Open in new window / Try shogun cloud
--- Log opened Wed May 30 00:00:41 2012
n4nd0good night people00:43
-!- n4nd0 [] has quit [Quit: leaving]00:45
-!- av3ngr [av3ngr@nat/redhat/x-xnbshjkyvrpdknhi] has joined #shogun02:11
-!- ckwidmer [] has quit [Remote host closed the connection]03:22
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]03:28
-!- uricamic [~uricamic@2001:718:2:1634:a0dd:d302:38df:5709] has joined #shogun08:50
-!- heiko [] has joined #shogun09:51
-!- abn_ [av3ngr@nat/redhat/x-tijaowgwtfpejqol] has joined #shogun09:51
-!- blackburn [~blackburn@] has joined #shogun09:52
-!- av3ngr [av3ngr@nat/redhat/x-xnbshjkyvrpdknhi] has quit [Ping timeout: 245 seconds]09:55
sonne|workheiko: whats up?10:00
heikosonne|work, hi10:00
sonne|workhey ...10:00
heikocurrently: NIPS, exam on friday, (so far: exam yesterday)10:01
sonne|workhope you survive this10:01
heikobut also reading stuff to implement eigenspectrum based threshold estimate for quadratic MMD10:01
heikoyeah me too :)10:01
sonne|workyou did ping me yesterday?10:01
heikoyes, I was just asking whether shogun had eigenvalue methods, but then I found them myself10:01
heikoI added this little convienience methods using SGVectors though10:02
heikowhats up in Berlin?10:02
sonne|workheiko: yeah I wanted to move more functions from CMath -> SGVector etc10:02
sonne|workso you patch is very welcome10:02
heikoso all of them should be converted?10:02
sonne|workheiko: in two days I will be on paternal leave10:03
heikoI would love to see the display methods10:03
sonne|workso just finishing up stuff10:03
sonne|workheiko: yes exactly10:03
heikopaternal leave?10:03
sonne|workyou know play time with kids ;-)10:03
heikosonne|work, ok then Ill convert all methods I touch10:03
heikoah ok10:03
heikodont forget to do that over all the coding :D10:03
sonne|workheiko: no worries will be on holidays too so no time for coding next week10:04
heikosonne|work, nice!10:04
heikosonne|work, I really like these director classes, havent tried them yet though, but its awesome10:05
sonne|workyeah I think so too10:05
sonne|workonly problem is speed penalty10:05
sonne|workso factor 10 slower...10:05
sonne|workbut hey could be worse10:05
heikoyes, but I mean10:06
heikosometimes people that dont want to get involved into internals just want a new kernel10:07
heikoor a new other thing10:07
heikoand now thats possible10:07
blackburnheiko: do you need full spectrum?10:08
sonne|workyeah. but we have to enable that one by one10:09
heikoblackburn, no, eigenvalues are fine so far, still experimenting a bit10:09
sonne|workit will only be possible case by case10:09
blackburnheiko: all eigenvalues?10:09
heikosonne|work, but thats fine10:09
sonne|workwe need to figure out for which cases we need this stuff...10:09
heikoblackburn, yes, so far, although I will sort them and only use most numerically stable ones10:10
blackburnfor max/min you can use arpack I mean10:10
heikoah ok10:10
heikocurrently using the lapack based compute_eigenvalues thingi10:10
blackburna lot of options10:10
blackburnheiko: dsyev?10:11
heikoyes that one10:11
heikoisnt arpack more for sparse matrices?10:11
heikowill be back in 5 mins ...10:12
blackburnheiko: no - reverse interface - it is like COFFIN for eigenstuff10:12
blackburnit depends on structure of matrix10:12
blackburnbasically it is pretty related to krylov subspace method10:13
blackburnwhich require span of [ Ax A^2 x .... ]10:13
blackburnso for example like in power method for largest eigenvalue it would require only few matrix vector products10:14
sonne|workblackburn: btw I was thinking about getting DirectorDOTFeatures to work10:14
blackburnwhat do you have in mind10:14
sonne|workand I think one way to do it is to define all these add/dense_dot etc operations again but based on SGVector10:14
sonne|workso then the low-level functions would call the sgvector ones10:15
sonne|workand these are the ones that via typemaps etc are nicely exposed to python and other langs10:16
blackburnyes obvious10:16
blackburnhmm while you are available10:17
blackburnwhat do you think about openmping it gradually?10:17
blackburnI attempted to openmp everything but it needs to be rebased now totally10:17
