Open in new window / Try shogun cloud
--- Log opened Fri Jun 22 00:00:20 2012
--- Day changed Fri Jun 22 2012
@sonney2kyeah the guy at the very left that appears a couple of times is a ML person00:00
blackburnheh00:00
@sonney2kthere are others but hard to descibe :D00:01
-!- 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 #shogun06:11
-!- uricamic [~uricamic@89.13.broadband14.iol.cz] has joined #shogun09:11
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun11:04
-!- gsomix [~gsomix@188.168.2.228] has joined #shogun11:06
-!- vojtech_ [9320543b@gateway/web/freenode/ip.147.32.84.59] has joined #shogun11:34
n4nd0sonney2k, alexlovesdata: about comparing the results of my SO code with SVM-struct for multiclass11:43
n4nd0I am not sure if it is possible since they don't have exactly the same formulation of the problem11:43
n4nd0sonney2k: can it be reliable to compare results of the primal formulation with results of the dual?11:44
@sonney2kn4nd0, well now that uricamic merged his MC example based on his BMRM implementation it is probably easier to compare against that11:45
@sonney2kvojtech_, am I right?11:45
n4nd0sonney2k: I found something that might work though, they provide code to solve a 1-slack formulation of the problem in the primal11:46
n4nd0sonney2k: what we have is a n-slack formulation11:46
@sonney2kn4nd0, btw is nico on vacations?11:47
n4nd0sonney2k: BUT in the paper they say that a w solution of one problem is solution of the other11:47
@sonney2kn4nd0, IIRC this is all equivalent11:47
vojtech_sonney2k: yes, Michal reimplemented BMRM which we have been using many times and it seems to work correctly11:47
n4nd0sonney2k: he said that his vacations will start from the middle of the next month11:47
n4nd0sonney2k: btw, will there be a problem with that and this mid-term report?11:49
@sonney2kn4nd0, so you should work with uricamic to compare your mosek based code against his impl.11:49
@sonney2kyou should both get very similar objective11:50
n4nd0sonney2k: that sounds better, I am going to talk to him11:50
@sonney2kand then all is good11:50
@sonney2kvojtech_, thanks - as I understand uricamic is now working on the new fancy thing right?11:50
@sonney2kvojtech_, 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
@sonney2kvojtech_, 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
@sonney2kvojtech_, what I meant is some fancy example that can be put into shogun11:55
@sonney2kyou know better what is easy / complex enough :D11:55
vojtech_ok, I'll think about it11:56
@sonney2kn4nd0, btw if you are stuck for too long - investigate intensively with nico11:59
n4nd0sonney2k: he told me he's preparing a document with HMSVM info that should be ready by the end of this week11:59
n4nd0sonney2k: until that, I want to test the so-multiclass and get familiarized with the Viterbi impl. in Shogun12:00
@sonney2kn4nd0, for how long are you trying to figure out whether the MC - SO SVM is correct?12:11
n4nd0sonney2k: I have been with it basically this week and the last one ...12:12
n4nd0I think it has been a loss of time to tell the truth12:12
@sonney2kn4nd0, yeah so you should have a meeting with nico, uricamic *ASAP*12:13
-!- pluskid [~pluskid@202.130.113.141] has joined #shogun12:13
@sonney2kn4nd0, email them asking them if they have time this afternoon and CC me...12:13
n4nd0sonney2k: I have no time this afternoon12:14
@sonney2kn4nd0, well when do you have time now?12:14
n4nd0sonney2k: today it is midsommar in Sweden, national holiday12:14
@sonney2kn4nd0, what I am saying is ASAP - otherwise you will become demotivated12:14
@sonney2ktrying for 2 weeks to find out what is wrong is just to long12:15
@sonney2ktoo12:15
@sonney2k1 week is already very tiring12:15
n4nd0sonney2k: yeah12:15
n4nd0but Nico told me to wait until he has ready the file with the new work to do12:15
@sonney2kn4nd0, so you definitely need help here!12:15
@sonney2kn4nd0, no12:15
@sonney2kthe document can wait12:16
@sonney2kn4nd0, or do you have something else to do?12:16
n4nd0sonney2k: apart from shogun you mean?12:16
