--- Log opened Thu Jul 19 00:00:17 2012 | ||
-!- heiko [~heiko@host86-177-183-33.range86-177.btcentralplus.com] has left #shogun [] | 00:09 | |
wiking | sonney2k: noup | 01:05 |
---|---|---|
wiking | sonney2k: just got stuck with one part of the code | 01:05 |
-!- romi_ [~mizobe@200-232-221-127.dsl.telesp.net.br] has quit [Remote host closed the connection] | 01:16 | |
-!- puffin444 [62e3926e@gateway/web/freenode/ip.98.227.146.110] has quit [Quit: Page closed] | 01:16 | |
blackburn | wiking: here? | 01:21 |
wiking | blackburn: yes | 01:21 |
wiking | wazza? | 01:21 |
blackburn | wiking: I need some kind of help with QP task | 01:21 |
wiking | ok | 01:21 |
wiking | listining | 01:21 |
blackburn | problem is \min_s \sum_i (s_i - s*_i) | 01:22 |
blackburn | s.t. \sum_i s_i = k | 01:22 |
blackburn | 0 <= s_i <= 1 | 01:22 |
wiking | u need a solver for that? | 01:22 |
blackburn | I am confused | 01:22 |
blackburn | is it a QP task? | 01:22 |
wiking | about? | 01:22 |
blackburn | argh ()^2 | 01:23 |
blackburn | not () | 01:23 |
wiking | ah ok | 01:23 |
wiking | \min_s \sum_i (s_i - s*_i)^2 | 01:23 |
wiking | right? | 01:23 |
blackburn | QP task is x^T H x + f x | 01:23 |
blackburn | yes | 01:23 |
blackburn | so what is H and f here? | 01:23 |
blackburn | if it is a qp | 01:23 |
wiking | u mean u wanna solve this with libqp? | 01:23 |
blackburn | yeah fits perfectly exepct H part | 01:24 |
blackburn | :D | 01:24 |
wiking | :DDD | 01:24 |
blackburn | I actually even have mosek calling code for that | 01:24 |
wiking | for which? | 01:25 |
blackburn | for solution of this problem | 01:25 |
wiking | aaah and you wanna have it instead with libqp? | 01:26 |
blackburn | okay actually it is (x-y)*I*(x-y)^T, right? | 01:26 |
blackburn | yeah would be cool to avoid mosek | 01:26 |
wiking | mmm | 01:26 |
wiking | ok | 01:28 |
wiking | only question is what is s*_i? | 01:28 |
blackburn | it is given | 01:28 |
wiking | ah cool | 01:28 |
wiking | so you only have s as a vector | 01:28 |
blackburn | everything is known | 01:28 |
wiking | and another s* vector | 01:28 |
blackburn | except s | 01:28 |
blackburn | yeah | 01:28 |
wiking | and wanna minimize your s* | 01:28 |
wiking | i mean s | 01:28 |
wiking | ;p | 01:29 |
wiking | ok get is | 01:29 |
wiking | get it | 01:29 |
blackburn | the other option is to use nlopt but I am confused whether it can be considered as qp | 01:29 |
wiking | well | 01:30 |
wiking | x = s | 01:30 |
wiking | x^T = s* | 01:30 |
wiking | and define Q that it creates s-s* | 01:30 |
wiking | only question here with the ^2 part | 01:31 |
wiking | but the problem is that in qp definition | 01:31 |
wiking | you have actually xHx^t | 01:32 |
wiking | so that x is the same | 01:32 |
wiking | :S | 01:32 |
blackburn | hmm wait | 01:33 |
blackburn | what if x = s - s* | 01:33 |
blackburn | then H = identity | 01:33 |
wiking | yes | 01:34 |
blackburn | \sum_i x_i = k - \sum_i s*_i | 01:34 |
wiking | u can do that | 01:34 |
wiking | H is positive semi-def in that way | 01:34 |
blackburn | s*_i < x_i < 1+s*_i | 01:34 |
blackburn | does it make sense to you? | 01:34 |
wiking | it was 0 < s_i < 1 | 01:35 |
wiking | right? | 01:35 |
blackburn | yeah | 01:35 |
wiking | yep | 01:35 |
wiking | try it | 01:36 |
wiking | :) | 01:36 |
blackburn | looks legit | 01:36 |
blackburn | wiking: funny thing is that problem size is usually <100 here | 01:44 |
blackburn | but I just don't know how to work with mosek :D | 01:45 |
wiking | :P | 01:54 |
wiking | i dont approve mosek | 01:54 |
wiking | http://imgur.com/a/ng9XT | 01:56 |
wiking | :>>> | 01:56 |
blackburn | birds are like sponges for oil :D | 01:57 |
-!- blackburn [~blackburn@109.226.92.17] has left #shogun [] | 03:04 | |
-!- ckwidmer [45c9b18c@gateway/web/freenode/ip.69.201.177.140] has joined #shogun | 03:46 | |
-!- ckwidmer [45c9b18c@gateway/web/freenode/ip.69.201.177.140] has quit [Client Quit] | 03:47 | |
shogun-buildbot | build #33 of nightly_default is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_default/builds/33 | 03:55 |
shogun-buildbot | build #23 of nightly_all is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_all/builds/23 | 04:23 |
-!- splade [~chatzilla@118.186.207.147] has joined #shogun | 06:33 | |
-!- gsomix [~gsomix@188.168.3.115] has quit [Ping timeout: 248 seconds] | 07:29 | |
-!- pluskid [~pluskid@111.120.81.118] has joined #shogun | 08:12 | |
-!- gsomix [~gsomix@r206-10.smr.ru] has joined #shogun | 08:46 | |
