Open in new window / Try shogun cloud
--- Log opened Wed Nov 14 00:00:17 2012
@sonney2kptizoom, excellent work as usual...00:09
blackburnla la00:09
-shogungit:#shogun- [shogun] sonney2k pushed 3 new commits to master: https://github.com/shogun-toolbox/shogun/compare/742b70c55a8c...2f3bb628950200:10
-shogungit:#shogun- shogun/master c3dd767 ptizoom: perl_modular: fixed matrix SG to PDL conversion....00:10
-shogungit:#shogun- shogun/master f566eba ptizoom: tests/perl_modular: increase max_train_time of linear classifiers00:10
-shogungit:#shogun- shogun/master 2f3bb62 Soeren Sonnenburg: Merge pull request #822 from ptizoom/perl_121113...00:10
-!- blackburn [~blackburn@46.20.65.191] has quit [Quit: Leaving.]00:56
-!- ptizoom [~christian@85.210.94.175] has quit [Quit: Ex-Chat]03:13
-!- ptizoom [~christian@85.210.94.175] has joined #shogun03:14
-!- zxtx [~zv@cpe-75-83-151-252.socal.res.rr.com] has joined #shogun07:48
ptizoomblackburn:yet another exemple of backward propagation training...reward from the master, jealousy from the others....08:55
sonne|workwork == reward, more work == more reward10:44
sonne|workwiking: I will ask today.10:44
sonne|workthis situation is not really helpful10:45
wikingsonne|work: yo10:53
wikingsonne|work: yeah but i think they have their 'good reason' for it10:53
wikingso by asking we are not going to go any further10:53
wikingi guess the latest we'll know something about it on the 3rd of december :D10:53
wikingwhen we realise that we are not there or when we are ther ;)10:54
wiking*there10:54
sonne|workhehe10:56
wikingsonne|work: you should make a shogun-foundation ;)10:57
wikingsonne|work: where people can donate :>10:57
sonne|workwiking: well I asked upfront and carol said that they know that the schedule is extremely tight so they want to do the decision w/i a week10:57
wikingthey can even make donation for a given feature ;)10:58
wikingthat'd be cool10:58
wiking:D10:58
sonne|workhaha10:58
sonne|workcarol is on vacations10:58
wikinghehehe10:58
wikingmm isn't thanksgiving is about to happen soon?10:58
sonne|workreturning just before the meeting :)10:59
wikingoh yeah 22nd of nov10:59
wikingso maybe she went on hunting some turkey ;P11:00
wikingsonne|work: here?13:27
sonne|workyes?13:27
wikingsonne|work: any suggestion which machine would be adecvate to solve this dual problem:13:28
wikingi hate pdf13:28
wikingsorry gimme a sec to formulate the problem13:28
wikingsonne|work: http://test.izyba.com/latex?&space;D%28\alpha_*,\beta_*%29=max_{\alpha,%20\beta}%20\sum_i%20\alpha_i%20+%20\sum_j\sum_h%20\beta_{j,h}%20-%20\frac{1}{2}%20||\sum_{\alpha}%20\alpha%20\Phi%28x_i,%20h_i%29%20-%20\sum_j%20\sum_h%20\beta_{j,h}%20\Phi%28x_j,%20h%29||^2%20\\%20\mattext{s.t.%200%20%3C=%20\alpha_i%20%3C=%20C_i,%20\forall%20i;%200%20%3C=%20\beta_{j,h}%20%3C=%20C_2,%20\forall%20j,%20\forall%20\in%20H}13:33
wikingshiiit13:33
wikingsonne|work: this is the problem http://pastebin.com/X2t8bEAd13:34
wikingand just copy-paste it there: http://www.codecogs.com/latex/eqneditor.php13:34
sonne|workhttp://mathbin.net/11369413:36
wikingyeeey13:39
wikingthat :D13:39
sonne|workit looks very SVMish13:42
wikingyeah it's a dual form13:42
sonne|workyeah but you also only have box constraints on alpha / beta13:50
sonne|workyou should make the quadratic term explicit btw13:51
sonne|workyou could certainly use ocas to solve that13:51
wikingocas?13:53
sonne|workwell libqp13:53
wikingaaah13:53
sonne|workocas assumes fixed Cs13:53
wikingmmm13:53
wikingi'm just checking libqp's formulation13:53
sonne|workI mean you could introduce gamma = [alpha beta] and reformulate the problem to use only that13:54
sonne|workthen you have just one variable to optimize over13:54
sonne|workdifferent box constraints for different parts13:54
sonne|workand some linear term13:54
sonne|workand quadratic term13:54
sonne|workhttp://cmp.felk.cvut.cz/~xfrancv/libqp/html/13:55
wikingyeah13:55
sonne|workyou have no bias term so no equality constraint even13:55
sonne|workso the problem is more simple but I don't know if libqp has something internally that works w/o one13:56
wikingbut this is minimizing the problem :S13:57
wikingi mean qp13:58
wiking*libqp13:58
sonne|workthen mimimize -qp :D13:59
wikingmmmm14:00
wikingi hope S_{equ} can be a zeroset14:01
sonne|workno not allowed14:04
sonne|workask vojtech14:04
wiking:S14:05
sonne|workI am pretty sure one of the sub-solvers for solving the problem w/ bias constraint needs a solver w/o one - so that one should be in libqp too14:06
sonne|workbtw I guess you tested this with mosek already?14:06
wikingno not at all14:07
sonne|workwhy not?14:08
wikingi've just seen this paper14:12
wikingand it would be great to have a solver that doesn't depend on mosek :)14:12
sonne|worksure but first build a reference then do the real impl.14:13
wikingmmm i guess mosek has a solver for this14:15
sonne|worksure it is a QP14:19
sonne|workbut you still need to create that gamma14:19
