heiko.strathmann@gmail.com>20:08
@wikingsonney2k: this is not helping at all....20:10
@wikingi mean the new PRNG20:11
@sonney2kwiking, wait - it did get a bit better right?20:14
@sonney2kI mean now the random stuff gives the same result or?20:14
@wikingsonney2k: well i'll check20:15
@wikingi'll now regenerate on linux20:15
@wikingand copy it to osx20:15
@wikingand check whether some at least got better or not20:15
@sonney2kwiking, some should have gotten better20:15
@sonney2klisitsyn, git mirror to shogun toolbox works20:17
@sonney2kbut not to code.google20:17
@sonney2kno idea why not20:17
@lisitsynehmm20:17
@sonney2kerror: unable to push to unqualified destination: HEAD20:18
@sonney2kThe destination refspec neither matches an existing ref on the remote nor20:18
@sonney2kbegins with refs/, and we are unable to guess a prefix based on the source ref.20:18
@sonney2kfatal: The remote end hung up unexpectedly20:18
@sonney2kwtf??20:18
@wikingyep20:18
@lisitsynnever seen that madness20:18
@wikingis it by github?20:18
-!- pickle27 [~kevin@rcv3-lab-pc.ee.queensu.ca] has quit [Quit: Leaving]20:19
@sonney2kwiking, no code.google.com20:19
@sonney2kcrazy!20:24
@sonney2kit works from my notebook20:25
@sonney2kbut not from our server20:25
@wikingsonney2k:20:25
@wikingsonney2k: 2 of all failed examples are not failing anymore :D20:25
@sonney2kwiking, hurray!20:25
@sonney2kwiking, so how many are left?20:26
@wikingabout 3020:26
@wikingbut of course the new PRNG introduced fails in tests20:26
@wiking:D20:26
@wikingsoooo fuck this shit20:26
@sonney2kwiking, no no it is the way to go(tm)20:28
@sonney2khey yeah works now20:28
@sonney2kwith newer version of git20:28
gsomixsonney2k, wiking: hey, can we add bot for building python_modular with python3 bindings?20:29
@sonney2kgsomix, I don't have a python3 aware system20:30
@sonney2kwiking, do you?20:30
@sonney2klisitsyn, so we have a mirror now!20:30
@sonney2kgtg20:31
@sonney2kcu20:31
-!- monalisag [6eeaec2c@gateway/web/freenode/ip.110.234.236.44] has quit [Ping timeout: 245 seconds]20:32
shogun-buildbotbuild #1038 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/1038  blamelist: Heiko Strathmann <heiko.strathmann@gmail.com>20:36
-!- travis-ci [~travis-ci@ec2-50-16-143-238.compute-1.amazonaws.com] has joined #shogun20:39
travis-ci[travis-ci] it's Viktor Gal's turn to pay the next round of drinks for the massacre he caused in shogun-toolbox/shogun: http://travis-ci.org/shogun-toolbox/shogun/builds/682720420:39
-!- travis-ci [~travis-ci@ec2-50-16-143-238.compute-1.amazonaws.com] has left #shogun []20:39
-!- vgorbati [d4029f22@gateway/web/freenode/ip.212.2.159.34] has joined #shogun20:40
vgorbatilisitsyn: here?20:48
@lisitsynvgorbati: yes20:48
vgorbatilisitsyn: I am thinking of implementation of manifold sculpting. Should I create a SparseMatrix of local distances to neighbors from distance object, or should just pass the distance object inside the actual implementation of the method?20:50
@lisitsynvgorbati: it depends how many times it is going to be used20:51
vgorbatilisitsyn: you mean whether it will compute the same distances over and over again?20:51
@lisitsynI mean if it is used on each iteration it is worth to compute and store20:51
@lisitsynyes20:51
vgorbatilisitsyn: I thought that the distance object performs some caching20:52
-!- iglesiasg [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun20:53
-!- mode/#shogun [+o iglesiasg] by ChanServ20:53
@lisitsynvgorbati: it might but I would not rely on that in general case20:53
-!- makokal [~bokal@212.201.44.247] has joined #shogun20:53
vgorbatilisitsyn: agree20:53
vgorbatilisitsyn: ok, then I will use the SparseMatrix. Should I put the method that gets SparseMatrix of distances to neighbors from distance object into the locally_linear.hpp?20:55
