--- Log opened Fri Jun 22 00:00:20 2012 | ||
--- Day changed Fri Jun 22 2012 | ||
@sonney2k | yeah the guy at the very left that appears a couple of times is a ML person | 00:00 |
---|---|---|
blackburn | heh | 00:00 |
@sonney2k | there are others but hard to descibe :D | 00:01 |
-!- blackburn [d557894d@gateway/web/freenode/ip.213.87.137.77] has quit [Quit: Page closed] | 00:47 | |
-!- zxtx [~zv@cpe-75-83-151-252.socal.res.rr.com] has joined #shogun | 06:11 | |
-!- uricamic [~uricamic@89.13.broadband14.iol.cz] has joined #shogun | 09:11 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 11:04 | |
-!- gsomix [~gsomix@188.168.2.228] has joined #shogun | 11:06 | |
-!- vojtech_ [9320543b@gateway/web/freenode/ip.147.32.84.59] has joined #shogun | 11:34 | |
n4nd0 | sonney2k, alexlovesdata: about comparing the results of my SO code with SVM-struct for multiclass | 11:43 |
n4nd0 | I am not sure if it is possible since they don't have exactly the same formulation of the problem | 11:43 |
n4nd0 | sonney2k: can it be reliable to compare results of the primal formulation with results of the dual? | 11:44 |
@sonney2k | n4nd0, well now that uricamic merged his MC example based on his BMRM implementation it is probably easier to compare against that | 11:45 |
@sonney2k | vojtech_, am I right? | 11:45 |
n4nd0 | sonney2k: I found something that might work though, they provide code to solve a 1-slack formulation of the problem in the primal | 11:46 |
n4nd0 | sonney2k: what we have is a n-slack formulation | 11:46 |
@sonney2k | n4nd0, btw is nico on vacations? | 11:47 |
n4nd0 | sonney2k: BUT in the paper they say that a w solution of one problem is solution of the other | 11:47 |
@sonney2k | n4nd0, IIRC this is all equivalent | 11:47 |
vojtech_ | sonney2k: yes, Michal reimplemented BMRM which we have been using many times and it seems to work correctly | 11:47 |
n4nd0 | sonney2k: he said that his vacations will start from the middle of the next month | 11:47 |
n4nd0 | sonney2k: btw, will there be a problem with that and this mid-term report? | 11:49 |
@sonney2k | n4nd0, so you should work with uricamic to compare your mosek based code against his impl. | 11:49 |
@sonney2k | you should both get very similar objective | 11:50 |
n4nd0 | sonney2k: that sounds better, I am going to talk to him | 11:50 |
@sonney2k | and then all is good | 11:50 |
@sonney2k | vojtech_, thanks - as I understand uricamic is now working on the new fancy thing right? | 11:50 |
@sonney2k | vojtech_, ahh and btw if you have some nice application for this SO stuff (something fancy/graphical) it would be cool to have it as illustration... | 11:51 |
vojtech_ | sonney2k, once he finishes implenmentation of plain BMRM (which is probably already done) he will implement original Bundle method - I don't know if this is sufficiently fancy but it works faster :) | 11:52 |
@sonney2k | vojtech_, well I meant something like word recognition / licence plate / image segmentation ... | 11:53 |
vojtech_ | yes, we have OS library for face landmark detection see: http://cmp.felk.cvut.cz/~uricamic/flandmark/ | 11:53 |
vojtech_ | license plate will not be public but we have lot of StructSVM examplables for image segmentation etc. | 11:54 |
@sonney2k | vojtech_, what I meant is some fancy example that can be put into shogun | 11:55 |
@sonney2k | you know better what is easy / complex enough :D | 11:55 |
vojtech_ | ok, I'll think about it | 11:56 |
@sonney2k | n4nd0, btw if you are stuck for too long - investigate intensively with nico | 11:59 |
n4nd0 | sonney2k: he told me he's preparing a document with HMSVM info that should be ready by the end of this week | 11:59 |
n4nd0 | sonney2k: until that, I want to test the so-multiclass and get familiarized with the Viterbi impl. in Shogun | 12:00 |
