Open in new window / Try shogun cloud
--- Log opened Sun Apr 08 00:00:19 2012
n4nd0sonney2k: covertree sucks :(00:10
n4nd0super slow00:11
n4nd0I am testing it with the files Sergey has sent00:11
n4nd0I am just trying to classify 10 vectors00:11
@sonney2kso blackburn slowed everything down with covertrees :D00:11
n4nd0the old KNN method takes about 2 second00:11
n4nd0with covertree ... it's not finishing ...00:12
n4nd0more than a minute now00:12
blackburnsonney2k: no, edrt stuff is faster with covertree00:12
blackburnn4nd0: do you precompute distance matrix there?00:12
n4nd0blackburn: any problem with cover trees for high dimensions?00:12
n4nd0blackburn: no00:13
blackburnn4nd0: no, it depends only on distance00:13
blackburnsome intrinsic dimensionality stuff is here00:13
blackburnhowever it is not of interest there I think00:13
n4nd0I don't know00:14
n4nd0but I think that even the covertree impl. we are using is not super good00:14
n4nd0this result is weird00:14
blackburnnot weird00:14
blackburnit computes all distances a few times00:14
blackburnit is time demanding00:14
blackburnhowever how long does it take to construct a tree?00:15
n4nd0I am going to check that now00:15
blackburndo you separate these processes?00:15
n4nd0I think that the problem is there00:15
n4nd0they are separate00:15
n4nd0construction is in train00:16
blackburnso how long train takes?00:16
n4nd0I think that the operation that takes a life time is training00:16
n4nd0it is not ending directly :)00:16
n4nd0we are talking about a huge bunch of vectors though00:16
n4nd0but does cover tree performance takes this into account?00:17
blackburnit should be log N00:17
n4nd0I mean, when they say that it is for KNN classification00:17
n4nd0do they take into account construction?00:17
blackburnno, construction is not considered there00:17
blackburnyou would need to construct it only once00:17
blackburnn4nd0: I can say that knn (k=10) takes a hour or so to compute outputs with naive quick sort00:18
blackburnon that data00:18
blackburnall vectors I mean00:18
n4nd0using 1000 vectors for training00:19
n4nd0it took 13 seconds to construct the tree00:19
blackburnand 2?00:19
n4nd0k = 800:20
n4nd0apply takes more time :(00:20
blackburnhow much?00:20
n4nd0two times more00:21
n4nd0but the number of vectors is really small00:21
blackburnactually this apply *can't* take more time than qsort one00:21
blackburnI do not understand how can that be00:21
blackburnqsort approach computes distances to each vector00:21
blackburnand covertree computes a few00:22
n4nd0it makes no sense with intuition but00:22
n4nd0there must sth there00:22
n4nd0I am going to try with more vectors fro apply00:22
n4nd0500 for apply and 1000 for training00:22
n4nd0I would expect here the cover tree00:23
n4nd0not taking into account training00:23
blackburnno need to take train time into account00:23
n4nd0is a shit00:23
gsomixgood night00:23
n4nd0sorry for the expression :P00:24
n4nd0gsomix: good night!00:24
n4nd0>>>>> Classification with quick sort took 2.9583039283800:24
n4nd0>>>>> Cover tree construction took 13.074337959300:24
n4nd0>>>>> Classification with cover tree took 5.4746079444900:24
blackburnn4nd0: haha00:24
blackburnokay 1000 apply?00:24
n4nd0I can tell you00:24
n4nd06 sec with quick sort00:24
n4nd015 with cover bush00:25
n4nd0just guessing00:25
n4nd0want to make your bet?00:25
blackburnit shouldn't be possible00:25
n4nd0>>>>> Classification with quick sort took 6.0068957805600:25
n4nd0>>>>> Cover tree construction took 13.040800094600:25
n4nd0>>>>> Classification with cover tree took 10.903674125700:25
n4nd0^ 1000 in train and 1000 in apply00:26
blackburnfactor of 200:26
n4nd0where do you want to end up :P?00:26
n4nd0train as well?00:26
blackburnboth ~200:27
blackburnjustinbeliebable lol00:27
