n4nd0 blackburn n4nd0 --- Log opened Wed Aug 22 00:00:05 2012 --- Day changed Wed Aug 22 2012 (I don't know how to write most today...) 00:00 it appear in risk 00:00 well ok ... it looks like we are not going to agree today 00:00 my guess is that you need to explicitily look for what is the value of that max_y 00:01 risk uses violating object 00:01 and not say that it is going to be given by argmax_y 00:01 decision function does not 00:01 in case of correct delta it should comply with max of the second term 00:02 I don't think so 00:02 I mean, I do not agree 00:02 what can I say it works for HMM :) 00:02 a good model will give you for a training pair (xi,yi) 00:03 will commit things pretty soon 00:03 an argmax_y for xi that is close to yi 00:03 then the delta will be close to zero 00:03 why we want to make delta zero 00:03 if we maximize it? 00:03 blackburn: commit so I can take it a look 00:03 will do once I check examples 00:04 rolf this yml format makes me crayz .. "The node does not represent a user object" 00:04 I hope I will manage to wake up tomorrow :D 00:05 today I was late for an hour 00:05 hehe 00:05 hehe 00:06 your commits today make me fell that one hour late didnt prevent you working on shogun during the day ^^ 00:07 I am trying to find a balance between not being fired and shoguning 00:08 n4nd0: how many iterations mosek does on that hm svm example? 00:09 blackburn: I am not sure right now, I think between 6-10, let me check 00:10 bmrm does 14 00:10 takes 0.63s here 00:10 it takes quite longer here 00:10 9 iterations 00:12 shogun: Sergey Lisitsyn master * r45840d7 / (12 files in 5 dirs): Refactored SO machine to not use its own features reference, fixed generic risk and bmrm result type - http://git.io/fj6dnQ 00:12 26 seconds 00:12 here you go 00:12 wow that's a large lambda :) 00:12 how did you guess that value? 00:12 shogun: Sergey Lisitsyn master * rf253152 / (2 files in 2 dirs): Improved multiclass tree guided logistic regression - http://git.io/1RuW7Q 00:13 no idea just entered big value 00:13 lol 00:13 it is not the best value btw 00:14 objective tends to be pretty big this way 00:14 okay 5 millions is too big 00:14 500 thousands tends to be overregularize too 00:15 hmm 50000 works just like 5000 00:15 0 is da best 00:15 will not converge 00:15 and furthermore will break you computer 00:16 I changed one already 00:16 n4nd0: with 0.01 accuracy is 0.9989 00:16 blackburn: nice 00:16 HM-SVM? 00:16 with 1e-5 it is 0.9991 00:16 yes 00:16 can it be 100%? 00:17 maybe 00:17 lets try 1e-37 00:17 oh no 00:17 noooo 00:17 0.9998 and crazy numbers 00:18 Fp=720255088080000357764535868940049436668877340672.000000, 00:18 what? 00:18 that is primal objective LOL 00:18 O_O 00:18 that looks pretty weird 00:18 maybe my accuracy thing is doing something wrong 00:18 no 00:18 that is ok for 1e-37 lambda 00:19 :D 00:19 n4nd0: w contain e+41 elements this way 00:20 n4nd0: okay let you compare results now :) 00:20 n4nd0: e+41 elements? 00:20 talking to myself again... 00:20 n4nd0: not number, but order 00:20 does the objective in bmrm minimize too? 00:21 n4nd0: do you understand what they sing about 00:21 http://www.youtube.com/watch?v=Q7xZ3cBPxIE&feature=player_detailpage 00:21 ? 00:21 :D 00:21 n4nd0: hmmmm really.. I don't know 00:21 looks like it maximiaze 00:21 blackburn: can you explain again why it is the argmax? :P 00:23 I just realized it is not feasible to check all the combinations y \in Y in HM-SVM 00:23 it just make sense to check all in the multiclass example 00:23 n4nd0: why loss is argmax? 00:24 because we find a most violating label for the vector 00:24 why max_y [ loss(yi,y) + < w, Psi(xi,y)-Psi(xi,yi) > ] 00:24 why the maximum is found for the argmax 00:24 ?? 00:25 why the maximum of the equation I have just written 00:25 is achieved with the argmax_y function 00:25 but the definition of the argmax is not to find the most violated label 00:25 it is to maximize 00:25 max_y [ loss(yi,y) + < w, Psi(xi,y)-Psi(xi,yi) > ] = max_y [ loss(yi,y) + < w, Psi(xi,y)> ] - 00:27 sure 00:28 now why 00:28 argmax_y [ loss(yi,y) + < w, Psi(xi,y) > ] = argmax_y < w, Psi(xi,y) > ?? 00:29 yes it is 00:29 for proper loss it is for sure 00:29 why? 00:29 they are both losses 00:30 first is 'label' loss 00:30 second is feature space 'loss' 00:30 that's not really a reason... 00:30 let me get a tea 00:30 :D 00:30 haha 00:31 we can talk about it tomorrow otherwise 00:31 maybe I am just stupid right now 00:31 and so stubborn as to see it 00:31 but this idea seems so simple 00:31 argmax_y should give a y close to yi for the input xi 00:32 I am kind of practical guy so I prefer to check if it works first :D 00:32 and a y close to yi implies Delta(yi,y) close to zero 00:32 blackburn: I cannot compile here 00:33 https://github.com/shogun-toolbox/shogun/blob/master/src/shogun/structure/PrimalMosekSOSVM.cpp 00:34 line 30 00:35 that constructor is not in 00:35 https://github.com/shogun-toolbox/shogun/blob/master/src/shogun/machine/LinearStructuredOutputMachine.h 00:35 uh ok 00:35 did you forget to add any of the files or? 00:36 I removed features - did you forgot? :) 00:36 yeah I know you did that :) 00:36 but since the build doesn't compile 00:37 -!- naywhaya1e is now known as naywhayare 00:37 I have to do a blind fix 00:38 build #431 of deb3 - modular_interfaces is complete: Failure [failed compile csharp_modular]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/431  blamelist: Sergey Lisitsyn 00:38 ? 00:38 blackburn: why a blind fix? 00:39 have no mercy^W mosek 00:39 aaah ok 00:39 true fact 00:39 whicho ne? 00:39 I can apply that if you want 00:39 it looks like that was the only one 00:40 okay let you check it 00:40 n4nd0: please try now 00:41 shogun: Sergey Lisitsyn master * rff37c90 / (5 files in 2 dirs): Updated PrimalMosekSOSVM - http://git.io/kTbSbg 00:41 n4nd0: inequality just after A.8 in teo 00:44 this one should make you believe 00:45 mmm 00:45 I didn't though about that 00:45 1 + 2 is majorizing thing for 1 00:45 yes 00:46 so it is a major of risk 00:46 but I don't understand why the loss is being majorized too! 00:46 haha 00:46 I am becoming crazy 00:46 there is a description out there 00:46 first of all dot product is positive okay? ;) 00:47 ok :) 00:47 second ineq comes from max 00:47 ? 00:48 orr 00:49 anyway 00:49 are you convinced now? 00:49 I am reading the explanation around here 00:49 Note that (A.2) majorizes ?(y, y * ), where y * := argmaxy w, ?(x, y ) [Tsochantaridis 00:49 et al., 2005 00:49 haha, magic copying from PDFs... 00:50 I have no idea if it has something formal 00:51 I have a  gut feeling about that :D 00:51 I know that would not work in science haha 00:51 shogun: Sergey Lisitsyn master * r416d360 / (2 files in 2 dirs): Doc fixes - http://git.io/VRU7oQ 00:51 blackburn: why not? 00:52 do you mean you? 00:52 what do you mean now? 00:52 :) 00:52 < blackburn> I know that would not work in science haha 00:52 I mean no one would believe me if I said I have a gut feeling :D 00:53 about quantum mechanics or economics or anything 00:53 aham 00:53 okay no moar mosek 00:53 I'm satisfied now 00:54 :) 00:56 build #432 of deb3 - modular_interfaces is complete: Failure [failed compile csharp_modular]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/432  blamelist: Sergey Lisitsyn 01:00 blackburn: I think the same still! 01:01 haha 01:01 I believe strongly that another argmax is required 01:02 it is infeasible 01:02 one that includes the term Delta(yi,y) 01:02 it depends on the application 01:02 for the HM-SVM it is probably feasible 01:02 infeasible for HM, right? 01:03 okay, infeasible for CRF? 01:03 in the same way that argmax_y < w, Psi(x,y) > 01:03 blackburn: my CRF knowledge is a bit limited for this 01:03 but my guess is that it is feasible for CRF too 01:03 since they have been used in real applications 01:04 you should taklk to some SO expert 01:04 blackburn: do you give me some confidence at least? :D 01:04 I am pretty sure it is correct now 01:04 but you are just basing it in the result of the accuracy 01:05 take into account that the dataset is not difficult 01:05 the underlying distribution is not really difficult 01:05 no, I am basing it on my understanding of all these things 01:05 and we are providing lot of training examples 01:05 but look to equation A.7 in Teo 01:06 and? 01:06 a hint could be that they do not use the same to denot 01:06 denote 01:07 y_bar 01:07 and y_* 01:07 y_* = argmax_y < w, phi(x,y) > 01:07 and y_bar is the thing of A.7 01:07 oh I have to make SO locked training work 01:07 ? 01:08 it is not supported now 01:08 anyway 01:09 time to sleep now 01:09 good idea 01:10 good night & good job :) 01:10 good night 01:10 -!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] 01:10 -!- zxtx [~zv@cpe-75-83-151-252.socal.res.rr.com] has quit [Ping timeout: 252 seconds] 01:16 -!- blackburn1 [~blackburn@188.168.14.172] has joined #shogun 01:20 -!- blackburn [~blackburn@37.61.181.133] has quit [Read error: Operation timed out] 01:21 -!- zxtx [~zv@cpe-75-83-151-252.socal.res.rr.com] has joined #shogun 01:27 shogun: Sergey Lisitsyn master * rd6b3a7d / (9 files in 3 dirs): Refactored apply of latent machines - http://git.io/9Omc-g 01:29 -!- blackburn1 [~blackburn@188.168.14.172] has quit [Quit: Leaving.] 01:38 build #433 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/433 01:42 -!- yooo [575b08cb@gateway/web/freenode/ip.87.91.8.203] has quit [Ping timeout: 245 seconds] 02:41 build #62 of nightly_none is complete: Failure [failed compile]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_none/builds/62 03:02 -!- av3ngr [av3ngr@nat/redhat/x-swaimdyxogxxwdlk] has joined #shogun 05:57 shogun: Sergey Lisitsyn master * rcf43532 / (2 files): Fixed compilation of data generator in case of no lapack available - http://git.io/QAFu6w 07:43 shogun: Sergey Lisitsyn master * r40564ac / src/shogun/latent/LatentSOSVM.cpp : Added missed lambda usage in LatentSOSVM - http://git.io/jfaIwQ 08:52 shogun: Sergey Lisitsyn master * r157fba3 / src/interfaces/modular/modshogun_ignores.i : An attempt to ignore display methods of SG datatypes - http://git.io/t04KgA 08:58 -!- uricamic [~uricamic@2001:718:2:1634:155e:7544:fba4:9878] has joined #shogun 09:03 -!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun 09:32 hey uricamic 09:44 how are you doing? 