-!- splade [~chatzilla@118.186.207.147] has quit [Ping timeout: 252 seconds] | 08:48 | |
-!- uricamic [~uricamic@2001:718:2:1634:7de6:7891:7fa4:5d68] has joined #shogun | 08:48 | |
-!- splade [~splade@117.79.233.241] has joined #shogun | 08:48 | |
-!- gsomix [~gsomix@r206-10.smr.ru] has quit [Read error: Connection reset by peer] | 08:53 | |
-!- gsomix [~gsomix@r206-10.smr.ru] has joined #shogun | 08:54 | |
gsomix | good morning | 08:54 |
gsomix | droopy, nya | 08:54 |
droopy | gsomix! | 08:55 |
-!- gsomix_ [~gsomix@r206-10.smr.ru] has joined #shogun | 09:06 | |
-!- gsomix [~gsomix@r206-10.smr.ru] has quit [Read error: Connection reset by peer] | 09:06 | |
-!- gsomix_ is now known as gsomix | 09:32 | |
-!- pluskid [~pluskid@111.120.81.118] has quit [Ping timeout: 248 seconds] | 09:40 | |
-!- pluskid [~pluskid@202.130.113.141] has joined #shogun | 09:40 | |
-!- pluskid [~pluskid@202.130.113.141] has quit [Ping timeout: 276 seconds] | 10:15 | |
-!- pluskid [~pluskid@111.120.81.118] has joined #shogun | 10:35 | |
-!- blackburn [~blackburn@109.226.92.17] has joined #shogun | 11:29 | |
-!- gsomix [~gsomix@r206-10.smr.ru] has quit [Remote host closed the connection] | 11:31 | |
blackburn | uricamic: hey, around? | 12:08 |
-!- splade [~splade@117.79.233.241] has left #shogun [] | 12:18 | |
uricamic | blackburn: yep | 12:39 |
blackburn | uricamic: I have some struggles with libqp | 12:39 |
uricamic | go on | 12:40 |
blackburn | I am solving following problem | 12:40 |
blackburn | \min_s \sum_i (s_i - s*_i)^2, s.t. \sum_i s_i = k, 0 <= s_i <= 1 | 12:40 |
blackburn | s* and k are given | 12:40 |
uricamic | ok | 12:42 |
blackburn | so I introduced x = s - s* | 12:42 |
blackburn | then the problem became | 12:42 |
blackburn | \min_x \sum_i x_i^2 s.t. \sum_i x_i = k - \sum_i s*_i | 12:42 |
blackburn | 0-s*_i <= x <= 1-s*_i | 12:43 |
blackburn | uricamic: does it look ok? | 12:43 |
blackburn | I am unsure such thing is allowed here | 12:43 |
uricamic | it seems to be ok | 12:43 |
uricamic | but I have to look on it more carefully | 12:44 |
blackburn | uricamic: main problem is that I got solution not in bounds with gsmo for some reason | 12:44 |
uricamic | aha | 12:45 |
uricamic | unfortunately Vojta is not here, so I can't ask him, but I will try to look on it more carefully | 12:46 |
blackburn | it sets bounds | 12:46 |
blackburn | [-0.5,0.5] | 12:46 |
blackburn | argh bad example | 12:46 |
blackburn | it looks like solution stays at initial point here :D | 12:46 |
uricamic | :) | 12:47 |
blackburn | uricamic: does it depend on initial guess? | 12:47 |
uricamic | not sure, what does the return value says? | 12:47 |
blackburn | 1 iteration, KKT | 12:47 |
blackburn | objective = 0.0 | 12:47 |
blackburn | so | 12:49 |
blackburn | lb = -0.5 | 12:49 |
uricamic | I mean the exitflag | 12:49 |
blackburn | ub = 0.5 | 12:49 |
blackburn | exitflag = 4 | 12:49 |
blackburn | b = 1.0 | 12:49 |
blackburn | but solution is x = [0,0] | 12:49 |
uricamic | I see, I am not sure if that exitflag is good | 12:50 |
uricamic | but I have been just using libqp_splx only so far | 12:50 |
blackburn | uricamic: well it is the only non-emergency code here | 12:51 |
blackburn | argh no wrong | 12:51 |
blackburn | -1 is no memory available | 12:51 |
blackburn | 0 is max iterations code | 12:51 |
blackburn | and 4 is KKT satisfied code | 12:51 |
uricamic | I see, so it should be ok, but it is strange though | 12:52 |
uricamic | in gsmo the exitflag cannot be 1, 2 or 3, tight? | 12:53 |
uricamic | right? | 12:53 |
blackburn | yeah it seems so | 12:53 |
uricamic | in that case 4 should be ok | 12:53 |
blackburn | I set initial x to [10,10] | 12:54 |
blackburn | and it did nothing with it actually | 12:54 |
blackburn | I think there is some illformed vector here | 12:55 |
uricamic | I am looking on the webpage of libqp now, and it seems that your constraints are not ok for it now | 12:55 |
uricamic | there is just x^T a = b | 12:56 |
blackburn | yeah but what is wrong with it? | 12:56 |
blackburn | I set a to ones | 12:56 |
uricamic | but in your task you said that constraints are sum_i x_i = k - sum_i s^*_i | 12:57 |
blackburn | but s* is given | 12:57 |
uricamic | I see, ok | 12:58 |
blackburn | so I just b = k - sum_i s*_i | 12:58 |
uricamic | sure | 12:58 |
uricamic | ok, in that case it should be ok | 12:58 |
uricamic | have you tried without using initial solution? | 12:58 |
blackburn | how to do that? | 12:59 |