n4nd0>>>>> Classification with quick sort took 12.092339038800:27
n4nd0>>>>> Cover tree construction took 13.075815916100:27
n4nd0>>>>> Classification with cover tree took 21.870846033100:27
n4nd01000 train, 2000 apply00:27
n4nd0maybe I am doing sth wrong00:28
-!- gsomix [~gsomix@] has quit [Ping timeout: 246 seconds]00:28
PhilTilletHmm, I can't find any knn c++ example, is there any?00:29
PhilTilletthat's why I can't find one :p00:29
n4nd0I think that if we use covertree + opencl support00:29
blackburnn4nd0: all is ok..00:29
n4nd0maybe we can beat quick sort00:29
blackburnn4nd0: nooooooo00:30
n4nd0I was ironic / joking :)00:30
n4nd0I think it must be an error in my integration in KNN00:31
blackburnI just glanced over it00:32
blackburnall seems to be ok00:33
@sonney2kno one beats my quicksort :)00:33
n4nd0sonney2k: hahaha00:33
@sonney2kit is 100 times faster than qsort() from libc *lol*00:33
blackburnsonney2k: wtf?00:34
blackburnhow can that be00:34
n4nd0sonney2k: serious?00:34
n4nd0sonney2k: I think you should tell them ...00:34
@sonney2knot serious00:34
@sonney2kit is faster but not sure how much00:34
blackburnwhy it is?00:35
@sonney2kit is because they do advances stuff00:35
@sonney2kI don't00:35
n4nd0that are never used?00:35
@sonney2kn4nd0, just benchmark it ... maybe it got fixed by now00:35
n4nd0sonney2k: what is your intuition about the cover tree?00:36
blackburnn4nd0: btw you may use [:,:1000]00:36
blackburninstead of range00:36
blackburnrange is a bad thing00:37
n4nd0you think so?00:37
n4nd0I actually changed :1000 to range00:37
blackburnyes, range computes whole range00:37
n4nd0mmm I don't really know why00:37
@sonney2kI think they do quicksort first then switch to bubble for small bin sizes00:37
n4nd0let's focus on doing a toolbox for sorting00:38
blackburnn4nd0: what if you take say 10 features?00:39
n4nd0blackburn: train or apply?00:39
blackburnis covertree faster then?00:40
n4nd0I guess you say for training00:40
blackburnjust truncate feature matrices00:40
n4nd0ah ok00:40
n4nd0I am trying first with the whole bunch of sample00:41
n4nd0but that may take a lot00:41
blackburnit looks like we have shity covertree :)00:42
n4nd0we are almost there00:44
n4nd0>>>>> Classification with quick sort took 0.35922598838800:45
n4nd0>>>>> Cover tree construction took 0.40572118759200:45
n4nd0>>>>> Classification with cover tree took 0.39050197601300:45
n4nd0get worse with more samples ...00:45
n4nd0>>>>> Classification with quick sort took 3.5772619247400:46
n4nd0>>>>> Cover tree construction took 0.40811395645100:46
n4nd0>>>>> Classification with cover tree took 4.0205910205800:46
@sonney2kn4nd0, try with few training examples and *lots* for testing00:46
n4nd0>>>>> Classification with quick sort took 0.71573090553300:47
n4nd0>>>>> Cover tree construction took 0.0095810890197800:47
n4nd0>>>>> Classification with cover tree took 1.5841479301500:47
n4nd0100 training examples, 100000 test vectors00:47
n4nd0using just 10 features00:47
blackburnwhere did you get so many00:47
@sonney2kok then screw this cover tree - blackburn, gsomix get JL's to work00:48
n4nd0blackburn: I did testdat[:10, :100000]00:48
n4nd0blackburn: none complained :)00:49
n4nd0blackburn: I don't understand why it does not complain00:51
n4nd0blackburn: testdat.shape = (580, 12630)00:51
blackburnn4nd0: it cuts to 10000000:52
blackburnokay LLE00:52
blackburnso just takes all00:52
blackburnwith 4000 vectors00:52
blackburn0.6s to compute neighborhood matrix00:52
blackburnand with 8000 vector00:52
blackburnkind of wrong asymptotic...00:53
blackburnfuck that00:53
blackburnwe really need proper covertree then00:53
blackburnn4nd0: okay postpone this shit for now00:54
n4nd0blackburn: ok00:54
n4nd0blackburn: fuck, I lost time doing the integration :(00:55