sonne|workblackburn: regarding openmp - maybe you can openmp the most difficult part, e.g. dotfeatures first. if this works then I think it will work for all stuff10:19
sonne|work(works == same speed as w/ pthreads)10:20
blackburnyes but shall I commit it then?10:21
blackburnreally hard to track one more branch10:22
heikoblackburn, I will just have a look how it works and compare against my matlab implementation and then get back to you10:23
heikohave to study kernel covariance operator now :)10:25
-!- abn_ [av3ngr@nat/redhat/x-tijaowgwtfpejqol] has quit [Ping timeout: 248 seconds]10:40
-!- blackburn [~blackburn@] has quit [Ping timeout: 244 seconds]11:38
-!- heiko1 [] has joined #shogun13:06
-!- heiko [] has quit [Ping timeout: 260 seconds]13:07
-!- gsomix [~gsomix@] has joined #shogun14:18
-!- alex_thedatawhis [82955843@gateway/web/freenode/ip.] has joined #shogun14:52
-!- heiko1 [] has quit [Ping timeout: 245 seconds]14:58
-!- alex_thedatawhis [82955843@gateway/web/freenode/ip.] has quit [Quit: Page closed]15:18
gsomixsonne|work, hey. I'm sorry, I will get back to work at evening. Exam... tomorrow. Preparing is needed.15:19
-!- heiko [] has joined #shogun15:46
-!- heiko [] has quit [Ping timeout: 245 seconds]16:16
-!- n4nd0 [] has joined #shogun16:38
n4nd0summer, yaay!!16:39
-!- alexdatawhispere [82955843@gateway/web/freenode/ip.] has joined #shogun17:04
-!- blackburn [~blackburn@] has joined #shogun17:28
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has joined #shogun17:29
-!- uricamic [~uricamic@2001:718:2:1634:a0dd:d302:38df:5709] has quit [Quit: Leaving.]17:29
-!- heiko [] has joined #shogun17:36
-!- zxtx [] has joined #shogun17:52
-!- heiko [] has quit [Ping timeout: 245 seconds]18:21
-!- heiko [] has joined #shogun18:36
-!- heiko [] has quit [Ping timeout: 245 seconds]18:40
blackburnn4nd0: did you finish your exams?18:55
n4nd0blackburn: yeah!19:11
blackburnI just came with cool game - everyone say some math word and I have to put some sentence with it to my thesis :D19:39
gsomixblackburn, topology19:39
blackburnn4nd0: your turn19:41
n4nd0blackburn: let me think...19:41
gsomixblackburn, ok. then I think I need to consult with Verbitsky, hehe19:42
n4nd0blackburn: curl19:43
blackburnn4nd0: hah ok!19:44
blackburnok currently diffeomorphism, curl, Darboux sum, topology (gsomix want to change?)19:48
blackburnpuffin444: surprise me19:48
blackburnsonney2k: and you :)19:48
puffin444Noetherian Rings19:48
gsomixblackburn, nope.19:49
blackburnoh sh19:49
puffin444Did I break your thesis?19:49
blackburnyes probably it is not about svms anymore19:49
blackburnhowever one guy said Klein group already19:51
blackburnuh I asked him to change :D19:51
blackburnaffinity now19:51
blackburnpuffin444: could you please change? :D19:51
blackburnsounds like a bit of hardcore to put rings to svms :D19:52
puffin444Uhh... Dense Set19:52
blackburnhah ok!19:52
blackburnthanks :)19:53
blackburnin summary, I need to add affinity, curl, dense set, topology, diffeomorphism, darboux sum19:55
blackburnokk sounds crazy :D19:57
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]20:02
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has joined #shogun20:13
puffin444Anybody here an expert in blas?21:53
blackburnwell not expert but have some experience21:54
blackburnwhat is the issue?21:54
puffin444So I am calling cblas_dposv, Solves Ax=B where A is a positive symmetric matrix21:54
-!- cwidmer [] has joined #shogun21:55
puffin444It is somehow possible that the contents of A get modified during the function?21:55
blackburnyes it modifies A21:55
blackburnstores cholesky upper triangular L21:55
puffin444Should I be calling clapack_dposv or dposv_?21:57
blackburnpuffin444: in mathematics/lapack.h we provide impl of clapack_dposv that calls usually fortran extern dposv_22:14
blackburnif atlas is available no need to define clapack_dposv so it will be used from atlas with the same interface22:14
puffin444I think I have been just using the wrong function all along22:15
puffin444There got it now. I should have been using clapack_dpotrs as I already had the factorization22:18
blackburnah yes22:18
blackburndpotrf and dpotrs22:18
blackburndposv is dpotrf and dpotrs basically22:18