blackburn | I just tried x = [1000 , 1000] and libqp made it a solution.. strange | 12:59 |
uricamic | it depends also how have you set the TolKKT parameter | 13:00 |
blackburn | in doc here | 13:00 |
blackburn | I set it to 1e-6 | 13:00 |
uricamic | and of course on the initial solution provided | 13:00 |
blackburn | it is stated here that x 'must be feasible' | 13:00 |
blackburn | does it mean initial solution should satisfy bound? | 13:01 |
uricamic | sure, I guess | 13:03 |
blackburn | should it satisfy a'x=b then? | 13:03 |
uricamic | yep | 13:04 |
blackburn | uh that is another optimization problem then :D | 13:05 |
droopy | blackburn, ^_^ | 13:05 |
uricamic | :) | 13:05 |
blackburn | uricamic: but seriously - how to do that initial guess | 13:07 |
-!- yoo [575b08cb@gateway/web/freenode/ip.87.91.8.203] has joined #shogun | 13:07 | |
-!- yoo [575b08cb@gateway/web/freenode/ip.87.91.8.203] has quit [Client Quit] | 13:07 | |
blackburn | hmmm apparently upper bound would work in that case it seems | 13:08 |
blackburn | anyway, thanks uricamic | 13:10 |
uricamic | no problem blackburn | 13:11 |
CIA-18 | shogun: Sergey Lisitsyn master * r759fdb7 / (17 files in 4 dirs): Refactored MALSAR-based multitask algorithms and added clustered multitask logistic regression - http://git.io/Nk_Abw | 13:18 |
-!- blackburn [~blackburn@109.226.92.17] has left #shogun [] | 13:22 | |
wiking | uricamic: here? | 14:34 |
uricamic | yep | 14:35 |
-!- pluskid [~pluskid@111.120.81.118] has quit [Ping timeout: 248 seconds] | 15:07 | |
-!- pluskid [~pluskid@202.130.113.141] has joined #shogun | 15:08 | |
-!- ermutarra [55578d64@gateway/web/freenode/ip.85.87.141.100] has joined #shogun | 15:12 | |
ermutarra | hi | 15:12 |
-!- ermutarra [55578d64@gateway/web/freenode/ip.85.87.141.100] has left #shogun [] | 15:14 | |
-!- pluskid [~pluskid@202.130.113.141] has quit [Read error: Connection reset by peer] | 15:19 | |
-!- uricamic [~uricamic@2001:718:2:1634:7de6:7891:7fa4:5d68] has quit [Quit: Leaving.] | 16:16 | |
-!- heiko [~heiko@host86-177-183-33.range86-177.btcentralplus.com] has joined #shogun | 16:17 | |
-!- gsomix [~gsomix@37.61.181.69] has joined #shogun | 17:04 | |
gsomix | good evening | 17:04 |
heiko | sonney2k, are you around? | 18:03 |
-!- os252 [~stegle@e179243147.adsl.alicedsl.de] has joined #shogun | 18:36 | |
-!- heiko [~heiko@host86-177-183-33.range86-177.btcentralplus.com] has quit [Quit: Leaving.] | 19:04 | |
-!- blackburn [~blackburn@109.226.92.17] has joined #shogun | 19:34 | |
-!- blackburn [~blackburn@109.226.92.17] has quit [Quit: Leaving.] | 19:54 | |
-!- puffin444 [a6f986d8@gateway/web/freenode/ip.166.249.134.216] has joined #shogun | 20:23 | |
puffin444 | hello os252 | 20:23 |
os252 | hi puffin444 | 20:24 |
puffin444 | how are things? | 20:24 |
os252 | ok, still little time due to family matters. Sorry for having been difficult to reach. | 20:25 |
os252 | How are things on your end? | 20:25 |
puffin444 | that's okay | 20:25 |
os252 | good. | 20:25 |
os252 | Fortunately you are on top of everything, so I am not really needed ;-) | 20:26 |
puffin444 | good finally done with the gradient stuff | 20:26 |
os252 | Nice,so including all the ARD variants ? | 20:26 |
puffin444 | now working on some vis. demos and fitc | 20:26 |
puffin444 | yes the ard works nicely | 20:26 |
os252 | perfect | 20:27 |
os252 | its separate kernels? | 20:27 |
puffin444 | there are ard linear and gaussian kernels | 20:27 |
os252 | yeah, I guess that's all for nwo. | 20:27 |
os252 | I was wondering whether it's worth porting the matern kernel. | 20:27 |
puffin444 | the weights can be learned thtough gradient search | 20:27 |
os252 | It is a bit rougher and not quite as smooth as Gaussian. | 20:27 |
os252 | Statisticians like it... | 20:27 |
os252 | and has some good properties. | 20:28 |
os252 | But I'd put somewhere further down on the list - there are probably more exciting things than implementing exotic kernels :-). | 20:28 |
puffin444 | interesting. never heard of that kernel | 20:28 |
os252 | The periodic kernels are also sort of cool. | 20:28 |
os252 | If you want to get inspiration: | 20:29 |
os252 | http://www.gaussianprocess.org/gpml/chapters/RW4.pdf | 20:29 |
os252 | (4.31) is a cool kernel. | 20:29 |
os252 | It basically is like a squared exponential with prior knowledge that the function you are regressing is periodic. | 20:30 |