@sonney2kn4nd0, no for GSoC I mean12:16
n4nd0no12:17
@sonney2kand the viterbi in shogun is very hard to digest12:17
@sonney2keven worse I would say12:17
n4nd0the plan was to check viterbi12:17
n4nd0mmm but I'll need to use it in any case12:17
n4nd0maybe I should get familiarized with it at least12:17
@sonney2kn4nd0, I tend to say that writing your own is easier!12:17
@sonney2kthe one in shogun is very specialized12:18
@sonney2kand not just for HM-SVM12:18
@sonney2kbut for semi-HM SVM12:18
n4nd0ok12:18
@sonney2k-> much more difficult12:18
@sonney2kn4nd0, ask nico and uricamic now!12:18
n4nd0yeah, I studied your paper about it a couple of days ago12:18
@sonney2kand don't work on comparing this stuff ... it doesn't help to kill more time with this. you need some principled investigation now12:19
n4nd0what do you mean with principled investigation?12:19
n4nd0what do you want me to do?12:19
@sonney2kn4nd0, have a session with uricamic/nico12:19
n4nd0sonney2k: what should we disscus on it?12:20
@sonney2k- compute objective in your / uricamic's MC SO12:20
@sonney2k- ensure that you are both solving the same problem12:20
n4nd0I have just asked uricamic about that12:21
@sonney2k- compare if objectives deviate by no more than epsilon12:21
n4nd0it looks that the formulation is not the same12:21
n4nd0his problem is about risk minimization12:21
n4nd0I don't think it is just the algorithm that changes but also the objective12:21
@sonney2kn4nd0, well you both solve the true MC problem w/o bias12:22
@sonney2kthe CS formulation right?12:22
@sonney2kso when you have done that you get as result a couple of w's12:22
n4nd0he has just told me that his formulation is12:22
n4nd0min_{w} 1/2 ||w||^2 + R12:22
@sonney2kusing these w's you can compute an unconstrained objective12:22
@sonney2kand compare it12:22
@sonney2kn4nd0, looks like the same as yours or?12:23
n4nd0mine uses sum_i xi_i instead of R12:23
n4nd0I am not completely sure what his R is though12:23
n4nd0probably it is the risk12:23
n4nd0and one can choose the regularizer for it12:24
n4nd0but the constraints are probably different12:24
@sonney2kn4nd0, it is the risk term12:25
@sonney2kn4nd0, sum of slacks weighted with C12:25
@sonney2kn4nd0, one can solve the problem in an unconstrained and a constrained way12:26
@sonney2kuricamic, does it the unconstrained you the constrained way12:26
@sonney2kbut you can compute the unconstrained objective and compare with his result12:26
@sonney2kall you need for that is the w's12:26
n4nd0sonney2k: I think uricamic went for a while, maybe he's not back yet12:26
@sonney2kn4nd0, probably lunch time now12:27
-!- nicococo [~nico@lacedcoffee.ml.tu-berlin.de] has joined #shogun12:28
n4nd0hi nicococo, how are you?12:29
nicococohi n4nd0: so mc svm is fine and we are now approaching the hmsvm ?12:37
n4nd0nicococo: 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 this12:38
n4nd0nicococo: but yes, we can start approaching hmsvm12:38
n4nd0nicococo: I studied the paper about semi hm-svm as you suggested to find out a bit more about these plifs12:39
n4nd0nicococo: but I am afraid I have not clear why the pop up in hm-svm yet12:39
nicococook, its good to get a general idea but of course the paper is much more advanced than our next step12:40
n4nd0yeah12:40
nicococoas i told  you, i am preparing a document with detailed information (and nice pictures)12:40
n4nd0yeah, that sounds awesome12:41
n4nd0about the Viterbi implementation12:41
n4nd0I don't know if you checked the irc logs12:41
n4nd0but sonney2k has told me some minutes ago that it would be better to do a new implementation for it12:41
nicococono..12:41
n4nd0http://www.shogun-toolbox.org/irclogs/%23shogun.2012-06-22.log.html12:42
n4nd012:1712:42
nicococono problem. i can give you a 30 lines of c code doing what we need12:42
n4nd0ok12:42
nicococounfortunately the document is not finished yet :(12:43
n4nd0no problem12:44
n4nd0for me it's ok to follow with my plan of comparing the results with svm-struct for multiclass12:45
n4nd0they provide a solver for the 1-slack formulation12:45