wikingah yeah14:22
wikingthe question still the same14:22
wikingsonne|work: https://gist.github.com/407203814:23
wikingline 19714:23
wikingi want to commit this solver tday14:24
wikingso i need a solution ;)14:24
sonne|workits only in a .cpp so I don't care14:35
wikingsonne|work: ah sure?14:38
wikingok then here goes the PR14:38
-shogungit:#shogun- [shogun] vigsterkr opened pull request #823: Add SO-SVM solver based on Chun-Nam Yu's implementation (master...so) https://github.com/shogun-toolbox/shogun/pull/82314:58
sonne|workwiking: did you ifdef with HAVE_MOSEK ?14:59
sonne|workwiking: and could you do w/o vector in the .h?14:59
wikingi've done ifdef for mosek15:00
sonne|workapart from that ok15:00
wikingsince this solver could use SVMlight15:00
sonne|workcannot see it15:00
wikingit's only within cpp at a given place15:00
wikinghttps://github.com/shogun-toolbox/shogun/pull/823/files#L0R32915:00
wikingah ok i'll remove the std::vector header includ ethere...15:01
wikingbut we need to update svmlight somehow for that15:04
wiking:)15:04
wikingsonne|work: ok i cannot move the include ;)15:07
wikingsonne|work: https://github.com/shogun-toolbox/shogun/pull/823/files#L1R13715:08
wikingas the u can see the resize_cleanup function args contain std::vector15:08
-shogungit:#shogun- [shogun] ptizoom opened pull request #824: perl_modular: fixed string_from_pdl<char> conversion. (master...perl_121114) https://github.com/shogun-toolbox/shogun/pull/82416:19
sonne|workwiking: couldn't you use SGVector for that or dynarray?16:23
-shogungit:#shogun- [shogun] sonney2k pushed 2 new commits to master: https://github.com/shogun-toolbox/shogun/compare/2f3bb6289502...0cdc7eb45a4416:29
-shogungit:#shogun- shogun/master d2765ab ptizoom: perl_modular: fixed string_from_pdl<char> conversion....16:29
-shogungit:#shogun- shogun/master 0cdc7eb Soeren Sonnenburg: Merge pull request #824 from ptizoom/perl_121114...16:29
wikingsonne|work: i'd need to define the template for SGVector<SGSpsarseVector...>16:36
wikingbut i need to send a fix anyways for that PR16:37
wikingi'm just checking whether it compiles now16:37
sonne|workwiking: btw do you understand why std::vector works even with SGVector objects?16:39
sonne|workI mean it is on the stack16:39
sonne|workor is it using new[] ?16:39
sonne|workwiking: anyway you could use DynArray16:40
sonne|workbut you need these sh... inplace constructors16:40
wikingsonne|work: it's on the heap16:40
wikingthere... let's see if travis can compile it now16:43
wikingok at least now it compiles on travis16:48
ptizoomsonne*:red concerns on formatting...I cannot promise a good style, but suggest to indent at the end of development and doc....but will try anyway!17:01
ptizoomsonne*:there are still many big problems in perl....17:02
ptizoomsonne*:sparse stuff is not tested at all! inf and nan values handling.17:03
ptizoomsonne*: also, string_to_pdl does not work as expected...I was looking at that.17:04
-!- blackburn [~blackburn@37.61.179.13] has joined #shogun17:48
blackburnwiking: oppan gangnam style solver?17:48
blackburnsonne|work: yes std vector uses constructors18:26
blackburnso no problem with sgvector18:26
blackburnstd is safe you should believe :)18:27
wikingblackburn: :DDDDDDDDDDDDDDDDD19:09
@sonney2kwiking, I just learned - tonight (timeout of your choice) we will get some notification email22:02
@sonney2kso so after lunch tomorrow CET22:03
@sonney2ks/timeout/timezone22:03
wiking:>22:18
wikingi guess that means we are not in :)22:18
wikingotherwise they would have sent that yeah you are in but we'll announce it officially tonight :)22:19
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun22:37
* wiking having the flow with boost23:10
blackburnwiking: boost is illegal in germany23:12
wikingidontcare23:16
wikingi'm working on my shit23:16
wiking:)23:16
blackburn:D23:16
blackburnwe are all working on our own shit23:16
wikingsomebody figure out wha tto do with that std::vector so that it can be merged23:17
blackburnwiking: I have nothing against it in .cpp23:18
wikingits in .h23:18
wikingas for a func args23:18
blackburnah23:18
blackburnwiking: may be some vodka?23:19
wiking:>23:19
blackburnhard to speed up brain activity w/o23:19
wiking:>23:23
wikingyes yes23:23
wikinghahaha i think the most i like about c++11 is the auto variable :)23:33
blackburnwhy?23:33
wikingno more fucking23:33
wikingstd::map<something long, somethingevenlonger>::const_iterator23:34
blackburntypedef23:34
wikinglolface23:34
blackburnyou should not use such types in your code I think23:34
wikingi'm not creating industry standard code23:34
blackburnI use typedefs for that23:34
wikingjust trying to test something as fast as possible23:34
wikingand usually at that moment23:34
blackburnI'd rather have ParameterMap than std::unordered_map<fdsfsdfdsf,fsdfsdfs>23:34
wikingmy code changes like everyfuckingtime23:35
--- Log closed Thu Nov 15 00:00:17 2012