os252 | It\s cool to do spectral decomposition form a just a handful of points;you can optimize the hyper parameter that corresponds to the period length and read of what's happening. | 20:30 |
puffin444 | hmm... I was hoping to implement fitc/laplacian first if thats okay | 20:31 |
os252 | Could be means to get good demos - but you really decide what you like, there is infinite freedom to put some cool demos together. | 20:31 |
-!- blackburn [~blackburn@109.226.92.17] has joined #shogun | 20:32 | |
os252 | Totally. I just thought I through up some ideas for demos that illustrate what you can do with GPs and cannot easily with SVMs etc. | 20:32 |
puffin444 | that way adding exotic kernels/likelihoods would be an easy task | 20:32 |
puffin444 | yes i am looking at gpml for inspirarion for the denos | 20:32 |
puffin444 | demos* | 20:32 |
os252 | Agree - and as I already pushed for the ARD stuff, let's defer that one for later. But take a look at that stuff. | 20:33 |
puffin444 | demos* | 20:33 |
puffin444 | i am hoping to do something similar via matploylib | 20:33 |
os252 | What have you got so far? | 20:33 |
puffin444 | matplotlib | 20:33 |
puffin444 | frankly i dont have demos yet, been looking at fitc | 20:34 |
os252 | Yes, I understand - more fun to do. | 20:35 |
puffin444 | but i plan to have at least some simple demos by the rnd of the week | 20:35 |
os252 | I have some old code for sausage plots in matplotlib | 20:35 |
os252 | https://github.com/PMBio/pygp/blob/master/pygp/plot/gpr_plot.py | 20:35 |
os252 | some super simple simulation, but it's similar to things in gpml | 20:35 |
os252 | is here | 20:35 |
os252 | https://github.com/PMBio/pygp/blob/master/pygp/demo/demo_gpr.py | 20:35 |
os252 | As suggested before. I'd vote for simple demos now and more complex things at the end of all this. | 20:36 |
os252 | Have you read the unfitting sparse GP approximations paer? | 20:36 |
puffin444 | yes these might be quite useful | 20:36 |
os252 | I think it will be very helpful to get an efficient implementation of a general framework for sparse approximations (including fits). | 20:37 |
os252 | *fitc | 20:37 |
puffin444 | yes i have looked at it briefly a fee times | 20:37 |
puffin444 | gew* | 20:37 |
puffin444 | few* lol | 20:37 |
os252 | :-) | 20:37 |
os252 | It may make sense to discuss this if you like. | 20:38 |
os252 | The approximations are certainly non-trivial. | 20:38 |
os252 | and take a bit of time to get into. | 20:38 |
puffin444 | there was a paper i found that compared many sparse methods, but i think its safest now to look at fitc | 20:38 |
os252 | yes, there are several comparisons. FITc is definitely high up in the ranking. | 20:39 |
puffin444 | as i can compare it directly to gpml | 20:39 |
os252 | Yeah, it's brill that they have this all implemented :-) | 20:40 |
puffin444 | lots of potential math error implementing this | 20:40 |
os252 | But let me know if you want to discuss. The initialization of the pseudo points etc. can be a bit of a tricky business. | 20:40 |
os252 | Another question is how to handle the pseudo points. | 20:40 |
puffin444 | yes its wonderful they have it in gpml | 20:40 |
os252 | Effectively they are parameters like everything else in the covariance. | 20:41 |
os252 | The question is how to best incorporate this into your framework. | 20:41 |
os252 | Definitely. That's why C++ is not the best prototyping language (at least for me). | 20:41 |
-!- puffin444_ [a6f986d8@gateway/web/freenode/ip.166.249.134.216] has joined #shogun | 20:43 | |
puffin444_ | yes it is something that i have been thinking about | 20:44 |
-!- puffin444 [a6f986d8@gateway/web/freenode/ip.166.249.134.216] has quit [Ping timeout: 245 seconds] | 20:45 | |
os252 | B.t.w Sergey has switched shogun to Eigen, right? | 20:45 |
os252 | You could consider coding up the maths for this stuff in Eigen. | 20:45 |
os252 | It will be a lot less pain than native blas/lapack calls. | 20:45 |
puffin444_ | yes that would be wonderful | 20:45 |
os252 | I am not aware of the latest state of the play but this: http://comments.gmane.org/gmane.comp.lib.eigen/3344 suggests it's in there yet. | 20:46 |
blackburn | os252: moreover I am already using it for my algorithms | 20:46 |
os252 | I think Eigen is superb in particular if you are porting code from matlaby/python like scripting languages. You can almost copy & paste thins. | 20:46 |
os252 | blackburn: are the libraries copied into shogun? | 20:47 |
os252 | or is there an external dependency? | 20:47 |