09:44 uricamic: hey - you should convince n4nd0 today :D 09:49 haha 09:49 _____________: uricamic and I will convince you :D 09:50 _____________: let me ask you 09:53 why do you do CDotFeatures* in 47 of PrimalMosekSOSVM.cpp 09:53 n4nd0: why not CFeatures? 09:54 _____________: yeah 09:54 because I didn't know what is restriction 09:54 please change if it is not correct 09:54 ok 09:54 just wanted to ensure that there was no special reason 09:54 good 09:54 but remember this restriction was the one that started all the refactoring 09:55 because in BMRM it was CDotFeatures and we needed CFeatures there 09:55 yes 09:56 -!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has joined #shogun 10:13 -!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has quit [Client Quit] 10:13 shogun: iglesias master * rb985ee1 / src/shogun/structure/PrimalMosekSOSVM.cpp : * fix compilation error with MOSEK support - http://git.io/TtNrZw 10:23 shogun: Sergey Lisitsyn master * r8122bd8 / src/shogun/structure/PrimalMosekSOSVM.cpp : Merge pull request #751 from iglesias/master - http://git.io/0aIxew 10:23 hi, sorry, I have been away 10:28 uricamic: we were arguing HARD yesterday 10:28 uricamic: help us to resolve that :) 10:28 uricamic: okay - the problem is 10:29 ok :) 10:29 is argmax of 10:29 equal to argmax of 10:29 \Delta(y_i, y) + 10:29 see what I mean? 10:29 uricamic: put some light into here please :D 10:30 I am not sure what u mean, but wait a sec, I will figure it out :) 10:30 uricamic: I've added generic risk function 10:30 to structured model 10:30 uricamic: when one needs to compute the risk function 10:30 have you seen? 10:30 it works like 10:30 not yet, but I will check it now 10:30 for each feature vector: find argmax, update risk and update subgrad 10:31 but n4nd0 thinks it is wrong to find argmax without loss 10:31 yep, it is wrong 10:31 :) 10:32 u need to add loss there, because u are trying to find the most violated constraint there 10:32 and adapt to it 10:32 it is similar to perceptron algorithm 10:32 and the argmax doesn't give you the most violated 10:32 it can if u add loss term there 10:33 to all examples 10:33 okay it is a luck it works now then :) 10:33 e.g. I have been using it in one of my applications of so-svm, I have designed argmax function in matlab with the possibility to pass also the losses 10:33 _____________: and on which instance u have checked that? 10:34 HM-SVM and multiclass 10:34 n4nd0: is it feasible to implement argmax with loss? 10:34 for HM-SVM? 10:34 it should be possible 10:35 n4nd0: can you fix my code then? 10:36 what is needed is 10:36 _____________: it should be done modifying the Viterbi I implemented there 10:36 to add argmax_loss or so 10:36 uricamic: I am trying out the DualLibQPBMSOSVM and it never ends :( 10:36 I am doing something wrong for sure 10:37 change risk to call that, remove loss addition in risk 10:37 n4nd0: I haven't checked your code for HM-SVM yet, but wouldn't it be possible to just ass loss to the features in argmax? 10:37 and then implement argmax_loss for multiclass and HM 10:37 n4nd0: which algorithm and what lambda have u tryied? 10:37 *tried 10:37 uricamic: I have not set any algorithm explicitely to tell the truth 10:39 _____________: btw: I wanted to ask about the changes in my code for modular interfaces, because of java_modular failure 10:39 so it should use BMRM 10:39 uricamic: I thought that would use BMRM by default 10:39 I fixed that 10:39 that one has proven convergency 10:39 uricamic: lamda equal to 0.01 10:39 it doesn't finish though 10:39 I see, that could be the problem 10:39 uricamic: do I understand right C is 1/lambda? 10:40 Like it goes through 1000 iterations and did not stop? 10:40 uricamic: ah one ask for your code too 10:40 could you please add !CSignal::cancel_computations() to while condition? 10:40 uricamic: 294 iterations and it just froze 10:40 I guess not just 1/lambda, it depends on the formulation of the task with C, do you have it somewhere? I guess it could be n/lambda 10:41 to make it possible to Ctrl+C that 10:41 froze? 10:41 hmm, that is strange 10:41 it might mean that the qp task is really hard to solve 10:41 and it could take really a long wjile 10:41 *while 10:41 I am using the so_multiclass example, in libshogun 10:42 n4nd0: it does not converge for so_multiclass ? 10:44 it was converging on my machine ehm 10:44 very fast 10:44 I am doing something wrong for sure then 10:45 but what? 10:45 http://pastebin.com/F6q9rrVS 10:46 just find bundle to look for the parts that use bmrm 10:46 try to increase lambda 10:47 a lot 10:47 because, I probably haven't mentioned it anywhere yet 10:47 but because of P3BMRM 10:47 we are using unnormalized risk 10:47 i.e. without division by number of examples 10:47 which changes the lambda a lot 10:48 to what number? 10:48 try something really big for start and then u can try lower 10:48 trying with 1000 10:48 like 1e4 10:48 yep, or that one 10:48 the problem is that without this unnormalized risk, P3BMRM would not be identical to BMRM 10:49 ok it finished now, thank you :) 10:49 it would require division of number of cp_models used depending also on the number of examples in the cp_model 10:49 n4nd0: you are welcome 10:49 I get this accuracy 10:50 SO-SVM: 82.80% 10:50 BMRM:   69.90% 10:50 is that normal? 10:50 but when using unnormalized risk, they are equal 10:50 could be, try lower lambda now :) 10:50 model selection for this would be nice :D 10:51 I have finished benchamrk for MNIST data yesterday, but I will probably have to change my code a bit to be able to send it in PR 10:52 I mean  I need to return history of Fp, Fd and wdist from the solver, but in current form it crashes java_modular 10:53 so I guess, it will be sufficient to add getters for these to DualLibQPBMSOSVM class and everything should be fine then 10:54 uricamic: it works now 10:56 I believe it does 10:57 _____________: how come?  I guess it is impossible :) 10:57 uricamic: you didn't declare constructor and it was not inherited from sgobject 10:57 because now in modular there is no support to return bmrm_return_T 10:57 so I had to declare ctor/dtor and add dummy save/load_serializable 10:57 I restored it 10:57 oh, I see 10:58 just check latest 10:58 I haven't noticed it 10:58 ok, thanks, I will check it 10:58 it was crashing python graphical example of multiclass 10:58 but now it works without any glitches 10:58 uricamic: could you implement a multiclass agrmax with loss? 10:59 _____________: I will look on it 10:59 so the design will be to have overloaded argmax one without loss and the other with it? 11:00 I think that is the idea 11:02 I'd like to skype with Nico first to ensure it 11:02 yes 11:03 I was pretty sure it would work without loss 11:03 but okay if it is not I don't mind 11:03 :) 11:03 well, I am still wondering that it was working, but it could happen :) 11:04 but I am sure, that the loss has to be there before argmax 11:05 n4nd0 thought it is only because of simple cases 11:05 I could believe so 11:06 yep, that is possible 11:06 I am preparing this case for multiclass classification to compare 11:06 http://25.media.tumblr.com/tumblr_m95fwdIEnn1qg3gfeo1_500.jpg hahah 11:07 blackburn why u no break somebody else's code? :D 11:07 lol 11:08 man it is easy to apply that patch 11:08 the lol was for the picture :D 11:08 wtf man 11:08 you broke the example :) 11:08 how? 11:08 i'll paste bt of gdb 11:08 [DEBUG] entering LatentSVM::apply(LatentFeatures at 0x10260bd90) 11:09 libc++abi.dylib: terminate called throwing an exception 11:09 [ERROR] Unknown problem type 11:09 Program received signal SIGABRT, Aborted. 11:09 0x00007fff8caff212 in __pthread_kill () 11:09 oh okay 11:09 that can be let 11:09 me fix it 11:09 i'mjustaying 11:10 wiking: well you are not very active with that code so I didn't think I could make it hard to rebase :) 11:10 :))) 11:11 all changes in your code are just 11:11 apply -> apply_latent 11:11 it doesn't mean that i'm not preparing a patch 11:11 yeah i know 11:11 uricamic: I don't manage to make so_multiclass_BMRM work properly 11:11 that is the right way to do that so just rename it and you wouldn't really need to rebase 11:11 and as well the fuckup of yesterday's primalmosek :) 11:11 uricamic: I execute on command line ./so_multiclass_BMRM data.out 2 1000 500 1 0.01 BMRM 11:11 n4nd0: what do u mean properly? 11:11 and it stays in iteration 0 forever 11:11 i mean the only thing is that please commit if it compiles 11:12 and not when it doesn't 11:12 and not having mosek is not an excuse when u edit that code ;) 11:12 * wiking feeling bossy today :) 11:12 shogun: Sergey Lisitsyn master * r5347c30 / src/shogun/machine/Machine.cpp : Added missed latent problem handling in apply of machine - http://git.io/8EdOQQ 11:12 _____________: I have to support wiking there, I am sorry :S 11:12 n4nd0: hmm,  I think I have changed that example to expect data in svmllight format 11:12 but let me check that 11:12 uricamic: aaahm I am just using the data from the example I run 11:13 I am just compiling the latest code 11:13 uricamic: ok, let me know someting 11:13 well guys I have nothing to say if you are so worried with breaking your code :) 11:13 :) 11:14 :> 11:14 I guess that we all want to rush a bit the things in order to be ready for release 11:14 btw what has happened with fixing up the static analyzer problems? 11:14 I have been waiting for generic risk function for a while and nobody did it 11:15 hmm, it seems that I have to modified that example, cos' it refused to compile now :D 11:15 furthermore we had wrong interface with redundant features 11:15 _____________: a PR + wait for comment would be nice... it really worked out for me till now quite well 11:15 so having no mosek is excuse, it is not project of mine but I did that 11:15 _____________: but you've broken the code, and you couldn't see the problem as that code never compiled on your machine 11:16 uricamic: oh yes, you have to change the constructor 11:16 DONT FUCKING BREAK THE CODE! :) 11:16 uricamic: just leave out features 11:16 uricamic: and thank _____________  :P 11:16 n4nd0: I see :D 11:16 uricamic: :DDD 11:16 wiking: I know what to do better 11:17 _____________: u mean u know better? :) 11:17 breaking mosek is not the issue when we are in rush to release sooner 11:17 but then we have a broken release :) 11:17 especially that mosek is not tested on any of the bots 11:18 n4nd0:  ok it works for me now 11:18 _____________: we have to face that it will probably be difficult to have everything working with bundle methods for release 11:18 i.e. nobody else will pick up this error just who really wants it to work :) 11:18 n4nd0 will always test it because he works with it 11:18 n4nd0: I can't see any difficult there 11:18 you said HM argmax with loss is easy? 11:18 ok i was mostly joking till now 11:18 I didn't say it was easy 11:19 I am just guessing 11:19 so you can't do that? 11:19 but seriously is it really hard to wait 6 hours for comments on a PR _____________ ? 