blackburnn4nd0: no, interface is similar00:55
n4nd0blackburn: at least :)00:55
blackburnwould be some patch rather than new reintegration00:55
blackburnit works in that covertree_point_t way as well00:55
blackburnn4nd0: gsomix actually was near to finish that00:56
blackburnI hope he didn't delete that stuff00:56
n4nd0blackburn: let's see00:56
blackburnokay time to got to bed00:57
n4nd0blackburn: curiosity, do you remember what the author of this covertree said?00:57
blackburnn4nd0: hmm? that one we have?00:57
blackburnI found that on github00:57
blackburnI didn't contact him00:57
n4nd0time to sleep00:57
blackburnbut probably it is a problem00:57
blackburnsee you00:58
n4nd0good night people!00:58
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has quit [Quit: Page closed]00:58
-!- n4nd0 [] has quit [Quit: leaving]00:58
* sonney2k starts to watch revision live stream00:58
-!- flxb [] has joined #shogun01:39
-!- flxb [] has left #shogun []01:40
PhilTilletgood night everybody!01:47
-!- PhilTillet [] has quit [Quit: Leaving]01:47
shogun-buildbotbuild #460 of java_modular is complete: Success [build successful]  Build details are at
-!- pluskid [~chatzilla@] has joined #shogun02:33
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]02:58
pluskidwhy is sonney2k  always online? :D03:40
pluskidI'm scanning some of your chatting logs last night03:40
pluskidconcerning cover tree, I accidently encounter this implementation the other day:
pluskidThe author also has a blog post on this:
pluskidMaybe you could compare with this implementation to identify where the problem is?03:42
-!- pluskid [~chatzilla@] has quit [Quit: ChatZilla [Firefox 11.0/20120312181643]]04:06
-!- pluskid [~chatzilla@] has joined #shogun04:08
-!- pluskid [~chatzilla@] has quit [Quit: ChatZilla [Firefox 11.0/20120314124128]]04:14
-!- pluskid [~chatzilla@] has joined #shogun04:16
-!- pluskid [~chatzilla@] has quit [Ping timeout: 252 seconds]04:21
-!- pluskid [~chatzilla@] has joined #shogun04:21
@sonney2kpluskid, which timezone are you in?06:19
pluskidsonney2k: I'm in Asia06:20
@sonney2kpluskid, so it is sth like 14 hrs at your place?06:26
pluskid12:16 PM actually06:27
@sonney2kahh ok06:27
pluskidare you here for all 24 hours ? :D06:27
pluskidIt seems you are always online06:27
@sonney2kbtw, the buildbot figured out there are a few warnings with your patch06:27
pluskidI will fix it06:28
@sonney2kpluskid, yeah well znc :)06:28
@sonney2kpluskid, if you can come up with some other small patches within the next weeks it would be helpful06:29
pluskidOK, I see06:29
pluskidI will try to find something to work on06:29
@sonney2kwe don't know #slots yet but we will try to come up with some kind of pre-selection by the end of tomorrow06:30
pluskidif I didn't find interesting small stuffs, I will come and ask here06:30
@sonney2kok gtg06:30
pluskidthat's cool06:30
-!- gsomix [~gsomix@] has joined #shogun08:40
gsomixgood morning :)08:46
-!- pluskid [~chatzilla@] has quit [Remote host closed the connection]08:49
-!- pluskid [~pluskid@] has joined #shogun08:53
-!- n4nd0 [] has joined #shogun09:20
n4nd0pluskid: hey! I just read your mail, thank you for your findings :)09:20
pluskidn4nd0: :)09:22
-!- wiking [] has joined #shogun09:22
-!- wiking [] has quit [Changing host]09:22
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun09:22
pluskidhope the problem could be solved soon09:22
-!- pluskid [~pluskid@] has quit [Quit: Konversation terminated!]09:22
-!- pluskid [~chatzilla@] has joined #shogun09:23
n4nd0pluskid: but it is tru that it might because of the dataset09:23
pluskidyes, it might be09:24
pluskidbut if it is the case09:24
pluskidat least we could figure out cover tree is good for what kind of dataset09:24
pluskidand document that09:24