blackburn | os252: no, we rely on dependency | 20:47 |
blackburn | it can be a problem though | 20:47 |
os252 | is it optional or a requirement, i.e. can core code use Eigen ? | 20:47 |
puffin444_ | yes this might make things muxh rasier | 20:47 |
blackburn | what is core? | 20:47 |
os252 | Well, not sure. GP support of course ;-) | 20:48 |
puffin444_ | well already gp only works with lapack installed | 20:48 |
blackburn | I don't actually know whether we can make it mandatory | 20:49 |
puffin444_ | is there a performance difference with eigen | 20:49 |
blackburn | yes, and performance of coding in particular | 20:49 |
os252 | Yeah, numerically it's bit faster than black/lapack, coding is orders of magnitude ;-) | 20:50 |
blackburn | puffin444_: it uses SSE stuff if you know what it is | 20:50 |
puffin444_ | oh wow, so do we keep lapack just because it runs on more pkayforms? | 20:50 |
os252 | I think they really tackled the last bottle necks. You can even now put Intel MKL underneath to get multi-core support for large matrices (which Eigen does not have yet). | 20:50 |
blackburn | puffin444_: we still have lapack code.. | 20:51 |
os252 | but Eigen has lapak interfaces. Blackburn: have you explored whether those are sufficient for Shogun? | 20:51 |
blackburn | os252: problem is that they are provided only in standalone packages | 20:52 |
blackburn | so if you install it on say debian/ubuntu | 20:52 |
blackburn | you get no blas/lapack headers provided | 20:52 |
os252 | but otherwise they work? | 20:52 |
os252 | If yes: I am really wondering why not copy all of Eigen into shogun and kick all the lapas/black dependencies. | 20:52 |
blackburn | let me check | 20:53 |
os252 | Linalg is sufficiently "central" to justify to few kilobyte of third party code I guess. | 20:53 |
blackburn | I am not big fan of integrating code actually | 20:53 |
blackburn | eigen3 is being developed pretty fast still | 20:54 |
os252 | It surely is a tradeoff - I completely understand. | 20:54 |
blackburn | lapack and blas haven't changed for years | 20:55 |
os252 | But either way: if you add stuff to shogun that depends on Eigen I see now reason why puffin444 should code up sophisticated PG linalg in lapack. | 20:55 |
blackburn | no just guard it with HAVE_EIGEN3 | 20:55 |
blackburn | it is not a problem actually | 20:56 |
os252 | vgood. | 20:56 |
os252 | The fitc stuff is pretty complicated and it will take a lot of time to code this up in lapak otherwise. | 20:56 |
os252 | *lapack | 20:56 |
os252 | puffin444: have you used Eigen before? Keen on going down that route? | 20:57 |
puffin444_ | yes especially the derivatives i found out that i have my work cut out for me | 20:57 |
blackburn | os252: puffin444_: see shogun/lib/malsar folder - a few examples of eigen3 code are here | 20:57 |
puffin444_ | okay | 20:57 |
os252 | Sorry. I was not aware that it's already sorted and ok to use Eigen in shogun; otherwise would have pointed this out earlier. Surely all the linalg you need is pretty messy... | 20:58 |
puffin444_ | Yes but it is workable | 20:59 |
os252 | I'd suggest to then use Eigen for the GP stuff. You can also mix and match as you see fit. | 20:59 |
puffin444_ | will do | 21:01 |
puffin444_ | os252, anything more you would like to discuss? | 21:02 |
blackburn | actually you can gain pretty much with eigen3 | 21:02 |
blackburn | e.g. afaik lapack has no LDL' | 21:03 |
blackburn | I remember you were using cholesky somewhere | 21:03 |
-!- cwidmer [8ca3fe9d@gateway/web/freenode/ip.140.163.254.157] has joined #shogun | 21:03 | |
blackburn | oh famous cwidmer is in | 21:03 |
blackburn | :) | 21:03 |
puffin444_ | yes actually everwhere :) | 21:03 |
droopy | 8) | 21:04 |
blackburn | puffin444_: LDLT is something more stable IIRC | 21:04 |
blackburn | works for semi-definite | 21:04 |
puffin444_ | okay | 21:04 |
os252 | Yes, Eigen has LDLT which is quite nice. It's also quite fast. | 21:05 |
os252 | I have some basic GP code for a very applied project in Eigen; I'd be happy to just send it. You could get some inspiration of Eigen for this purpose. | 21:05 |
puffin444_ | sure that would be great | 21:07 |
os252 | I'll just put it into an email and you'll see what you can use. | 21:07 |
puffin444_ | okay | 21:07 |
os252 | sent | 21:11 |
os252 | ok, anything else we should discuss now? | 21:12 |
os252 | Definitely ping when you need input on the sparse stuff. | 21:12 |
puffin444_ | nothing else on my mind | 21:12 |
puffin444_ | will do | 21:12 |