11:19 if yeas 11:19 then mea culpa 11:19 n4nd0: so the input data are expected in svmlight format 11:19 _____________: I didn't say I can't 11:19 I have no fucking idea when will you appear wiking 11:19 _____________: i dont have to appear to comment on a PR on githug 11:20 what to wait for? 11:20 _____________: I just say that maybe it cannot be done for today 11:20 *github 11:20 I will do that then 11:20 i.e. on each line first number is the label and then non-zero elements of feature vectors in format idx:value 11:20 especially if you add a @ into the comment 11:20 as that person will get an email most probably 11:20 as that's the default setting of github... 11:20 wiking: what should I wait for? 11:21 _____________: comment on your intended PR 11:21 from the author of that actual code 11:21 which PR? 11:21 apply_latent? 11:21 _____________: i'm talking about future 11:21 that if u change code 11:21 which u really dont feel sure about 11:21 then maybe go with a PR and wait 3-6 hours 11:21 in future I would never get you back 11:21 what is unsure in that PR? 11:22 and if there's no comment 11:22 it should be apply_latent but not apply 11:22 did you ever check how machine works? 11:22 it's not about that given commit 11:22 same goes with primal mosek per se 11:22 and this is not a personal vendetta 11:22 did you check logs where I asked n4nd0 to check if it compiles? 11:22 so please do not start to patronize me with "did you ever check how machine works?" 11:23 i'm just suggesting 11:23 that maybe it would be better... 11:23 but if you know better 11:23 then it's ok 11:23 should I spend time on installing mosek just to check if it compiles when I could ask n4nd0 and he don't mind to do that? 11:23 sorry mate to bringing up this idea 11:23 I do not understand what you do blame me about 11:23 i dont blame u 11:24 i'm just saying that it would be great that if you start changing code that you haven't authored and it's not really an obvious assertation 11:24 then maybe it would be great that you somehow assure that it's ok 11:24 what is not obvious? removing features or renaming apply to apply_latent? 11:25 _____________: i've given there the example: not obvious = not an assertation 11:25 you claim I do that - where did I something that needs an author comment? 11:25 all these commits 11:26 that actually broken something 11:26 which these? 11:26 see the mentioned ones above 11:26 or now you really want to have sha-1 lines? 11:26 i mean seriously you are now going to take this as a personal matter 11:26 ? 11:26 your example should work now it is just a missed line 11:26 because sorry mate i really like you so i rather just shut up 11:27 I do not understand why making it correct is so much an issue for you 11:27 yes because you are blaming me in things I didn't do 11:27 I am spending time on things I am unsure you will spend on 11:28 lately I have been fixing doc - wiking your code too - I don't know whether you would do that 11:28 this conversation makes me sad :( 11:28 as for your example I planned to run it later anywhay 11:29 I just wanted to make your projects more suitable before release that is coming already 11:30 ok 2 more questions: a) have you commited code that broke examples and compilation of primalmosek? (and you have your answer on your line of 'blaming you') b) why do you feel attacked here, i think this really a work matter and nothing personal? but anyhow if you feel in any ways offended, then i'm honestly sorry and let's just forget this... it's really not worth it... 11:30 you are not in active development (at least it looks like that for me) 11:30 a) yes, and fixed that after n4nd0's report 11:31 b) because you are attacking me - pretty obvious 11:31 n4nd0: can you finally assure wiking I didn't break your code without your participation at all? 11:32 as said in b) and earlier it wasn't meant to be attacking at all just a suggestion what would be great... 11:32 or you think too I am the breaker that makes your projects worse? 11:32 dude 11:32 you completely take this the wrong way 11:33 nobody ever told that "I am the breaker" 11:33 and that you are making the project worse 11:33 you are taking this out of proportion 11:33 i haven't told any of this 11:33 uricamic, _____________ : nice! I finally finished the test 11:34 n4nd0: great :) 11:34 and wasn't even nearly implying it... 11:34 uricamic, _____________ : so I have compared the error when using the CMulticlassModel::risk and the error when using CStructuredModel::risk 11:34 the same data, same epsilon, same lambda, everything the same except from that function 11:35 when using CMulticlassModel::risk we get an error equal to 21.4% 11:35 and when using the other around 40% 11:35 _____________: so I finally conclude that bot argmax are not the same :) 11:36 n4nd0: I don't mind that but just do that correct 11:36 n4nd0: yep, of course they can't be :) 11:36 I am going for lunch now, will be back hopefully in less than 1 hour 11:36 uricamic: bye 11:37 _____________: c'mon man, don't say it so :( 11:37 how so? 11:37 never mind 11:37 I just mean if you have no time for that 11:38 just tell me 11:38 I will try to do by myself 11:38 look _____________, I have time for that 11:38 but time in my time frame 11:38 I may have done it by the end of this week, tested and everything 11:38 because first I want to consult with Nico 11:39 if you do not like that ..... just do it yourself 11:39 I want that to be in release I am sorry I am pushing you 11:39 _____________: I will do my best then 11:39 it seems to be the best new feature of shogun 11:40 _____________: then I will get more motivation :) 11:40 so I don't want to fuck up there 11:40 all right 11:40 so argmax is just viterbi in hm? 11:41 yes 11:41 can't you just put losses in that loss matrix? 11:41 I am newbie in HM still but just ask 11:41 let me check one thing in the hmsvm toolbox 11:42 it is easy to implement for multiclass 11:43 shogun: Soeren Sonnenburg master * r6d43e98 / src/shogun/features/DataGenerator.cpp : include lib/config.h when checking for a DEFINE_TO_BE_SET - http://git.io/cTCq-w 11:43 really easy 11:43 http://pastebin.com/KPkuwjta 11:44 take a look to that code 11:44 sonney2k: it is already included in .h - why? 11:44 _____________: you can see there are two calls to best_path 11:45 hah so it uses shogun's viterbi? 11:45 in the past yes, not in that version 11:45 lol 11:45 because our HMM is broken 11:45 and not because of me wiking 11:45 I think it was just to make it standalone... 11:46 at least they say so in the doc 11:46 n4nd0: 59-64 lines are the thing, right? 11:47 yes 11:47 and that is actually what I do in my code 11:47 I think that my argmax is looking for the MMV 11:47 what is MMV? 11:48 it is always doing the pred_path_mmv and not the pred_path 11:48 maximal margin violater 11:48 it appears in that snippet, in a comment 11:48 _____________ as i said earlier i'm sorry that i've brought up this whole topic; it was completely unrespectful from me to say anything like that.... 11:48 ah right sure 11:48 wiking: you are right in what you are saying 11:49 but 11:49 I didn't break someone code if it is not obvious fix 11:49 or doc fix 11:49 or I didn't talk to author 11:49 obvious fix is your case 11:49 in case of mosek svm I was in touch with n4nd0 - no idea why he refuses to say that :D 11:50 because I do not agree in the way you are handling this 11:50 but anyway ... let's move on, this is not so important 11:51 okay one more guy against me 11:51 so 11:51 as i said it was my fault so i hope we can get over this 11:51 _____________: you see, it is not against you or with you man 11:51 JESUS 11:51 why didn't you say I should install mosek and do not touch it yesterday? 11:51 please 11:51 _____________: I don't know how you say that if I am not even taking part .... 11:51 you said okay I will compile' 11:51 didn't you say that? 11:51 -!- emrecelikten [~emre@trir-5d8000e4.pool.mediaWays.net] has joined #shogun 11:52 all right guys 11:52 let's just get over this and i'm sorry for bringing this up 11:52 it's not worth any of this 11:52 sonney2k: captain, bring some peace please :) 11:53 I give up guys 11:53 _____________ there's nothing to give up, keep on as is and it's all good 11:53 let you work on that separetely then if you are so hurted 11:53 -!- emrecelikten [~emre@trir-5d8000e4.pool.mediaWays.net] has quit [Read error: Connection reset by peer] 11:53 _____________ nobody is hurt here.... 11:54 except maybe you 11:54 with commits I have to do without much checks 11:54 but that wasn't any of my intention 11:54 and since it escalated into this that you think i'm saying that you are makeing bad for the project 11:54 which is OBVIOUSLY not the case... it's the opposite 11:55 you are pushing this project like nobody else here 11:55 which is awesome 11:55 +1 11:56 so please just try to forget my comments and https://i.chzbgr.com/completestore/12/6/17/u5J1lVbU6UabQ5QinFqv0A2.jpg 11:56 I partly do understand your unsatisfaction 11:57 but I do not understand what made you so worried about that 11:57 neither of that has happened in me :) 11:57 neither unsatisfaction nor worried 11:57 it was really just a comment 11:57 no more no less 11:58 and u do whatever you want with it 11:58 just please dont take it as a personal vendetta against you and your personality 11:58 n4nd0: why didn't you say it is bad to do that blind fix? 11:59 and why do you feel it so bad if you could test that - as I asked you 12:00 I do not understand that 12:00 nyipp 12:01 _____________: I think I told you I could fix the MOSEK thing to ensure that it will compile, but you insisted 12:01 ????? 12:01 *booom* and the last 20 minutes disappeared :DDD 12:01 nonono 12:01 I don't think it is bad to the blind fix either 12:01 llalalalal 12:01 lalal 12:01 a 12:01 insisted what? 12:01 alalalal 12:01 aa 12:01 bad sounds quite bad 12:01 alalla 12:01 alala 12:01 lalalala 12:01 lalalalal 12:01 ala 12:01 ala 12:01 I wanted you to test if it compiles after removing features 12:01 (do not let you speeak :DDD ) 12:02 is that wrong for you? 12:02 no 12:02 of course not 12:02 what is? 12:02 ? 