n4nd0I thought that if we use for our test any of the one referenced here09:24
n4nd0on the last page09:24
n4nd0cover tree should be better then09:25
pluskidyes, that would be a good try09:26
pluskidthey used larger dataset09:27
pluskidas mentioned in the paper09:27
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]09:53
-!- n4nd0 [] has quit [Ping timeout: 260 seconds]10:02
-!- gsomix [~gsomix@] has quit [Quit: ????? ? ?? ??? (xchat 2.4.5 ??? ??????)]10:19
-!- gsomix [~gsomix@] has joined #shogun10:20
-!- wiking [] has joined #shogun10:22
-!- wiking [] has quit [Changing host]10:22
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun10:22
-!- hpsoar [7c108901@gateway/web/freenode/ip.] has joined #shogun10:44
-!- hpsoar [7c108901@gateway/web/freenode/ip.] has quit [Client Quit]10:47
-!- gsomix [~gsomix@] has quit [Quit: ????? ? ?? ??? (xchat 2.4.5 ??? ??????)]11:04
-!- gsomix [~gsomix@] has joined #shogun11:04
-!- blackburn [~qdrgsm@] has joined #shogun12:08
gsomixblackburn, yo12:15
blackburngsomix: hi, I'll merge your stuff in a min12:15
blackburngsomix: as for covertree I have one suggestion12:16
blackburngsomix: did you delete your covertree modifications? I hope not12:16
gsomixblackburn, I did not delete.12:17
blackburnsee KNN sources12:17
blackburnactually all you would need12:18
blackburnis to change block at KNN.cpp:11312:18
blackburnwhere covertree is constructed12:18
gsomixblackburn, do you mean KNN in shogun?12:19
blackburnand 42612:19
blackburnwhere neighbors are enqueued12:20
gsomixblackburn, and..?12:21
blackburngsomix: hmm my suggestion is to use your adopted covertree there :)12:22
pluskidblackburn: hi, what's the status of regression support in shogun?12:22
pluskidit seems incomplete?12:22
blackburnas we measured some stuff and it turns that dncrane's covertree is rather slow12:22
blackburnpluskid: in which means?12:22
pluskidfor example12:23
pluskidwhy train_machine() in LinearRidgeRegression protected?12:23
pluskidand LeastSquareRegression is almost empty ?12:24
blackburnpluskid: in shogun machines train delegates all the stuff to virtual protected train_machine()12:24
blackburnoh I see12:24
pluskidI mean, when a method is protected, it cannot be invoked by others12:25
blackburntrain can12:25
blackburnthat's ok12:25
pluskidoh, train invokes train_machine ?12:25
blackburngot it12:25
blackburnpluskid: that's ok with leastsquares12:25
blackburnit is LinearRidge with tau = 012:26
pluskidgot it :)12:26
-!- gsomix [~gsomix@] has quit [Ping timeout: 246 seconds]12:26
blackburnthat would be hard to approach any large scale with it though12:28
pluskidI'm considering to add LARS, if n4nd0 hasn't started to work on it yet12:31
CIA-64shogun: Evgeniy Andreev master * r126a1de / examples/undocumented/python_modular/ : minor fix -
CIA-64shogun: Evgeniy Andreev master * rc5c7117 / (3 files in 2 dirs): fixed set_triangle_kernel_matrix_from_full in custom kernel -
blackburnno idea12:31
pluskidmlpack seems to got an implementation of LARS12:31
pluskidwith elasticnet12:32
CIA-64shogun: Sergey Lisitsyn master * r79a91ad / src/shogun/evaluation/ClusteringEvaluation.cpp : Merge branch 'normalized-mutual-info' of git:// -
-!- gsomix [~gsomix@] has joined #shogun12:33
-!- PhilTillet [] has joined #shogun12:35
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]13:12
pluskidlapack is horrible...13:19
pluskidwhere can I find a function list, with doc on which parameters is what?13:19
pluskidfor lapack & clbas13:19
-!- wiking [] has joined #shogun13:36
-!- wiking [] has quit [Changing host]13:36
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun13:36
wikinghappy easter y'all, except blackburn, you'll have it later :P13:46
pluskidwhat is easter y' ?13:58
gsomix>> except blackburn | and me. I live in a floor lower than blackburn. :]14:03