os252 | Great. Then happy coding! | 21:13 |
puffin444_ | ok! :) | 21:13 |
os252 | cu | 21:14 |
-!- os252 [~stegle@e179243147.adsl.alicedsl.de] has left #shogun [] | 21:15 | |
-!- os252 [~stegle@e179243147.adsl.alicedsl.de] has joined #shogun | 21:15 | |
@sonney2k | wiking, so what is the problem / where are you stuck? | 21:24 |
@sonney2k | gsomix, how is it going? | 21:24 |
gsomix | sonney2k, working and reading numpy code. | 21:24 |
wiking | sonney2k: so solver | 21:24 |
wiking | sonney2k: need to do some redesigning there to be able to handle both primal mosek and the bmrm solver | 21:24 |
wiking | sonney2k: had a chat today with uricamic about it, and we've agreed on some changes | 21:25 |
@sonney2k | wiking, but you know what to do? | 21:26 |
@sonney2k | ok | 21:26 |
wiking | yep | 21:26 |
wiking | now i'm trying it out | 21:26 |
@sonney2k | ok then | 21:27 |
@sonney2k | gsomix, still about the slicing / dictionary? | 21:27 |
gsomix | sonney2k, yes, this numpy code is not well commented. | 21:28 |
gsomix | so, I'm doing experiments (as physicist hehe) | 21:28 |
-!- emrecelikten [~emre@176.40.251.10] has joined #shogun | 21:32 | |
-!- puffin444_ [a6f986d8@gateway/web/freenode/ip.166.249.134.216] has left #shogun [] | 21:37 | |
@sonney2k | gsomix, but can't you determine the types of the PyObjects | 21:37 |
@sonney2k | wiking, is your PR ready to be merged or will it break things? | 21:38 |
@sonney2k | I meant will it still compile and other (unrelated) stuff still work. | 21:39 |
wiking | sonney2k: it compiled for me | 21:39 |
wiking | so it shouldn't break anything | 21:39 |
@sonney2k | wiking, ok then | 21:40 |
gsomix | sonney2k, I can and I know that [a:b, c:d] in my_cool_features[bla, bla] is tuple | 21:40 |
@sonney2k | wiking, btw could you figure out why the linking error occurs on *bsd | 21:40 |
wiking | sonney2k: not yet | 21:40 |
@sonney2k | gsomix, and the sub-elements? | 21:40 |
wiking | sonney2k: that's a really weird thing | 21:40 |
wiking | sonney2k: i think i have a conflict there... since most of the .so are compiled with cc and c++ (which is 4.2.x) | 21:41 |
gsomix | sonney2k, hm, at know I can say that it is not array or sequence. | 21:41 |
wiking | so that's why ldd shows that libshogun.so will depend on that other libstdc++ | 21:41 |
@sonney2k | wiking, hmmhh mixture of gcc 3.X and 4.X? | 21:46 |
wiking | yeah | 21:50 |
wiking | sonney2k: since we have to use 4.6 gcc on that machine because of that gcc bug | 21:50 |
wiking | but maybe i'm just doing something fishy there | 21:51 |
@sonney2k | wiking, well can you explicitly force gcc 4.6 for the examples then? | 21:52 |
gsomix | sonney2k, yes! | 21:53 |
@sonney2k | gsomix, yes?! | 21:53 |
gsomix | sub_elements are PySlices, huh. | 21:54 |
gsomix | obviously | 21:54 |
@sonney2k | ok then :) | 21:54 |
gsomix | but little deep in python docs. | 21:54 |
-!- rieck [~rieck@paul.ml.tu-berlin.de] has quit [Quit: ZNC - http://znc.sourceforge.net] | 21:54 | |
wiking | sonney2k: retrying again something | 21:55 |
wiking | gmake[3]: Leaving directory `/usr/home/wiking/shogun-buildbot/bsd1_-_libshogun/build/examples/undocumented/libshogun' | 22:01 |
wiking | Error expanding embedded variable. | 22:01 |
-!- rieck [~rieck@134.76.96.43] has joined #shogun | 22:01 | |
wiking | so sonney2k | 22:01 |
wiking | i don't think that this patch did the job | 22:01 |
wiking | https://github.com/shogun-toolbox/shogun/commit/98f4788015fdf04e71cd0a93748707a52e52a6c3 | 22:01 |
wiking | it's still calling make :( | 22:02 |
@sonney2k | wiking, please replace the test -n with test -z in check.sh | 22:04 |
@sonney2k | in the make line | 22:04 |
CIA-18 | shogun: Soeren Sonnenburg master * r96a80ed / examples/undocumented/libshogun/check.sh : use test -z not -n to use configured make - http://git.io/7-0mjA | 22:04 |
wiking | ok | 22:05 |
wiking | i've did this | 22:05 |
wiking | LIBRARY_PATH="LD_LIBRARY_PATH" LIB_PATH="//usr/local/lib" INC_PATH="//usr/local/include" LIBS="-Xlinker --no-undefined -lshogun" LD_LIBRARY_PATH=//usr/local/lib INCLUDES="-I. -I.. -I/usr/local/include/json -I/usr/local/include/libxml2 -I/usr/local/include -I../libshogun -I.." MAKE="gmake" ./check.sh | 22:05 |
wiking | make | 22:05 |
wiking | in check.sh i did an echo $MAke | 22:05 |
wiking | i mean echo $MAKE | 22:05 |
wiking | and it's make and not gmake | 22:05 |
wiking | which is very weird | 22:05 |
@sonney2k | wiking, well either git pull again or change the test -n "$MAKE" line | 22:09 |