12:03 I want to understand what is you feel wrong in handling 12:03 you said you disagree 12:03 with what? 12:03 how you have reacted 12:04 but it was nothing bad 12:04 have i told that ever since a chick moved into my office my office became a geek magnet and all the guys are coming into my office chatting... :)) 12:05 :D 12:05 i'll be a true cockblocker :) 12:05 haha 12:05 i still need to work on some lines :) 12:06 _____________: seriously, I don't think you did wrong doing any fix there 12:06 then I am completely lost what I am wrong with 12:06 _____________: nothing! 12:07 that's the point :) 12:07 and i was just practicing my trolling skills :D 12:07 I understand you are comfortable with your own code and it is rather you should do changes 12:07 but it would take a while and I want to push latest must have features before release 12:08 and make it correct 12:08 _____________: agree 12:08 so is it over yet? 12:08 let's work together and have as many stable things as possible before release 12:08 I'll let you polish your code by yourselves then 12:09 wiking: so is your example worknig now? 12:12 _____________ yeah awesom-ooo-ly thnx! 12:12 wiking: why latentsosvm has training but has no apply? 12:13 _____________ because it's WIP 12:13 it'll be no WIP by this weekend 12:13 I see 12:14 we have this 'little' ongoing problem with structmodel and latentmodel and merging them 12:14 as currently creating a working latentSOSVM 12:14 takes too much user defined stuff 12:14 you have to define at least 4 different classes 12:14 and i think we shoudl take it down to 3 12:15 ok 12:15 I am sorry I am a bit offensive but I still don't really get why did you both blame me 12:15 I BLAME YOU FOR MISUNDERSTANDING ME! 12:16 u r the MISUNDERSTANDER! 12:16 :D 12:16 do not be that one :> 12:16 comeon man seriously 12:16 can we get over it? 12:16 yes but I want you to understand I am not a psycho 12:17 heheh i know you are not 12:18 :D 12:18 never was the case 12:18 you are just passionate about shogun 12:18 that would make you a geek but i would really like to think that you are not that type of a geek who would come into my office to try to chat with this chick :))) 12:19 :D 12:19 it is hard to talk about structured output machine with chick 12:20 hahahah 12:20 i hope they won't find these chat logs 12:20 :))) 12:20 or maybe i hope they'll find it :DD 12:20 I am sure no chick would ever try to read IRC LOG of MACHINE LEARNING TOOLBOX, I am pretty sure 12:21 * wiking works at a very unpleasant place so that's why his implied complaning :) 12:21 which office you are talking about? your university? 12:21 heheheh she is doing phd in maths... 12:21 i mean it's not soooo unprobable 12:21 :))) 12:21 yeah my uni :) 12:22 ah yeah cool 12:22 i know how i'll cockblock 12:22 i'll use banners 12:22 :D 12:22 i'll print some banners and just raise them sometimes when i get annoyed with the jibberish :) 12:22 _____________: I think that the argmax I implemented does: argmax_y [ Delta(yi,y) + ] in training 12:23 _____________: and argmax_y [ ] during prediction 12:23 n4nd0: it should do that yes 12:23 lol 12:23 so if you have some good ideas for the banners let me know :) 12:23 wiking: but the definition of argmax is: y* = argmax_y or? 12:24 n4nd0: well that's the thing 12:24 n4nd0: can't we assume training is for +LOSS and not-training is w/o loss? 12:24 _____________: do you mean for every model, or just for HM-SVM? 12:24 n4nd0: it should be that in your case but i think you are trying to find the most violating ones 12:24 for all models 12:24 _____________: I am not sure about all models 12:24 _____________: let me check what I did in the multiclass one 12:25 not-training w/o loss? 12:25 no comply in me please explain 12:25 wiking: there are two modes in argmax 12:25 first is training one 12:25 _____________: in the multiclass the loss is not taking into account 12:25 n4nd0: we should change that then 12:25 but classification (not-training?) is just w*psi(x,y) 12:26 n4nd0: okay wait - so HM is correct because it was correct alreadY? 12:26 wiking: yes 12:26 _____________: why? I don't know which one is the correct :D 12:26 wiking: yes 12:26 n4nd0: you said it does loss+w max 12:26 so what else is not-training? :))) 12:26 prediction 12:26 exactly 12:26 well then u don't need that 12:26 :D 12:26 when you don't have the true label, you cannot get the loss 12:26 it's just w*psi() 12:26 no, it is argmax of w*psi 12:27 really? 12:27 ah ok 12:27 i get it 12:27 most probable label 12:27 or i would get it in latent case 12:27 but in simple struct as well? 12:27 http://www.cs.cornell.edu/people/tj/publications/tsochantaridis_etal_04a.pdf 12:27 look guys at algorithm 1 12:27 oh shit I am getting crazy confused 12:28 what to do with it? 12:28 n4nd0: if you are optimizing with loss term already 12:28 lets just remove it from generic risk 12:28 and put loss to multiclass argmax on training=true 12:29 I don't understand, sorry 12:29 n4nd0: you said you compute argmax with loss already, right? 12:29 in the HM-SVM 12:30 yes 12:30 not in the multiclass 12:30 we can fix multiclass - it is not the issue 12:30 I mean currently it is being added to risk 12:30 in generic risk 12:30 https://github.com/shogun-toolbox/shogun/blob/master/src/shogun/structure/StructuredModel.cpp line 170 12:31 n4nd0: if we assume argmax includes loss on training we remove line 170 12:31 and add loss to multiclass 12:31 then everything becomes correct and we are all happy 12:31 the thing is that I am not sure whether the argmax should include the loss 12:32 even if we are in training 12:32 ehm 12:32 you was sure yesterday 12:32 and uricamic said it should 12:32 I was unsure it should 12:32 mmm 12:32 well risk is being used in bmrm code 12:33 okay I am off for food 12:33 and there the primal objective in some way will include using argmax 12:33 alas the risk 12:33 or? 12:33 _____________: yesterday I was sure that the argmax doesn't include the loss 12:33 _____________: that's why I said that it is not the same the argmax that was already implemented in the model than the one we need for risk 12:34 n4nd0 what i dont get is that how u want to have a generic risk, where what the bmrm solves is actually labmda*||w|| + RISK 12:34 so it pretty much depends on you what RISK -> p.o. is 12:34 but the risk can be written in terms of an argmax and the psi function 12:35 well it depends what's RISK's form 12:35 :) 12:35 so the generic risk calls these model dependent functions 12:35 yeah i get it 12:35 but the risk can be pretty much anything 12:35 :) 12:35 or? 12:35 R({\bf w}) = \sum_{i=1}^{m} \max_{y \in \mathcal{Y}} \left[ \ell(y_i, y) + \langle {\bf w}, \Psi(x_i, y) - \Psi(x_i, y_i)  \rangle  \right] 12:36 until it's convex... ;) 12:36 I thought the risk is that, and only that 12:36 this is from multiclass right/ 12:36 ? 12:36 (paste here for better reading) 12:36 http://www.codecogs.com/latex/eqneditor.php 12:36 i know this one 12:37 wiking: why just the multiclass one? 12:37 but from where did you get this comment? 12:37 MuticlassModel.h 12:37 yeah 12:37 afaik it's a good risk function for multiclass 12:37 it is the same that appears in Teo 12:38 yeah i know :) 12:38 page 103, equation A.2 12:38 and a lot of other places with SO 12:38 they don't talk there just about multiclass, right? 12:38 noup 12:38 wait a sec 12:38 wait 12:39 i'll just dig something out 12:39 ok so here's another thing 12:39 hahahah A FUCKING GEEK CAME TO MY ROOM BUT SINCE SHE IS NOT HERE HE DIDNT EVEN ENTER THE ROOM 12:39 how fucking lame is this :))) 12:39 they all want the pussy :))) 12:39 so yeah n4nd0 look 12:40 :D 12:40 http://www.cs.cornell.edu/~cnyu/papers/icml09_latentssvm.pdf 12:40 page 3 eq(7) 12:40 for me that is the RISK 12:40 with the -C\sum_.... 12:40 the whole thing after 1/2||w|| 12:41 aham 12:41 yeah, that is definitely different 12:41 that term is new 12:41 and the Dual solver 12:41 basically is intented to solve 12:41 anything that is convex and in labmda*||w||^@ + RISK 12:42 w/o constraints 12:42 so 12:42 until risk is convex 12:42 you can have anything 12:42 as risk 12:42 I see 12:42 this is my understanding and that's why i thought that StructuredModel::risk should not be implemented 12:42 n4nd0: but in multiclass case 12:42 you are totally right though 12:42 you can actually define the risk function with the help of psi and argmax 12:43 but what if is your risk is something completely different but still convex 12:43 see for example my case 12:44 but maybe this risk function we use in multiclass, and probably in HM-SVM too, may work as well 12:44 and yes in my case \hat{y} and \hat{h} should be defined/given by argmax 12:44 yeah it will work 12:44 until you want that risk function 12:44 then we just make it virtual 12:45 yeah 12:45 but i think you should note it somewhere that this is just an example ;) 12:45 giving the possibility to be overridden 12:45 yeah as until now 12:45 wiking: sure 12:45 in any case 12:45 probably it is not that great to have this generic risk for efficiency reasons 12:46 but i thought that it's a bit 'narrowing down' when you already implement a risk function in an abstract class whilst your risk function could be anything 12:46 anything = the PO is convex... 12:46 yes, I see your point and I agree with you 12:46 but 12:47 it is not conceptually right I think 12:47 there could be a protected default risk function ?:) 12:47 in structuredmodel 12:47 and then anybody who inherits from this class 12:47 will have already a function at hand 12:47 and he can just chuck that in into his risk function 12:48 (call it from it) if he really doesn't know what is he doing :))) 12:48 sounds like a good idea 12:48 or just wants to compare 12:48 bmrm vs mosek 12:48 btw have any of you run that ? 12:48 what's the accuracy difference between mosek and bmrm 12:49 on one given data set 12:49 not using an important data set 12:49 hehe yeah 12:49 just randomly generated Gaussian 12:49 but any dataset 12:49 yeah 12:49 for the multiclass example 12:49 and what are the numbers? 12:49 yeah, I have prepared that before 12:49 how far r they from each other ? 12:49 running again right now .. 