@sonney2k | n4nd0, for how long are you trying to figure out whether the MC - SO SVM is correct? | 12:11 |
n4nd0 | sonney2k: I have been with it basically this week and the last one ... | 12:12 |
n4nd0 | I think it has been a loss of time to tell the truth | 12:12 |
@sonney2k | n4nd0, yeah so you should have a meeting with nico, uricamic *ASAP* | 12:13 |
-!- pluskid [~pluskid@202.130.113.141] has joined #shogun | 12:13 | |
@sonney2k | n4nd0, email them asking them if they have time this afternoon and CC me... | 12:13 |
n4nd0 | sonney2k: I have no time this afternoon | 12:14 |
@sonney2k | n4nd0, well when do you have time now? | 12:14 |
n4nd0 | sonney2k: today it is midsommar in Sweden, national holiday | 12:14 |
@sonney2k | n4nd0, what I am saying is ASAP - otherwise you will become demotivated | 12:14 |
@sonney2k | trying for 2 weeks to find out what is wrong is just to long | 12:15 |
@sonney2k | too | 12:15 |
@sonney2k | 1 week is already very tiring | 12:15 |
n4nd0 | sonney2k: yeah | 12:15 |
n4nd0 | but Nico told me to wait until he has ready the file with the new work to do | 12:15 |
@sonney2k | n4nd0, so you definitely need help here! | 12:15 |
@sonney2k | n4nd0, no | 12:15 |
@sonney2k | the document can wait | 12:16 |
@sonney2k | n4nd0, or do you have something else to do? | 12:16 |
n4nd0 | sonney2k: apart from shogun you mean? | 12:16 |
@sonney2k | n4nd0, no for GSoC I mean | 12:16 |
n4nd0 | no | 12:17 |
@sonney2k | and the viterbi in shogun is very hard to digest | 12:17 |
@sonney2k | even worse I would say | 12:17 |
n4nd0 | the plan was to check viterbi | 12:17 |
n4nd0 | mmm but I'll need to use it in any case | 12:17 |
n4nd0 | maybe I should get familiarized with it at least | 12:17 |
@sonney2k | n4nd0, I tend to say that writing your own is easier! | 12:17 |
@sonney2k | the one in shogun is very specialized | 12:18 |
@sonney2k | and not just for HM-SVM | 12:18 |
@sonney2k | but for semi-HM SVM | 12:18 |
n4nd0 | ok | 12:18 |
@sonney2k | -> much more difficult | 12:18 |
@sonney2k | n4nd0, ask nico and uricamic now! | 12:18 |
n4nd0 | yeah, I studied your paper about it a couple of days ago | 12:18 |
@sonney2k | and don't work on comparing this stuff ... it doesn't help to kill more time with this. you need some principled investigation now | 12:19 |
n4nd0 | what do you mean with principled investigation? | 12:19 |
n4nd0 | what do you want me to do? | 12:19 |
@sonney2k | n4nd0, have a session with uricamic/nico | 12:19 |
n4nd0 | sonney2k: what should we disscus on it? | 12:20 |
@sonney2k | - compute objective in your / uricamic's MC SO | 12:20 |
@sonney2k | - ensure that you are both solving the same problem | 12:20 |
n4nd0 | I have just asked uricamic about that | 12:21 |
@sonney2k | - compare if objectives deviate by no more than epsilon | 12:21 |
n4nd0 | it looks that the formulation is not the same | 12:21 |
n4nd0 | his problem is about risk minimization | 12:21 |
n4nd0 | I don't think it is just the algorithm that changes but also the objective | 12:21 |
@sonney2k | n4nd0, well you both solve the true MC problem w/o bias | 12:22 |
@sonney2k | the CS formulation right? | 12:22 |
@sonney2k | so when you have done that you get as result a couple of w's | 12:22 |
n4nd0 | he has just told me that his formulation is | 12:22 |
n4nd0 | min_{w} 1/2 ||w||^2 + R | 12:22 |
@sonney2k | using these w's you can compute an unconstrained objective | 12:22 |
@sonney2k | and compare it | 12:22 |
@sonney2k | n4nd0, looks like the same as yours or? | 12:23 |
n4nd0 | mine uses sum_i xi_i instead of R | 12:23 |
n4nd0 | I am not completely sure what his R is though | 12:23 |
n4nd0 | probably it is the risk | 12:23 |
n4nd0 | and one can choose the regularizer for it | 12:24 |
n4nd0 | but the constraints are probably different | 12:24 |