@sonney2k | to test -z | 22:09 |
wiking | sonney2k: i'll wait now | 22:10 |
wiking | since it's building | 22:10 |
wiking | :) | 22:10 |
wiking | and i was using the build dir of the bot to do the tests | 22:10 |
wiking | 3 more mins | 22:12 |
wiking | sonney2k: is there a way to do gmake -j4 ? | 22:16 |
wiking | or soemthing like this | 22:16 |
wiking | sonney2k: ok so the problem is now the following | 22:17 |
wiking | ok let me check it for 100% before i start telling b.s. | 22:18 |
wiking | huh so first of all | 22:19 |
wiking | not the right c++ is used for building the libshogun examples | 22:19 |
@sonney2k | wiking, set MAKEFLAGS="-j4" | 22:19 |
wiking | sonney2k: i mean that u should change that in the buildbot | 22:20 |
wiking | since there's really free cpu there in the machine | 22:20 |
wiking | so | 22:20 |
wiking | CC=c++ | 22:20 |
wiking | in examples/undocumented/libshogun/Makefile | 22:20 |
wiking | which will not be ok, since i've used c++46 for building the libshogun | 22:21 |
wiking | so that should be set as well by ./configure script not just when building libshogun itself | 22:21 |
wiking | but the examples as well | 22:21 |
wiking | but even then i'm seeing some weird shit | 22:21 |
wiking | LD_LIBRARY_PATH=/usr/home/wiking/shogun-buildbot/bsd1_-_libshogun/build/src/shogun ./basic_minimal | 22:22 |
wiking | this give me | 22:22 |
wiking | /usr/lib/libstdc++.so.6: version GLIBCXX_3.4.11 required by /usr/home/wiking/shogun-buildbot/bsd1_-_libshogun/build/src/shogun/libshogun.so.12 | 22:22 |
wiking | i'm trying to find info about this | 22:23 |
wiking | ok i think i see wtf | 22:24 |
wiking | just a sec i'll try to experiment with some extra compiler flags | 22:25 |
wiking | ;) | 22:25 |
wiking | ok i think the problem occures here when it's linking the libshogun.so | 22:28 |
wiking | that it doesn't link it to the right libstdc++.so.6 | 22:28 |
-!- os252 [~stegle@e179243147.adsl.alicedsl.de] has left #shogun [] | 22:29 | |
gsomix | >> You probably do not want to use this function. | in python docs :) | 22:30 |
wiking | woah i've just seen in a printed book a screenshot of netscape navigator :DDD | 22:30 |
wiking | gsomix: then u most defintely should use it :) | 22:31 |
droopy | cool!?! ;> | 22:31 |
gsomix | droopy, what do you think about higgs boson? | 22:32 |
droopy | i'll re-pm you on this serv if pm is actually 'private'!?! | 22:32 |
* gsomix sighs | 22:33 | |
CIA-18 | shogun: Soeren Sonnenburg master * rfa6ce12 / (2 files in 2 dirs): pass COMP_CPP through for compiling libshogun examples - http://git.io/sudcZg | 22:38 |
@sonney2k | wiking, another attempt... | 22:39 |
wiking | sonney2k: okey! will check | 22:39 |
wiking | or wait till it builds | 22:39 |
@sonney2k | yeah better that then | 22:40 |
@sonney2k | wiking, btw do you have some nice graphics produced w/ the help of shogun? | 22:40 |
@sonney2k | I am trying to create some kind of gallery | 22:40 |
@sonney2k | illustrating various applications/use cases/benchmarks whatever for the new website | 22:41 |
wiking | sonney2k: will do it for you | 22:41 |
blackburn | I have one idea but I need someone blender skilled :D | 22:41 |
droopy | grin ^_^ | 22:41 |
wiking | sonney2k: i'll create some screenshots for the latent svm | 22:41 |
wiking | sonney2k: but we can even for fun make a webservice :) | 22:42 |
@sonney2k | blackburn, heh | 22:42 |
@sonney2k | wiking, you mean with security audited code like shogun's :D | 22:42 |
wiking | sonney2k: let say there's a cat regonizer :) and people can just upload any images to see if it locates the cat (if there's one) in the pic | 22:42 |
@sonney2k | rieck, woudl cry :D | 22:42 |
wiking | sonney2k: DDDD | 22:42 |
wiking | sonney2k: :> | 22:42 |
wiking | well then just screenshots | 22:42 |
@sonney2k | wiking, well ok that is certainly safe | 22:43 |
@sonney2k | well to some extend | 22:43 |
wiking | blackburn: :>>>>> | 22:43 |
@sonney2k | (bugs in image parsing lib) | 22:43 |
@sonney2k | I think such examples are really highly needed | 22:43 |
@sonney2k | what shogun can do is too abstract for everyone... | 22:44 |
@sonney2k | except the authors :D | 22:44 |
wiking | :P | 22:44 |
wiking | ok i think i'll need to add this flag to the linker | 22:44 |
blackburn | I still convinced shogun is used by ten guys | 22:44 |
wiking | -Wl,-rpath,/usr/local/lib/gcc46 | 22:44 |
wiking | but i'll test this now | 22:44 |
wiking | blackburn: :DDD | 22:44 |