12:50 but I don't think these results are very trustworthy 12:50 I just plugged a lambda at random 12:50 SO-SVM: 92.60% 12:50 BMRM:   88.60% 12:50 MC:     40.00% 12:50 okey 12:51 so bmrm is not so far at all 12:51 btw: labmda = 1/c 12:51 no 12:51 what's your c in primal? 12:51 just use lambda = 1/c 12:51 wiking: michal said something about that BUT also that he was not regularizing with the #examples 12:51 C = 1 in multiclass classification her 12:52 but it's pretty cool though 12:52 for the PrimalMosekSOSVM 12:52 and for mosek? 12:52 ah ok 12:52 then use lambda =1 12:52 :) 12:52 I am interested in knowing how faster it is 12:52 n4nd0: CMath::time 12:52 :P 12:52 but still it's awesome that there's bmrm that can actually give similar values as mosek but it's fully opensource 12:53 it is pretty slow with lambda = 1 12:53 https://fbcdn-sphotos-a-a.akamaihd.net/hphotos-ak-ash3/580085_462759493754616_184994591_n.jpg 12:54 :D 12:54 ok i'm off to pick up something to eat 12:54 lol 12:55 brb in 10 12:55 for a new run with other data and lambda = 1000 (#examples) 12:55 SO-SVM: 85.30% 12:55 BMRM:   79.20% 12:55 MC:     54.10% 12:55 >>>> PrimalMosekSOSVM trained in   22.0718 13:01 >>>> BMRM trained in    8.6328 13:01 SO-SVM: 85.10% 13:01 BMRM:   81.90% 13:01 MC:     46.30% 13:01 http://29.media.tumblr.com/tumblr_lltzgnHi5F1qzib3wo1_400.jpg 13:01 back again, I had to read quite a big log :) 13:07 so for the argmax question 13:07 I said, that it is reasonable to have overloaded argmax function, which in training will get loss and therefore it will find the most violated constraint 13:08 and the standard argmax function which should be used for prediction of with trained W vector 13:09 uricamic: should it argmax including loss? 13:09 only in training 13:10 n4nd0: what is MC? 13:10 uricamic: sure 13:10 uricamic: my suggestion is 13:10 there is a training flag in argmax 13:10 but then for predictions it shouldn't use it 13:10 _____________: ok 13:10 in case of training it includes loss 13:10 _____________: linearMulticlassMachine with LibLinear 13:10 in other cases no 13:10 _____________: see libshogun/so_multiclass.cpp 13:11 uricamic: good for you? 13:11 _____________: ok, that should be fine then 13:11 yep 13:11 n4nd0: agree? 13:11 sure 13:11 then I had not fucking idea what the argmax should do 13:11 I feel stupid 13:11 ehhm I do not understand 13:11 I was wrong not you 13:11 because then it fits the design and it is easy to understand what is going on there 13:11 _____________: I thought the loss is not included in the argmax 13:12 ????? 13:12 so uricamic 13:12 what we were arguing about last night? 13:12 :D 13:12 uricamic: can you please check MulticlassModel.cpp 13:12 argmax should do exactly what the name says - find argument where the dot product is maximal 13:12 I was pretty sure my position is that argmax with loss is equal to argmax without 13:12 but it was wrong 13:12 _____________: yeah, but I am not talking about that 13:12 I thought that the argmax must not include the loss 13:13 to score? 13:13 and in training it has to find this maximum enhanced with losses -> i.e. find the most violated example 13:13 I understand 13:13 n4nd0: the version which now in shogun or somewhere else? 13:13 shogun: Sergey Lisitsyn master * r50ba31c / src/shogun/structure/StructuredModel.cpp : Removed redundant loss in generic risk - http://git.io/wcQu-w 13:13 uricamic: let me paste it for you 13:14 n4nd0: ok 13:14 http://pastebin.com/iK0xxBf1 13:14 that is used for training currently 13:14 and if I understand everything properly now 13:14 that is kind of wrong 13:14 since for training 13:14 the score should be 13:15 n4nd0: uricamic so should we add return_value->delta in risk? 13:15 uricamic: what is computed in that snippet in line 8 + the loss 13:15 n4nd0: yes it is wrong 13:15 it is not wrong for *applying* 13:15 but it is wrong for *training* 13:15 n4nd0: I miss the loss there 13:16 uricamic: exactly 13:16 uricamic: n4nd0 only in case of training right? 13:16 _____________: I believe so 13:16 I can fix that 13:17 it is ok for prediction but for training u have to add loss there to find the maximum score on features with loss 13:17 or anybody is willing to? 13:17 _____________: I can do it 13:17 okay please do then 13:17 uricamic: should we expect better classification results after this change? 13:17 n4nd0: I am unsure we need delta parameter in result type 13:17 _____________: that was designed by Nico 13:18 if it is being added in case of training to score 13:18 we won't use it anyway 13:18 n4nd0: well I would expect that, because if you haven't counted with loss in training u actually trained different classifier 13:18 agree? 13:18 _____________: I am not sure if it is used somewhere right now 13:18 _____________: did you check? 13:18 _____________: what do u mean by resturn_value->delta? 13:19 it is being set but it won't be used by risk 13:19 *return_value 13:19 because score contains it 13:19 uricamic: https://github.com/shogun-toolbox/shogun/blob/master/src/shogun/structure/MulticlassModel.cpp 109 13:19 we don't need it, right? 13:19 probably not 13:20 because we don't need in when applying 13:20 it* 13:20 and it is included when training 13:20 n4nd0: agree? 13:20 I wouldn't remove it right away 13:21 I guess Nico put it there for some reason 13:21 but no idea 13:21 okay no need to remove it 13:21 but lets remove it's setting in multiclass model 13:21 okay I'll wait for your code :) 13:26 n4nd0: we may do that with delta though 13:29 n4nd0: we then would need to add it in risk function but subtract from maximum score 13:29 see what I mean? 13:29 n4nd0: just a little bit of comment on CStructuredModel. dont u wanna pass const SGVector& w in argmax? 13:29 n4nd0: just to reserve some space on the heap 13:30 i mean stack 13:30 ;) 13:30 otherwise copy constructor is being called etc etc etc 13:30 please do that somebody 13:30 or I will 13:30 :D 13:30 and actually 13:31 in risk function 13:31 what is in risk function? 13:31 don't we want to use SGVector instead of the simple pointers? 13:32 where? 13:32 float64_t risk(float64_t* subgrad, float64_t* W, TMultipleCPinfo* info=0); 13:32 float64_t* -> SGVector 13:32 why to use it? 13:32 we could but any specific reason? 13:32 since we have sgvector as a wrapper for simple type arrays 13:33 no other reason 13:33 wiking: to make it explicit I would keep it 13:34 it makes you think you are going to write to given memory 13:34 we need to add director model before release 13:37 I will do that 13:37 oops I am at job 13:38 damn 13:38 _____________: working again? 13:38 yes starting from this week 13:39 same place? 13:39 yes 13:39 netcracker 13:39 it is worth to mention that netcracker supports shogun implicitly :D 13:39 \o/ 13:40 put it in the release notes :> 13:40 because I do shogun instead of my job 13:40 :D 13:40 (got that part ;P) 13:41 n4nd0: wtf is with CLatentData CData and CStructedData? 13:43 wiking: you talkead with sonney2k about that right? 13:45 he didn't like CData I think 13:45 uricamic: does BMRM use argmax somewhere? 13:45 n4nd0: ?? risk 13:46 BMRM uses risk function, and risk function should use argmax inside 13:46 call directly to argmax 13:46 n4nd0: i remember sonney2k telling us that he doesn't like the naming... that was as far as i understood :) 13:46 n4nd0: that's why i'm asking to what now? 13:46 _____________: dont u wanna fix this? :) 13:47 wiking: I don't know ... I am ok with anything regarding that to tell the truth :) 13:47 and i'm TOOTALy not being ironic here :) 13:47 wiking: fix what? 13:47 as much as it sounds :> 13:47 _____________: so that we have this shit going on 13:47 _____________: we have CLatentData (actually we had now it's CData) and CStructuredData 13:47 and basically they are serving the same exact purpose 13:48 uricamic, _____________: so ... the accuracy doesn't change at all in this multiclass example 13:48 _____________, uricamic: training with argmax using or not using loss 13:48 you mean with argmax containing losses in training? 13:48 _____________: so i thought it would be great to remove this redundancy and merge CLatentData and CStructuredData and just have CData 13:48 :) 13:48 _____________: but sonney2k didn't like the name 13:48 _____________ so we need a new name and that's it basically ;) 13:49 n4nd0: it might be for multiclass (as far as I understand) 13:49 wiking: why latent is not structureD? 13:49 _____________ because it's not :) 13:49 uricamic: yes 13:49 n4nd0:  but it could be true just for some particular example, definitely not generally 13:49 uricamic: ok 13:49 i mean they are semantically not the same 13:49 wiking: what is different in latent? 13:49 uricamic: I am using also loss equal to 1 or 0 so it might normal that it doesn't make a huge difference 13:49 _____________: that it's latent and not fixed contrary to structured 13:49 _____________: other than that it's the exact same implementation :) 13:49 n4nd0: yep, that is another important fact 13:50 _____________: it's just semantics 13:50 _____________: but imho important one... 13:50 when u use e.g. L2 loss the difference should be more obvious 13:50 wiking: okay lets get back to that later 13:51 uricamic: it converges faster including the loss though, for the PrimalMosekSOSVM case 13:52 n4nd0: what did you change now? 13:53 n4nd0: you mean less number of iterations? 13:53 I am just comparing what happens when the argmax includes the loss or not 13:53 uricamic: I am just looking at the training time 13:53 uricamic: but I guess it is because the number of iterations is smaller 13:54 n4nd0: I see, but it should be reasonable since without loss it could happen that sometimes not true most violated constraint is found and therefore it needs more iterations to converge 13:54 uricamic: yeah, could be 13:55 _____________: http://i.imgur.com/zQHv4.gif 13:55 :DDDDDDDDDDDDDDDDDDDD 13:55 uricamic: you are an expert in this man! 13:55 megalol 13:55 :> 13:55 n4nd0: :D I wouldn't call myself one, but I have already written one paper about that 13:56 actually 2 but unfortunately the second one was not accepted, since someone has already done something similar -> axed by reviewers immedeately 13:56 too bad 13:57 it is life :) 13:57 uricamic: btw, I would really really prefer if the MulticlassModel::risk calls to argmax and delta loss 13:57 uricamic: what do you think? 