@sonney2k | n4nd0, it is the risk term | 12:25 |
@sonney2k | n4nd0, sum of slacks weighted with C | 12:25 |
@sonney2k | n4nd0, one can solve the problem in an unconstrained and a constrained way | 12:26 |
@sonney2k | uricamic, does it the unconstrained you the constrained way | 12:26 |
@sonney2k | but you can compute the unconstrained objective and compare with his result | 12:26 |
@sonney2k | all you need for that is the w's | 12:26 |
n4nd0 | sonney2k: I think uricamic went for a while, maybe he's not back yet | 12:26 |
@sonney2k | n4nd0, probably lunch time now | 12:27 |
-!- nicococo [~nico@lacedcoffee.ml.tu-berlin.de] has joined #shogun | 12:28 | |
n4nd0 | hi nicococo, how are you? | 12:29 |
nicococo | hi n4nd0: so mc svm is fine and we are now approaching the hmsvm ? | 12:37 |
n4nd0 | nicococo: about the mc svm I just wanted to check the results with other implementation of the same thing to ensure it is ok, I think svm-struct can be useful for this | 12:38 |
n4nd0 | nicococo: but yes, we can start approaching hmsvm | 12:38 |
n4nd0 | nicococo: I studied the paper about semi hm-svm as you suggested to find out a bit more about these plifs | 12:39 |
n4nd0 | nicococo: but I am afraid I have not clear why the pop up in hm-svm yet | 12:39 |
nicococo | ok, its good to get a general idea but of course the paper is much more advanced than our next step | 12:40 |
n4nd0 | yeah | 12:40 |
nicococo | as i told you, i am preparing a document with detailed information (and nice pictures) | 12:40 |
n4nd0 | yeah, that sounds awesome | 12:41 |
n4nd0 | about the Viterbi implementation | 12:41 |
n4nd0 | I don't know if you checked the irc logs | 12:41 |
n4nd0 | but sonney2k has told me some minutes ago that it would be better to do a new implementation for it | 12:41 |
nicococo | no.. | 12:41 |
n4nd0 | http://www.shogun-toolbox.org/irclogs/%23shogun.2012-06-22.log.html | 12:42 |
n4nd0 | 12:17 | 12:42 |
nicococo | no problem. i can give you a 30 lines of c code doing what we need | 12:42 |
n4nd0 | ok | 12:42 |
nicococo | unfortunately the document is not finished yet :( | 12:43 |
n4nd0 | no problem | 12:44 |
n4nd0 | for me it's ok to follow with my plan of comparing the results with svm-struct for multiclass | 12:45 |
n4nd0 | they provide a solver for the 1-slack formulation | 12:45 |
n4nd0 | in the primal | 12:45 |
n4nd0 | we have the n-slack | 12:46 |
n4nd0 | for I am reading in the document that they are somewhat equivalent | 12:46 |
n4nd0 | :%s/for/but ^ | 12:46 |
n4nd0 | so I am just going to get the svm-struct running and check if w is close to my w | 12:48 |
nicococo | thats okay. if you have some time i would suggest, that you also start writing down how you would approach the hmsvm.. | 12:48 |
nicococo | maybe its easier to see where problems are.. what do you think? | 12:49 |
-!- uricamic [~uricamic@89.13.broadband14.iol.cz] has quit [Quit: Leaving.] | 12:50 | |
nicococo | oh.. and one more thing: forget the plifs :) | 12:50 |
n4nd0 | haha ok :) | 12:52 |
n4nd0 | maybe now is too fast to say how I would approach it in detail | 12:54 |
n4nd0 | but my general idea is to first have the three pieces that we would give to the PrimalMosekSOSVM | 12:54 |
n4nd0 | psi | 12:55 |
n4nd0 | argmax = viterbi | 12:55 |
n4nd0 | delta loss = ? | 12:55 |
nicococo | for delta we can use hamming loss | 12:55 |
n4nd0 | but how is it defined for sequences? | 12:55 |
n4nd0 | forget about what I have just said... | 12:56 |
nicococo | :) | 12:56 |
n4nd0 | but there must be more apart from this | 12:56 |
nicococo | i'll cover that in the document | 12:57 |
n4nd0 | ok | 12:57 |
n4nd0 | so do you know something about what I said of 1-slack and n-slack for the primal so-svm? | 12:58 |
n4nd0 | should I just expect that both give the same w (accepting some error for the difference in the optimizer, etc) | 12:59 |