n4nd0in the primal12:45
n4nd0we have the n-slack12:46
n4nd0for I am reading in the document that they are somewhat equivalent12:46
n4nd0:%s/for/but ^12:46
n4nd0so I am just going to get the svm-struct running and check if w is close to my w12:48
nicococothats okay. if you have some time i would suggest, that you also start writing down how you would approach the hmsvm..12:48
nicococomaybe 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
nicococooh.. and one more thing: forget the plifs :)12:50
n4nd0haha ok :)12:52
n4nd0maybe now is too fast to say how I would approach it in detail12:54
n4nd0but my general idea is to first have the three pieces that we would give to the PrimalMosekSOSVM12:54
n4nd0psi12:55
n4nd0argmax = viterbi12:55
n4nd0delta loss = ?12:55
nicococofor delta we can use hamming loss12:55
n4nd0but how is it defined for sequences?12:55
n4nd0forget about what I have just said...12:56
nicococo:)12:56
n4nd0but there must be more apart from this12:56
nicococoi'll cover that in the document12:57
n4nd0ok12:57
n4nd0so do you know something about what I said of 1-slack and n-slack for the primal so-svm?12:58
n4nd0should I just expect that both give the same w (accepting some error for the difference in the optimizer, etc)12:59
n4nd0or?12:59
nicococothis is another optimization approach. there are several possible but this aims in the direction of bundle methods12:59
nicococoyes, same data should produce the (almost) same output12:59
nicococodepends on the stopping criterion and so..13:00
nicococoyou can compare the objective values also13:00
-!- uricamic [~uricamic@89.13.broadband14.iol.cz] has joined #shogun13:01
n4nd0what do you mean with the objectives values?13:01
nicococothe objective function value at w^*13:02
n4nd0got it13:02
nicococookay.. i go to lunch now. if there are probs just write an email.. i will be here again later ..13:03
n4nd0ok, not me probably13:03
n4nd0today is midsommar :)13:03
nicococoenjoy ;)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
pluskidsonney2k: how to disable pthread when compiling? I'm having some strange bugs here: it hangs when computing dense_dot_range of DotFeatures13:26
-!- uricamic [~uricamic@89.13.broadband14.iol.cz] has joined #shogun13:28
@sonney2kpluskid just Set number of threads to 114:05
pluskidsonney2k: where to set?14:06
@sonney2kobj.parallel.set_num_threads(1)14:48
@sonney2kpluskid, ^14:48
pluskidsonney2k: ok, thanks!14:48
@sonney2kpluskid, where does it hang btw?14:48
@sonney2kI am not aware of any issues...14:49
pluskidsonney2k: finally I traced it is not happening in dot product, but somewhere  SG_UNREF14:49
pluskidmaybe some memory issue14:49
pluskidwild pointers maybe14:49
pluskidI'm still trying to dig deeper14:49
@sonney2kpluskid, doing what?14:50
pluskidsonney2k: when I'm testing the ShareBoost implementation, when setting the features of a linear machine14:50
@sonney2kpluskid, just setting features you mean?15:43
pluskidsonney2k: yeah15:53
@sonney2ksounds weird...15:54
pluskidyeah, I think I should double check my code to ensure all memory operations are correct15:58
@sonney2kvalgrind is your friend :)15:59
pluskidit hangs on the SG_UNREF statement of setting features15:59
pluskidah, I almost forgot about valgrind, thx!15:59
@sonney2kpluskid, sounds like uninited memory16:02
@sonney2kprobably features wasn't inited with NULL16:02
pluskidseems not, it hangs on the 2nd time when setting features16:02
pluskid1st time when features are previously NULL is OK16:03
pluskidSG_UNREF on an non-NULL features hangs16:03
pluskidnot sure whether the features corrupted16:03
pluskidanyways, it is strange that the program hangs, instead of segfault or something16:03
@sonney2kpluskid, hmmhh well with memory corruptions anything can happen16:06
-!- 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 #shogun17:26
-!- ckwidmer [45c9b18c@gateway/web/freenode/ip.69.201.177.140] has joined #shogun17:32
-!- ckwidmer [45c9b18c@gateway/web/freenode/ip.69.201.177.140] has quit [Quit: Page closed]17:48