@lisitsynvgorbati: no it is not related to LLE I think20:55
@lisitsynjust put it to routines/manifold_sculpting.hpp for now20:55
@lisitsynit is not used anywhere more20:56
vgorbatilisitsyn: ok. Then I can (to make an 'embedManifoldSculpting' method more clear) pass in the distance object, and get the SparseMatrix from it inside the actual implementation function, what do you think?20:57
@lisitsynyes sure20:57
vgorbatilisitsyn: ok, thanks20:58
gsomix"so, we need a crane to lift another crane from a trench! wait... oh shi~" http://ic.pics.livejournal.com/human_oid/15606149/504066/504066_original.jpg20:58
gsomixhappy day of workers20:58
-!- van51 [~van51@athedsl-401908.home.otenet.gr] has joined #shogun21:04
van51hey all21:06
van51quick question, do I have to include a cv in the GSoC proposal?21:07
@iglesiasgvan51: no need I think, but feel free if you'd like to21:07
@lisitsynvan51: you don't have to21:07
van51ok, thanks!21:08
@sonney2klisitsyn, I am pretty pleased with the 28 we have now...21:21
@lisitsynsonney2k: 28 already?21:21
@lisitsynohh21:21
@sonney2kconsidering that we had almost 20 on the last day last year21:21
@sonney2kso maybe we get a handful more21:22
@sonney2kbut it is not such a drastic decrease then... already ~60% of what we had last year21:22
@lisitsynwe will get a few more I think21:27
@sonney2kI should write some blog post about this ride again21:28
-!- vgorbati [d4029f22@gateway/web/freenode/ip.212.2.159.34] has quit [Ping timeout: 245 seconds]21:42
van51I just submitted my proposal. Please let me know if you think I should change/clarify something :)22:10
@lisitsynvan51: looks pretty ok22:21
van51lisitsyn: cool, thanks for taking the time to read it22:22
@lisitsynvan51: you are the 29th submitted :)22:22
van51hehe should have waited a bit more, to have it rounded22:23
@lisitsynmay be I should apply by myself as I am eligible to be a gsoc student haha22:23
van51naah it won't be a challenge for you22:25
van51and i'm not saying that because i want a spot :P22:25
@lisitsynhaha22:25
@lisitsynmentor: me22:25
van51like some actors who write,direct and star in their movies22:28
van51you will propose,mentor and implement your proposal22:28
@lisitsynsounds like 'google, just give me money'22:29
@lisitsyn:D22:29
van51yea22:29
van51in a more polite way22:29
van51I'm off for today22:38
van51cya22:38
@lisitsynsee you22:38
-!- van51 [~van51@athedsl-401908.home.otenet.gr] has quit [Quit: Leaving.]22:38
-!- wencan [~wencan@c-71-61-182-121.hsd1.pa.comcast.net] has quit [Remote host closed the connection]22:40
-!- iglesiasg [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving]22:40
-!- iglesiasg [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun22:42
-!- mode/#shogun [+o iglesiasg] by ChanServ22:42
-!- makokal [~bokal@212.201.44.247] has quit [Read error: Connection reset by peer]22:42
-!- makokal [~bokal@212.201.44.247] has joined #shogun22:43
-!- shogun-notifier- [~irker@7nn.de] has quit [Quit: transmission timeout]22:45
-!- makokal [~bokal@212.201.44.247] has quit [Quit: makokal]22:49
@sonney2kwiking, poing?22:51
@sonney2kstill there?22:51
-!- makokal [~bokal@pD9EF09DC.dip0.t-ipconnect.de] has joined #shogun23:22
-!- cameron__ [~quassel@global-2-11.nat.csx.cam.ac.uk] has quit [Ping timeout: 252 seconds]23:33
-!- cameron__ [~quassel@global-2-11.nat.csx.cam.ac.uk] has joined #shogun23:40
cameron__cameronjml223:44
-!- cameron__ [~quassel@global-2-11.nat.csx.cam.ac.uk] has quit [Remote host closed the connection]23:44
@wikingsonney2k: pong23:53
--- Log closed Fri May 03 00:00:30 2013
« Wednesday, 2013-05-01 Index Friday, 2013-05-03 »

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