n4nd0 | or? | 12:59 |
nicococo | this is another optimization approach. there are several possible but this aims in the direction of bundle methods | 12:59 |
nicococo | yes, same data should produce the (almost) same output | 12:59 |
nicococo | depends on the stopping criterion and so.. | 13:00 |
nicococo | you can compare the objective values also | 13:00 |
-!- uricamic [~uricamic@89.13.broadband14.iol.cz] has joined #shogun | 13:01 | |
n4nd0 | what do you mean with the objectives values? | 13:01 |
nicococo | the objective function value at w^* | 13:02 |
n4nd0 | got it | 13:02 |
nicococo | okay.. i go to lunch now. if there are probs just write an email.. i will be here again later .. | 13:03 |
n4nd0 | ok, not me probably | 13:03 |
n4nd0 | today is midsommar :) | 13:03 |
nicococo | enjoy ;) | 13:04 |
-!- nicococo [~nico@lacedcoffee.ml.tu-berlin.de] has left #shogun [] | 13:04 | |
-!- uricamic [~uricamic@89.13.broadband14.iol.cz] has quit [Ping timeout: 248 seconds] | 13:15 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 13:25 | |
pluskid | sonney2k: how to disable pthread when compiling? I'm having some strange bugs here: it hangs when computing dense_dot_range of DotFeatures | 13:26 |
-!- uricamic [~uricamic@89.13.broadband14.iol.cz] has joined #shogun | 13:28 | |
@sonney2k | pluskid just Set number of threads to 1 | 14:05 |
pluskid | sonney2k: where to set? | 14:06 |
@sonney2k | obj.parallel.set_num_threads(1) | 14:48 |
@sonney2k | pluskid, ^ | 14:48 |
pluskid | sonney2k: ok, thanks! | 14:48 |
@sonney2k | pluskid, where does it hang btw? | 14:48 |
@sonney2k | I am not aware of any issues... | 14:49 |
pluskid | sonney2k: finally I traced it is not happening in dot product, but somewhere SG_UNREF | 14:49 |
pluskid | maybe some memory issue | 14:49 |
pluskid | wild pointers maybe | 14:49 |
pluskid | I'm still trying to dig deeper | 14:49 |
@sonney2k | pluskid, doing what? | 14:50 |
pluskid | sonney2k: when I'm testing the ShareBoost implementation, when setting the features of a linear machine | 14:50 |
@sonney2k | pluskid, just setting features you mean? | 15:43 |
pluskid | sonney2k: yeah | 15:53 |
@sonney2k | sounds weird... | 15:54 |
pluskid | yeah, I think I should double check my code to ensure all memory operations are correct | 15:58 |
@sonney2k | valgrind is your friend :) | 15:59 |
pluskid | it hangs on the SG_UNREF statement of setting features | 15:59 |
pluskid | ah, I almost forgot about valgrind, thx! | 15:59 |
@sonney2k | pluskid, sounds like uninited memory | 16:02 |
@sonney2k | probably features wasn't inited with NULL | 16:02 |
pluskid | seems not, it hangs on the 2nd time when setting features | 16:02 |
pluskid | 1st time when features are previously NULL is OK | 16:03 |
pluskid | SG_UNREF on an non-NULL features hangs | 16:03 |
pluskid | not sure whether the features corrupted | 16:03 |
pluskid | anyways, it is strange that the program hangs, instead of segfault or something | 16:03 |
@sonney2k | pluskid, hmmhh well with memory corruptions anything can happen | 16:06 |
-!- alexlovesdata [c25faea9@gateway/web/freenode/ip.194.95.174.169] has quit [Quit: Page closed] | 16:09 | |
-!- vojtech_ [9320543b@gateway/web/freenode/ip.147.32.84.59] has quit [Quit: Page closed] | 16:13 | |
-!- pluskid [~pluskid@202.130.113.141] has quit [Quit: Leaving] | 16:54 | |
-!- emrecelikten [~emre@176.40.230.182] has quit [Ping timeout: 264 seconds] | 17:11 | |
-!- emrecelikten [~emre@82.222.144.51] has joined #shogun | 17:26 | |
-!- ckwidmer [45c9b18c@gateway/web/freenode/ip.69.201.177.140] has joined #shogun | 17:32 | |
-!- ckwidmer [45c9b18c@gateway/web/freenode/ip.69.201.177.140] has quit [Quit: Page closed] | 17:48 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!