13:57 n4nd0: sure, it is no problem 13:57 I will get to this hopefully soon 13:57 uricamic: do you want to change that? 13:57 n4nd0: yep I can, but if u want to do that feel free :) 13:58 since I have to write TR for my work which could be then referenced by shogun tutorial 13:58 I think I am going to stop soon of shoguning today ... 13:59 I should have done something for the job this morning 13:59 wiking: yes that made all russian twitter day yesterdday 14:00 n4nd0: uricamic where did you stop? 14:01 is it ok to include loss in multiclass now? 14:01 can I commit that? 14:01 _____________: you mean loss in multiclass argmax? 14:01 yes 14:01 _____________: then yep 14:02 if nobody of you are willing to do that I'll do 14:02 :) 14:02 n4nd0: job? 14:02 _____________: yeah 14:03 my professor at KTH gave me some things to do to help out in the robotics course this year too :) 14:04 teaching is great 14:04 he didn't tell me if he needs me in the lab this year too, for the moment I am just setting up some material 14:05 so to sum up 14:06 now both Multiclass and HM-SVM should work good with BMRM methods 14:06 uricamic, _____________: agree? 14:06 n4nd0: well 14:10 if HM-SVM argmaxes with loss 14:10 does it? 14:10 yes 14:11 okay nice 14:11 then yes, but multiclass needs to be changed 14:12 n4nd0: does structure_hmsvm_bmrm.py produce 99% still with latest changes? 14:12 it is changed 14:12 changed where? 14:12 by you - you mean? 14:13 if so let me merge it ;) 14:13 PR going 14:13 python structure_hmsvm_bmrm.py 14:15 Accuracy = 0.9989 14:15 I ask because I removed loss from risk 14:15 I assume it is correct and score includes it 14:16 right? 14:16 let me think about it a moment 14:16 but I think one term needs to be substracted from the risk 14:18 psi_truth in particular 14:18 yes 14:18 multiclass has that problem 14:18 well not just psi truth 14:19 but 14:19 yes 14:19 is it true for HMModel too? 14:20 I am not sure for that 14:20 for example 14:20 I don't see directly that the score in the HMMModel includes the loss 14:20 although it may be since a matrix used for Viterbi decoding includes the loss term... 14:21 I am not sure regarding this 14:21 I will ask Nico about it tomorrow 14:21 or maybe uricamic can tell us about it 14:21 n4nd0: why did you make first -INFTY? 14:22 this way you disallow it to select 0th class 14:23 if you do it this way then c=0 initially 14:23 -!- av3ngr [av3ngr@nat/redhat/x-swaimdyxogxxwdlk] has quit [Quit: That's all folks!] 14:24 mmm but that is weird 14:29 the accuracy remained the same after that fix 14:29 yeah of course it'd remain the same if I don't compile .... 14:30 n4nd0: you talking to me? (c) ? 14:31 n4nd0: sorry, I have been away for a while, what is the question? 14:31 looking better now 14:31 wiking: talking alone I think... 14:31 n4nd0: hehe it was a quote from the taxi driver 14:32 n4nd0: c=0 14:32 uricamic: I need a rest now ... 14:32 please :) 14:32 n4nd0: ok :) 14:32 n4nd0: where is your power? 14:32 _____________: I updated the PR 14:32 n4nd0: http://www.youtube.com/watch?v=bqLyTdcMLhc 14:32 n4nd0: for your rest ;) 14:32 I lost my power 14:33 shogun: iglesias master * r1a87bf0 / (3 files): * fix MulticlassModel argmax, the loss must be included in training to find the MMV - http://git.io/l1c4ig 14:33 shogun: iglesias master * r80f82b5 / (3 files): + extension in so_multiclass example and * fix in multiclass BMRM - http://git.io/ihzbGw 14:33 shogun: iglesias master * rbbb3457 / src/shogun/structure/MulticlassModel.cpp : * wrong looop initialization - http://git.io/gxDe4Q 14:33 shogun: Sergey Lisitsyn master * r8cec5d0 / (6 files in 2 dirs): Merge pull request #752 from iglesias/master - http://git.io/GRG4TQ 14:33 I think i'll do the dishes during my rest instead 14:33 build #333 of bsd1 - libshogun is complete: Failure [failed test]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/bsd1%20-%20libshogun/builds/333  blamelist: iglesias , Sergey Lisitsyn 14:44 lol 14:45 shogun/io/streaming/StreamingAsciiFile.h 14:46 shogun: Sergey Lisitsyn master * r2f999e4 / examples/undocumented/libshogun/so_multiclass_BMRM.cpp : Update examples/undocumented/libshogun/so_multiclass_BMRM.cpp - http://git.io/PoHK1g 14:48 may the ILSVRC people rot in hell 14:48 ? 14:48 ILSVRC2012_val_00004220.JPEG 14:48 JPEG 14:48 and capital 14:48 wtf :)))) 14:48 oracle loves to capitalize words 14:49 and putting tars in tar is really a smart idea... 14:49 putting a tar archiver into a tar is way smarter 14:49 heheh 14:49 yeah that'dbe 14:49 but still now i'm running low actually on space on the cluster :) 14:50 i mean not the big hdfs 14:50 but the local little /home fs 14:50 how big it is untared? 14:50 well i guess it's the same size 14:50 yeah makes sense 14:50 so i would need another +138 gigs 14:50 but i think i've found a way around here :) 14:50 since the / has 800 gigs free on one node 14:51 that is part of the free space of hdfs 14:51 buuut i could use that for temp time 14:51 they put 1000 tars within that big tar 14:53 i think the 1000 tars is the different classes 14:53 lol 14:55 this is nice 14:55 actually they've provided a matlab file with all the dsift features within a class 14:55 * wiking wonders if it's all the images in the class 14:55 dense sift? it must be way too big to be large scale 14:55 it is but still 14:56 it's there 14:56 however what is the other way? 14:57 matching? 14:57 or bow like yeah 14:57 mmm lol yeah 14:57 we have already all the sifts if we want it :) 14:57 marvolous 14:57 wiking: what is deadline for ilsvrc? 14:58 23rd of sept 14:58 I see 14:58 -!- _____________ [5bdfb203@gateway/web/freenode/ip.91.223.178.3] has quit [Ping timeout: 245 seconds] 15:07 build #351 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/351  blamelist: iglesias , Sergey Lisitsyn 15:20 build #444 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/444  blamelist: iglesias , Sergey Lisitsyn 15:45 build #334 of bsd1 - libshogun is complete: Failure [failed test]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/bsd1%20-%20libshogun/builds/334  blamelist: Sergey Lisitsyn 15:50 -!- alexlovesdata [~binder@goldenezahl.ml.tu-berlin.de] has joined #shogun 15:58 build #445 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/445  blamelist: Sergey Lisitsyn 16:10 -!- blackburn [~blackburn@188.168.13.28] has joined #shogun 16:12 build #352 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/352  blamelist: Sergey Lisitsyn 16:12 time to fix things 16:13 blackburn: btw I have checked IRC log, BMRM cannot be used with lambda=0 16:16 uricamic: sure 16:16 that's clear 16:16 and also, it should not be even close to 0 16:16 well it converges with small lambda 16:16 but it makes no sense 16:17 it is just overfitting 16:17 blackburn: yep, I just wanted to point that out just for sure :) 16:17 the only reason I did that 16:17 to try to fit model perfectly 16:17 on training data 16:17 with no regularization at all 16:17 well, it can happen that for some instance and features the regularization term is not needed, by then it is not good idea to use BMRM 16:17 blackburn: I see, but BMRM is not designed for such things, lambda has to be >0 16:18 uricamic: btw multiclass model is not typical in bmrm, right? 16:18 I mean it is L2 of whole w 16:18 I am not sure that I understand the question now 16:19 ahhh nevermind 16:19 it is ||w||^2 so it is the same 16:19 I meant that in crammer-singer or weston one 16:20 \sum ||w|| is used 16:20 you mean the way how it is implemented in shogun now? 16:20 but it is the same in cases of ||w||^2 16:20 I see 16:20 read ^ as \sum ||w_i||^2 16:20 yep 16:21 uricamic: so are we ok now with multiclass bmrm? 16:22 uricamic: btw what is the difference between all these solvers? 16:22 I will check multiclass bmrm, soon, but I guess it should be ok now 16:23 the difference 16:23 BMRM is the standard method as described by Teo et al. 16:23 only the inactive cutting plane (ICP) removal strategy is added there 16:24 how one should choose a solver? 16:24 PPBMRM is BMRM enhanced by prox-term 16:24 kind of initial guess? 16:25 which is our method how to enforce possibility of some initial solution by restricting the L2 distance between consecutive W_t 16:25 our - your and vojta? 16:26 and the last one P3BMRM is the same  as PPBMRM plus it can use multiple cutting plane models, it is similar to Joachim's nslack 16:26 yep 16:26 when cp_models = 1 in P3BMRM then it is reduced to PPBMRM 16:27 and when you set K = 0 PPBMRM should reduce back to BMRM (but I haven't checked this particular one yet) 16:27 we have to make your example support default values 16:27 I mean in Shogun's implementation 16:27 you mean cpp example? 16:28 yes 16:28 ok, I put it there just for my convenience to test all algorithms quickly 16:28 I don't mind command line commands actually 16:28 but we need it able to run as is 16:29 like ./so_multiclass_BMRM and that's all? 16:29 yes 16:29 Makefile does that 16:29 when testing 16:29 i.e. to use randomly generated data and default settings 16:29 ok 16:29 yeah that would work I think 16:29 can you do that next time? 16:30 sure 16:30 the earlier the better because currently test is broken: http://www.shogun-toolbox.org/buildbot/builders/deb2%20-%20static_interfaces/builds/352/steps/test%20libshogun/logs/stdio 16:30 I think I will do this either in the evening or tomorrow 16:30 ahh, ok 16:30 so back to the solvers, if u are tuning the optimal lambda, it is nice to have some reasonable range of lambda's and in the cycle use PPBMRM or P3BMRM, since then it will use solution from the last lambda as initial one and converges much faster, especially for lower values of lambda 16:33 to put it in numbers, on MNIST data (10 classes, 784 feature dimension), 60k training examples 16:34 the difference between BMRM and P3BMRM on lambda=1 16:34 kind of hot start? 16:35 yep 16:35 nice 16:35 was from 2.01 hours to 2.09 hours for P3BMRM 16:35 shit 16:35 0.209 for P3BMRM :D 16:35 otherwise it would be so impressive :D 16:35 oh, again, wouldn't :D 16:36 2 hrs? 16:36 :D 16:36 in iterations it was from 5932 iterations to 408 16:36 uricamic: 0.209h sounds like it is pretty fast, something near to liblinear 16:37 and the other thing about P3BMRM is that it could be quite easily written to use more cores 16:37 blackburn: I haven't tried liblinear yet :) 16:38 but this multicore thing has probably reason only when computation of the risk takes a long time 16:39 I see 16:39 so in multiclass it would probably not give any speedup at all 16:39 that's clear 16:39 another thing is that both PPBMRM and P3BMRM tends to converge to more precise solution than BMRM because of our alpha setting strategy and termination condition 16:40 last two weeks I've been trying to employ L1/L2 tree regularization 16:41 no success :( 16:41 -!