blackburn | worst thing is that we have outdated version in ubuntu/debian | 22:45 |
wiking | blackburn: dude i think the train set is down! :D | 22:50 |
blackburn | wiking: what do you mean? | 22:50 |
wiking | it downloaded :))) | 22:50 |
wiking | that 135 gigs | 22:50 |
blackburn | oooohh | 22:51 |
blackburn | finally | 22:51 |
blackburn | okay now i have no idea what to do with it | 22:51 |
blackburn | :D | 22:51 |
blackburn | I mean RealFeatures(traindat) wouldn't work here LOL | 22:51 |
@sonney2k | blackburn, no just I use shogun - you all only work for me *eg* | 22:55 |
blackburn | sonney2k: yeah I tend to believe ;) | 22:55 |
wiking | blackburn: well now i'll upload the whole thing into the hdfs on the cluster | 22:59 |
wiking | and then we can run on it feature extractors | 22:59 |
wiking | you've said you'll create a c based hog extractor | 23:00 |
blackburn | aham I have to hurry up with hogs | 23:00 |
wiking | i'll port that retina or what the fuck was that descriptr :) | 23:00 |
blackburn | yeah FREAK | 23:00 |
blackburn | I didn't call you freak :D | 23:01 |
wiking | freak | 23:01 |
wiking | :) | 23:01 |
wiking | that's the one | 23:01 |
wiking | sonney2k: hmmm did that run well with other builders? | 23:02 |
wiking | i mean libshogun build this time | 23:02 |
wiking | ok | 23:07 |
wiking | sonney2k: almost... ;) | 23:07 |
shogun-buildbot | build #176 of deb3 - modular_interfaces is complete: Failure [failed test libshogun] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/176 blamelist: Soeren Sonnenburg <sonne@debian.org> | 23:07 |
wiking | error in gmake basic_minimal CC=c++46 LIBRARY_PATH=LD_LIBRARY_PATH LIB_PATH=//usr/home/wiking/shogun-buildbot/bsd1_-_libshogun/build/src/build/lib INC_PATH=//usr/home/wiking/shogun-buildbot/bsd1_-_libshogun/build/src/build/include LIBS=-Xlinker --no-undefined -lshogun | 23:07 |
wiking | error in ./basic_minimal | 23:07 |
wiking | ================================================================================ | 23:07 |
wiking | gmake: unrecognized option `--no-undefined' | 23:07 |
wiking | sonney2k: oh ok it's the same for you apparently | 23:08 |
shogun-buildbot | build #152 of deb2 - static_interfaces is complete: Failure [failed test libshogun] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb2%20-%20static_interfaces/builds/152 blamelist: Soeren Sonnenburg <sonne@debian.org> | 23:08 |
@sonney2k | wiking, grmpf | 23:15 |
CIA-18 | shogun: Soeren Sonnenburg master * r2442926 / examples/undocumented/libshogun/check.sh : escape flags - http://git.io/0YZxtg | 23:15 |
shogun-buildbot | build #153 of deb2 - static_interfaces is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb2%20-%20static_interfaces/builds/153 | 23:23 |
blackburn | wiking: hmm it looks like my substitution we discussed last night was wrong | 23:24 |
blackburn | sonney2k: are you busy with that building stuff? | 23:27 |
blackburn | wiking: probably I do not cover negative and positive cases when I transform s.t. part | 23:30 |
gsomix | sonney2k, ta dam! | 23:33 |
rieck | /msg *status loadmod colloquy | 23:37 |
rieck | rah! | 23:37 |
wiking | :) | 23:41 |
wiking | sonney2k: ok | 23:53 |
shogun-buildbot | build #177 of deb3 - modular_interfaces is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/177 | 23:54 |
wiking | sonney2k: i have the solution... but i really don't know how to add this into the configure script | 23:54 |
wiking | so basically i would need to give this compiler flag: -Wl,-rpath,/usr/local/lib/gcc46 | 23:54 |
wiking | i'll try to see how's this solved usually | 23:55 |
wiking | with the configure script | 23:55 |
wiking | because now for some reason when c++46 is doing the linking | 23:56 |
wiking | it uses the standard /usr/lib/ path for the libgcc and libstdc++ libs | 23:56 |
wiking | and that'll break of course | 23:56 |
wiking | since it's for c++ (which is 4.2..) | 23:56 |
wiking | this is the only thing left to be solved and then we can enable libshogun tests on freebsd | 23:57 |
gsomix | good night guys | 23:59 |
@sonney2k | wiking, yeah but that is wrong | 23:59 |
@sonney2k | gsomix, tada? | 23:59 |
gsomix | sonney2k, ta dam. | 23:59 |
gsomix | works | 23:59 |
@sonney2k | gsomix, cool | 23:59 |
gsomix | my_cool_features[a,b:c,d] | 23:59 |
@sonney2k | PR'd alreadY? | 23:59 |
gsomix | sonney2k, PR will be tomorrow. | 23:59 |
--- Log closed Fri Jul 20 00:00:05 2012 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!