blackburnhowever multiple calls of dposv make no sense22:19
blackburnit would factor factors22:19
puffin444Yay! I can calculate the gradients now!22:27
blackburnheh cool22:28
puffin444Ok. I think its time to take a break. See you later.22:30
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has quit [Quit: Page closed]22:37
-!- alexdatawhispere [82955843@gateway/web/freenode/ip.] has quit [Quit: Page closed]22:38
n4nd0blackburn: hey, I got one question22:44
n4nd0blackburn: why is it that we cannot have modular interface for matlab?22:44
blackburneasy question - swig does not support it :D22:44
n4nd0aham I see22:44
n4nd0my friend is using shogun for his thesis22:45
n4nd0he started with octave and it was all right22:45
n4nd0but he needed some OOP and it looks like octave is somewhat limited there22:45
blackburnpython is the key to everything22:45
n4nd0so he's now in matlab but found that there's no modular interface there22:45
n4nd0I think he will manage with static in any case22:46
-!- n4nd0 [] has quit [Quit: leaving]22:59
-!- heiko [] has joined #shogun23:05
gsomixgood night guys23:25
gsomixexam is tomorrow... networks23:26
blackburngood luck23:27
heikogood luck23:27
blackburnheiko: are you busy right now?23:29
heikono, resting :)23:30
heikoblackburn, but in generell yes23:30
heikoNIPS till friday23:30
heikoand I got an exam on friday also23:30
blackburnI have one ML question23:30
blackburnwhat is it is about?23:30
heikoexam or nips?23:30
heikoI mean, I if I can help you, I will enjoy doing it :) whats the question?23:31
blackburnheiko: do you know what is L1/Lq norm?23:32
blackburnboth exam and nips23:32
heikoblackburn, exam is about kernel theory: RKHS theory, some kernel algorithms, distributions in feature space, MMD, a bit of convex optimisation23:32
blackburnheh cool23:32
heikonips working title is kernel selection for large scale two sample tests23:32
heikoyes the course was pretty cool23:32
heikoI know Lq norm23:33
blackburnI have troubles deriving dual of some special task23:33
heikousing Lagrangian?23:33
blackburnL1/Lq is basically if you divide your vector to some cells and compute Lq of it23:33
blackburnsummarizing it with L123:33
blackburnL1/L2 of [1 2 3 4] = (1+4) + (9+16)23:34
blackburnif groups are first twos and second23:34
heikooh that is used in the group lasso right?23:34
blackburnheh exactly23:35
blackburnthe problem is I do not understand how ||w||^2 would looks like in terms of alphas :)23:35
blackburnwould look*23:35
heikodo you have like a formal description of the problem?23:36
blackburnno but let me try to latex it a little23:36
heikoyou can use gmailtex23:36
heiko(thats faster /easier than whole latex procedure)23:37
blackburnheiko: is that idea clear for you?23:37
blackburnlet G some indices space :)23:38
heikommh that graphic isnt properly displayed here23:38
blackburnwait I start to think my question is stupid23:39
heikoI cannot read it :(23:39
heikobackground is black instead of white23:40
heikoand font is also black23:40
blackburnheiko: ok nevermind I have to think about it a little more23:40
heikoI am only seeing the antialiased pixels23:40
blackburnthe thing I was confused is how to compute derivate of dual lagrangian by w23:41
blackburnif norm is complicated23:41
heikoah ok23:41
heikoderivatives of norms23:41
heikoI always do it component wise if this happens23:42
heikoand the re-assemble the result later once I understand whats going on23:42
blackburnin case of 2 norm it is pretty basic23:44
blackburnjust like scalars23:44
heikoare you talking about the SVM dual?23:45
blackburnjust want to get liblinear easy patched23:47
blackburnit is based on gradient so I think if I had it I would patch it23:47
heikommh, I dont really get it, but nevermind, have you solved it though?23:48
blackburnsolved what?23:48
blackburnI think I have to formulate questions more precisely :D23:48
blackburnjust confused you23:48
heikoyes :)23:49
heikowell, I gotta go to  bed anyway23:49
heikoWe can discuss tomorrow if you want23:49
blackburnok :)23:50
heikoI will be st23:50
heikoI will be online during the day since I have to study :)23:50
heikogood night23:51
-!- heiko [] has left #shogun []23:51
blackburngood night23:51
--- Log closed Thu May 31 00:00:41 2012