- blackburn1 [~blackburn@188.168.4.46] has joined #shogun 16:45 -!- blackburn [~blackburn@188.168.13.28] has quit [Ping timeout: 250 seconds] 16:47 at some point users must claim us as an OSS having best support ever 16:50 I will eat them if not 16:55 -!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] 17:02 -!- blackburn [~blackburn@83.234.54.112] has joined #shogun 17:03 -!- blackburn1 [~blackburn@188.168.4.46] has quit [Ping timeout: 244 seconds] 17:04 -!- blackburn1 [~blackburn@188.168.2.65] has joined #shogun 17:18 shogun: Sergey Lisitsyn master * r90d4424 / (9 files in 4 dirs): Refactored multiclass machines to support C parameter selection - http://git.io/rfAykA 17:20 -!- blackburn [~blackburn@83.234.54.112] has quit [Ping timeout: 252 seconds] 17:21 -!- uricamic [~uricamic@2001:718:2:1634:155e:7544:fba4:9878] has quit [Quit: Leaving.] 17:22 build #335 of bsd1 - libshogun is complete: Failure [failed test]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/bsd1%20-%20libshogun/builds/335  blamelist: Sergey Lisitsyn 17:27 ahahhahahahahaha 17:41 "otherwise my concern would be -- where is the unittest?   logic-wise 17:41 it should indeed accomplish the desired goal but imho test would help to 17:41 'guarantee' that ;) 17:41 " 17:41 :DDDD 17:41 people are damanind :) 17:42 demanding :P 17:42 so yeah where's the unit test :> 17:42 build #446 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/446  blamelist: Sergey Lisitsyn 17:55 build #353 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/353  blamelist: Sergey Lisitsyn 17:57 -!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has joined #shogun 18:02 hi all 18:02 yo 18:13 -!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has quit [Quit: Page closed] 18:28 -!- emrecelikten [~emre@trir-5d8000e4.pool.mediaWays.net] has joined #shogun 18:34 -!- Netsplit *.net <-> *.split quits: shogun-buildbot_, yoh, audy, emrecelikten, @sonney2k, naywhayare, sr___, wiking, zxtx, CIA-52, (+1 more, use /NETSPLIT to show all of them) 19:38 -!- Netsplit over, joins: CIA-52, sonney2k, audy, shogun-buildbot_, sr___, wiking 19:41 -!- naywhaya1e [~ryan@spoon.lugatgt.org] has joined #shogun 19:41 -!- Netsplit over, joins: yoh, emrecelikten, blackburn1, zxtx 19:41 -!- ServerMode/#shogun [+o sonney2k] by moorcock.freenode.net 19:41 build #336 of bsd1 - libshogun is complete: Failure [failed test]  Build details are at http://www.shogun-toolbox.org/buildbot/builders/bsd1%20-%20libshogun/builds/336  blamelist: Sergey Lisitsyn 19:43 -!- Netsplit *.net <-> *.split quits: yoh 19:45 -!- Netsplit over, joins: yoh 19:45 blackburn1: we were in a split 19:45 hehe 19:45 so any ideas there? 19:46 about?? 19:47 ah the func arg 19:47 have you got my two example cases for diff dims? 19:47 http://pastebin.com/VhXVs7qF and http://pastebin.com/p3EFtZdf 19:47 are two actual applications 19:47 this is *not* different dims case 19:48 just like in multiclass it is just sparse 19:48 blackburn1: noup http://pastebin.com/p3EFtZdf 19:48 check this one 19:48 score += sm->w[cur_class*sparm->size_hog+i+1]*hog[i]; 19:48 w = num_class*sparm->size_hog 19:49 it is just multiclass hog 19:49 i mean the dimension of w 19:49 otherwise the indexing would not work 19:49 dimension is W is always equal to dimension of Psi, but Psi can be sparse 19:49 that is the case 19:49 and multiclass is the case 19:49 ano 19:49 no 19:50 ok 19:50 u r right 19:50 dimension of Psi here is not equal to dimension of hog 19:50 it is dim of hog * n_classes 19:50 but everything out given class is 0 19:51 so it is not computed at all 19:51 my bad 19:51 sorry 19:51 actually if you have that code HOGSO Model is straightforward 19:53 mmm now i'm in trouble... 19:55 i need to calculate this for all examples: argmax_{(ybar,hbar)} [ + loss(y,ybar,hbar)]. 19:55 and i want to keep the COFFIN fw 19:56 line 25 is a dense dot 19:56 wiking: it is argmax alreadyy 19:58 which? 19:58 i mean now i'm writing a generic implementation i'm not talking about those pastebins 19:58 it's just that currnetly i have 1 method for psi 19:58 is that CDotFeatures* get_psi....() 19:58 i guess i could generate all the different CDotFeatures and generate a matrix of the w,psi(x,ybar,hbar)> + loss(y,ybar,hbar) values 20:01 and do there a max value search for each row/column (depending on who do i store [w,psi(x,ybar,hbar)> + loss(y,ybar,hbar)]) and then sum it up 20:02 :) 20:02 it's just a lot of memory waste... 20:02 blackburn1: input? 20:03 yeah sounds inefficient 20:04 but i dont have something like SGVector get_psi(x,y,h) 20:04 as sparse vector wouldn't work then 20:05 why don't you have it? 20:05 because i couldn't have sparse vector 20:05 this is why i threw out the first place the get_psi(x,y,h) and have only CDotFeatures* get_psi_vectors() 20:06 I see 20:11 build #447 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/447  blamelist: Sergey Lisitsyn 20:12 -!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun 20:20 wiking: I just read the the thread with the problem with risk function args :D 20:22 :D 20:28 -!- blackburn [~blackburn@83.234.54.163] has joined #shogun 20:42 -!- blackburn1 [~blackburn@188.168.2.65] has quit [Ping timeout: 248 seconds] 20:43 -!- sr___ [u5548@gateway/web/irccloud.com/x-fgcntzhdfnowhati] has quit [Read error: Connection reset by peer] 21:03 -!- sr___ [u5548@gateway/web/irccloud.com/x-ewmhdcvbknaunxsl] has joined #shogun 21:03 -!- blackburn [~blackburn@83.234.54.163] has quit [Quit: Leaving.] 21:15 -!- naywhaya1e is now known as naywhayare 21:38 -!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Read error: Connection reset by peer] 21:56 -!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun 21:58 any clue why that crash is happening in Arthur's question? 22:03 mailing list 22:03 n4nd0: swig version could be a possible problem 22:11 i mean i have 2.0.8 22:11 1.3.29 seems a bit old 22:11 wiking: that's true 22:12 wiking: should I answer him or do you want to do it? 22:13 mmm i think that should be tested though 22:13 and if that's the case then we should find the minimum swig version shogun requires to be able to compile the modular interfaces 22:14 the problem is that i have 0 knowledge about swig 22:15 i think gsomix would be the most competent to say anything about this 22:15 wiking: I remember once my compilation failed and sonney2k told me to update swig 22:16 it worked after that 22:16 I will just tell him in case it can solve his problem 22:16 n4nd0, wiking yes we need newer swig - but IIRC configure should fail if swig < 2.0.4 22:17 yeah the problem is i think that he would need to manually install a new version of swig as he now uses the standard distrib supplied swig 22:17 ooh yeah 22:19 I could do it via apt-get easily 22:19 _swig_version=${SWIG} -version 2>/dev/null | grep Version | cut -f 3 -d ' ' 22:19 if assert_version swig$_swig $_swig_version 2.0.4; then 22:19 echores "$_swig_version" 22:19 _swig=yes 22:19 n4nd0: it's RH :) 22:19 wiking: RH? 22:19 redhat 22:20 according to his email he uses redhat => no apt-get ;P 22:20 lol his configure log is 22:21 ============ Checking for SWIG ============ 22:21 Result is: 1.3.29 22:21 ########################################## 22:21 and the assert_version doesn't fail 22:21 -!- blackburn [~blackburn@188.168.13.61] has joined #shogun 22:21 so there's something wrong with that if line 22:22 discussing swig issue? I have to note he is using 1.1.0 22:23 I do not remember if it requires >2.0 but probably it does 22:23 blackburn: just joined 1 minute later 22:23 blackburn: yeah accoding to ./configure script it should actually fail because it's not 2.0.4 or later 22:24 but for some reason that if branch keeps being true 22:24 he seems to be a hardcore guy 22:24 using RHEL 22:24 n4nd0: did you graduate from kindness school? :) 22:26 blackburn: haha why so? 22:26 doh 22:27 it is free to be kind, not expensive :D 22:27 :D 22:27 if i set _swig_version by hand the configure script fails 22:27 aaah 22:28 i think he is using a rather old shogun version 22:28 where this check is not yet present in ./configure script 22:28 as Result is: gcc 4.1.2, ok 22:28 and afaik that should fail as well with < 4.3 22:28 i wonder why they want to use shogun @ ernst & young ;P 22:36 hey really 22:37 what the heck 22:37 maybe for tax auditing :DDD 22:37 why anybody in ernst & young would need shogun 22:37 i guess some classification stuff on tax auditing data ;P 22:37 lol 22:37 would be great to hear more what's his idea for application 22:38 hey we must get more E&Y users 22:38 I bet they have money 22:38 I don't mind to become rich :D 22:38 hahahahahha 22:38 yeah you should do a shogun spin-off ;P 22:39 and make money of shogun support :))))) 22:39 n4nd0: did you notice 22:39 i mean this guy should just fucking update his linux distrib 22:39 his name is sean 22:39 but email is arthur edge 22:39 ah lol 22:40 wtf? 22:40 maybe he is disguising himself ;P 22:40 haha yeah I noticed 22:40 how old is RHEL 5.8 actually? 22:40 I like footer part 22:40 I surprised myself when I thought, I am going to change the signature just in case it is not the same as the mail 22:40 wiking: 2010 probably 22:42 to check* not to change 22:43 duuude 22:45 2012!!!! 22:45 Any U.S. tax advice contained in the body of this e-mail was not intended or written to be used, and cannot be used, by the recipient for the purpose of avoiding penalties that may be imposed under the Internal Revenue Code or applicable state or local tax law provisions. 22:45 WTF?!?!?!?!? 22:46 wiking: really? 22:46 latest is 6.3 IIRC 22:47 blackburn: https://access.redhat.com/knowledge/docs/en-US/Red_Hat_Enterprise_Linux/5/html/5.8_Release_Notes/index.html 22:47 I had to work with 5.3 yesterday on dev server 22:48 blackburn: http://www.h-online.com/open/features/What-s-new-in-Red-Hat-Enterprise-Linux-5-8-1439890.html 22:48 and it's a 2012 february news 22:48 wtf 22:48 redhat ships with fucking gcc 4.1.x in 2012? 22:48 it is ENTERPRISE redhat 22:49 i.e. 22:49 slow 22:49 very slow 22:49 very very slow 22:49 so it must be filled with acient stuff? :) 22:49 exactly 22:49 very very slow developing 22:49 i guess they are following: 'the older the better' 22:49 ;) 22:49 they would use win95 22:49 if m\$ did support it still 22:49 ah lol 22:49 there's gcc44 package 22:49 so he should be ok 22:49 wiking: I know a company which forces my company to use ClearCase 22:49 heard about that? 22:49 hahahah YES 22:49 that's like fucking 2000 22:49 used? 22:49 :D 22:49 yes unfortunately 22:49 that's like 22:49 that's like nothing on the earth 22:49 it is pure love 22:49 yeah it's a pure beauty 22:49 :)))hahahhah 22:49 it is when you lose your mind of happiness 22:50 :D 22:50 but serious, did they really think that dance is a good  way to develop 22:50 you have to create an integration stream to view sources 22:50 hahahah 22:50 NOOOOOO 22:50 then you add activity there you specify what do you want 22:50 i dont want to hear about that shit! 22:50 i had physical pain 22:50 and then you need to create a development stream 22:50 when i had to use it 22:50 and join project with it 22:50 yeah 22:50 then checkout or hijack the file where you want to remove 2 lines 22:50 commit (or how is it called out there) 22:50 i mean seriously why would u not use git+github in 2012? 22:50 and write down 40 pages description of the commit 22:50 we are using svn 22:50 blackburn: and i hope u use git-svn ;) 22:50 but one project uses clearcase still 22:50 blackburn: just to make them cry :)))) 22:50 that is restricted and I didn't say that but it is TELUS 22:50 :D 22:50 wiking: you are the only one but me developer of shogun 22:50 who have a sexual contact with CC 22:51 had* 22:51 I have still though 22:51 :DDD 22:51 fucking clearcase 22:51 i know a guy 22:51 !!! 22:51 he just had to start to use it 22:51 :DDDD 22:51 in 2012 22:51 yep yep 22:51 sounds like torquemada is back 22:51 2 weeks ago or something 22:51 and he is full of will to punish somebody 22:51 yep 22:51 he seemed desparate on facebook 22:51 :P 22:51 that's quite normal to get a depression after using CC for a while 22:52 I am pretty sure there was a case of commiting suicide after trying to get that thing to get your changes 22:54 hehehe 22:54 i think a git-cc project would never be possible 22:54 I am not using git-svn btw 22:54 LOOOOL 22:54 i was proved WRONG 22:54 *kabooom* 22:54 https://github.com/charleso/git-cc 22:54 :DDDD 22:54 I actually feel comfortable with svn as well 22:55 "Warning 22:55 ======= 22:55 I wrote this purely for fun and to see if I could stop use Clearcase at work 22:55 once and for all." 22:55 for the thing we develop it is ok 22:55 :DDDDDDDDDDDDDD 22:55 blackburn: yeah i just don't like that u actually cannot commit a code in svn w/o having internet connection 22:55 "Actually what I would love to see 22:56 more is for Clearcase to die" 22:56 wiking: that is ok when you have pretty reliable internet connection at your job 22:56 lol this guy did feel the pain 22:56 I share his pain 22:56 okay I am going to sleep now probably 22:58 gnite 22:58 I am going to be at job 10 (11) o'clock :D 22:59 or sth like 10* 22:59 blackburn: good night 23:00 -!- blackburn1 [~blackburn@188.168.5.89] has joined #shogun 23:00 oops my cable didn't like that 23:00 wiking: I think git-cc is too dangerous to even try 23:00 CC looks like a space ship command panel 23:01 :))) 23:01 who knows may be one button could destroy the whole world 23:01 yeah i think it opens a new dimension if u try it :))) 23:01 one may end with nuclear war or some chemical disaster 23:01 if saddam would have used gitcc 23:02 haha you guys scare me sometimes :D 23:02 maybe they would have found some weapons of mass destructions :D 23:02 yeah CC with development stream 23:02 with all files hijacked 23:02 n4nd0: you should try CC 23:03 to achieve zen 23:03 in hofstadter book a theory that zen's stuff is just for destroy of all means 23:03 -!- blackburn [~blackburn@188.168.13.61] has quit [Ping timeout: 265 seconds] 23:04 CC is just the same, it destroys development as we know it 23:04 cc is a version control system right? 23:04 yes 23:04 n4nd0: nonono 23:04 blackburn1: dont confuse him 23:04 spaceship missile control system 23:04 I was confused at the beginning with the command cc 23:04 with version control support 23:04 n4nd0: it's like a nuclear reactor man 23:04 hahaha 23:05 have you seen star trek? 23:05 no 23:06 may be gayniggers from the outer space? 23:06 they had planet destruction stuff 23:07 it comes out of box with clear case 23:07 I don't know what the heck are you talking about :) 23:07 :D 23:08 n4nd0: never ever touch clearcase man 23:08 not even in your dreams 23:08 all right, I promise you that :) 23:08 I had heard about subversion and cvs but not about clearcase 23:08 when was that used? 23:08 n4nd0: in 1400-1450 23:09 when the inquisition was like fashion 23:09 hahahahha 23:09 blackburn1: columbus tried to teach CC for the indians 23:09 and they told him: i rather fucking die 23:09 :DD 23:09 ok i know i might be a bit offensive with this comment, wasn't inteded 23:10 then it was probably invented by Spanish people 23:10 n4nd0: it is not true jeanne d'arc was burnt 23:10 Spain was quite an important country back then 23:10 she had to commit to UNOPENABLE DEVELOPMENT STREAM 23:10 is that like a code in that system? 23:11 n4nd0: i think they thought that using CC would be the solution... that's why your ship sank against nelson admiral 23:11 yeah, they just forgot to complete VOBs on integration views 23:12 that usually leads to ship sink or sth like that 23:12 chernobyl is the similar case 23:12 :DDDD 23:12 somebody joined wrong project 23:12 haha 23:12 blackburn1: i bet gorbachov thought that using CC would be the solution as well 23:12 I think we should move shogun to CC 23:12 haha lol 23:12 fuck off with git 23:13 :D 23:13 CC sounds exciting 23:13 it is 23:13 bye guys 23:13 -!- alexlovesdata [~binder@goldenezahl.ml.tu-berlin.de] has left #shogun [] 23:13 CC is a masterpiece 23:13 already leaving? 23:13 ahh 23:13 we scared alex! 23:13 i must check images.google for CC!!! 23:13 okay now finally leaving you too 23:14 http://www.cmcrossroads.com/forum/31-clearcase/100442-clearcase-ucm-vs-git 23:14 CC vs git 23:14 have a nice night I will push you tomorrow to develop more code with clearcase 23:14 oh 23:15 lock before edit is a pros 23:15 is in pros* 23:15 that's really nice 23:15 wiking: fits for your issue about breaking code :D 23:15 First, I would answer that comparing git to UCM is unfair to ClearCase, and not only from the point of view of performance! 23:16 nobody would break your code because you can LOCK IT 23:16 :D ohmfg 23:16 :DDD 23:16 n4nd0: they all sound like dinosaurs 23:16 I've seen one BIG COMPANY WITH MANAGERS claimed that open source solutions sucks 23:17 recently 23:17 in 2012 23:17 blackburn1: yes 23:17 gosh 23:17 blackburn1: that's pretty much normal :D 23:17 haha 23:17 okay okay 23:17 I found an answer you would like more in another forum 23:17 see you tomorrow 23:17 :) 23:17 Now, there is some potential in ClearCase that might make it compare favourably to Git. 23:17 duuuude 23:17 " If not, I'll probably need to have some internal discussion before proceeding. 23:17 But it is not commonly known or used, and would require some support at last from IBM. 23:17 these guys are trying to use 23:18 I love the last part 23:18 shogun @ ey 23:18 :DDD 23:18 n4nd0: yeah nuclear missile button is absent in git 23:18 :D 23:18 https://www.ibm.com/developerworks/forums/thread.jspa?messageID=14787484� 23:18 blackburn1: :DDDDDDDDDDDDDDDDDDDDD 23:19 ROTFL 23:19 :DDDD 23:19 okay see you :) 23:19 -!- blackburn1 [~blackburn@188.168.5.89] has quit [Quit: Leaving.] 23:19 ok 23:20 let's support EY to use shoung 23:20 :D 23:20 erm shogun ;P 23:20 sure 23:21 wiking: do you know it gcc 4.1 will go well with shogun v.1.1.0? 23:22 yeah 23:23 i'm just checking 23:23 that why actually we had that bug 23:23 i must dig out the stupid commit log 23:23 ok yeah htis one 23:27 mmmm http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23287 23:27 n4nd0: can u open this one? 23:27 yes 23:28 it took a bit long though 23:28 yeah here too 23:28 mmm what was the code that actually caused this one 23:28 ah git grep will tell me 23:28 oh didn't know there is a grep in git 23:28 yeah 23:29 it's fucking great 23:29 ;) 23:29 sound really useful yes 23:29 thank you! 23:29 awesomefast 23:29 mmm do we have tags? 23:29 ah yeah cool we do 23:29 what are you executing exactly? 23:30 trying to find out 23:30 whether that bug actually present in 1.1.0 version 23:30 what command are you using? just to learn more about it 23:30 well now i've done this 23:30 git grep "\.~" 23:31 that'll give me back where the destructor was called explicitly in the code 23:31 and now i wanna do the same 23:31 on the 1.1.0 tag 23:31 but i cannot find tag for that 23:31 aham I see 23:31 git tag 23:31 did you go back to that commit? 23:31 n4nd0: yeah tag is for that 23:32 what was the gcc bug exactly? 23:32 n4nd0: that you cannot call explicitly a destructor 23:32 on a templated class 23:32 all right 23:32 i.e. something.~SGVector() 23:32 will not compile 23:32 ok 23:32 the problem is that there's no tag for 1.1.0 release 23:33 too bad 23:33 i'll check 1.0.0 23:33 maybe with the release date we can guess the last commit done by that time? 23:33 ok in 1.0.0 there was no such thing 23:34 mmm i cannot see it from the git log 23:35 which one supposed to be the 1.1.0 release 23:35 ok 23:36 how the fuck he downloaded 1.1.0 ? 23:36 ah ok 23:36 i was checking the wrong thingy 23:36 mmm 23:36 i really dont feel for downloading the tar.gz and do a grep -r on it 23:37 ;) 23:37 i'll just tell him to try to compile it with 4.1 23:37 and it might fail 23:37 if it does 23:37 then it's because we execute that bug :D 23:37 but 23:37 wasn't the bug fixed in gcc 4? 23:37 n4nd0: after 4.3 23:38 ahh al right 23:38 I thought it was after 3.x 23:38 noup 23:38 4.3 and later 23:38 n4nd0: check the comments on the bug page 23:38 ok this guy might be lucky 23:39 i've just checked it :D 23:39 let's see 23:40 ok 23:42 support email sent 23:42 good job, nice mail! 23:43 hahah another one 23:44 osx 23:44 it should fucking work 23:44 ok shit i forgot to cc the mailinglist 23:45 ;( 23:45 wonder what acient osx is he using 23:46 time to sleep for me 23:47 good night! 23:47 -!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] 23:47 --- Log closed Thu Aug 23 00:00:17 2012