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds]14:05
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun14:17
-!- V[i]ctor [] has joined #shogun14:31
-!- gsomix [~gsomix@] has quit [Quit: ????? ? ?? ??? (xchat 2.4.5 ??? ??????)]14:41
CIA-64shogun: Heiko Strathmann master * rb7b8742 / (21 files in 7 dirs):14:49
CIA-64shogun: Merge pull request #428 from karlnapf/master14:49
CIA-64shogun: subset stacks (+24 more commits...) -
-!- PhilTillet [] has quit [Read error: Connection reset by peer]14:54
-!- PhilTillet [] has joined #shogun15:04
-!- blackburn [~qdrgsm@] has quit [Ping timeout: 240 seconds]15:06
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has joined #shogun15:11
-!- V[i]ctor [] has left #shogun []15:13
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]15:25
pluskidblackburn: thanks!!15:26
-!- gsomix [~gsomix@] has joined #shogun15:35
shogun-buildbotbuild #694 of libshogun is complete: Failure [failed compile]  Build details are at  blamelist: heiko.strathmann@gmail.com15:39
-!- harshit_ [~harshit@] has joined #shogun15:53
-!- n4nd0 [] has joined #shogun16:08
n4nd0pluskid: I have not started with LARS yet, I am planning to fix the testsuite so fell free to work on that if you want16:11
pluskidn4nd0: OK, I'm starting on this16:11
pluskidbtw: Lapack and blas is really hard to use ...16:12
n4nd0you think so?16:13
n4nd0I am not an expert but I managed to make it work for the couple of things I have used it around here16:14
blackburnI don't think it is :)16:14
pluskidhaven't got familiar with it yet16:15
pluskiddoing something like matrix multiple16:15
pluskidand solving linear system16:15
pluskidmakes me headache16:15
n4nd0matrix multiple?16:15
pluskidC = A * B this easy16:16
n4nd0ah ok multiplication ;)16:16
pluskidah, yeah16:16
pluskidmaybe I need some time to get familiar with its naming convention16:17
pluskidit could be better if their doc are supplied with examples16:18
n4nd0now that we are talking about it16:18
pluskidfor example16:18
pluskidin a function for multiplication of two matrices16:19
n4nd0blackburn: what is the main difference between lapack and BLAS?16:19
blackburnn4nd0: BLAS is basic ops16:19
n4nd0I think that lapack includes BLAS but I am not sure about it16:19
blackburnlike Ax=b16:19
pluskidLapack built on BLAS16:19
blackburnand Lapack is solving/etc16:19
pluskidthey provide options to indicate that one of the matrix is transposed or not16:19
blackburnBLAS is all about three levels16:19
blackburnvector-vector operations16:19
blackburnlike dot16:19
pluskidand you have to indicate the shape (dimension) of the matrix16:19
blackburnthat is Level 116:20
pluskidbut I don't know whether the shape should be before transpose or after...16:20
blackburnmatrix-vector operations16:20
blackburnLevel 216:20
n4nd0and matrix-matrix level 316:20
shogun-buildbotbuild #695 of libshogun is complete: Success [build successful]  Build details are at
shogun-buildbotbuild #632 of r_static is complete: Failure [failed compile]  Build details are at  blamelist: heiko.strathmann@gmail.com16:21
n4nd0pluskid: I think that options are actually to change the operations like16:21
n4nd0pluskid: if you want to do A*v or A'*v you use the same function but you use an option for that16:21
shogun-buildbotbuild #654 of cmdline_static is complete: Failure [failed compile]  Build details are at  blamelist: heiko.strathmann@gmail.com16:21
pluskidn4nd0: yes, but in the function, you have to pass the shape (m * n) of A16:21
blackburnpluskid: in the doc they say that sizes are of Op(A)16:21
blackburnso Op(A) is A or A'16:22
pluskidif A is transposed, shall I pass m, n or n, m ?16:22
n4nd0pluskid: I'd say you pass those parameters before transposed16:22
pluskidI'll read the doc again16:22
pluskiddidn't find the detailed doc before blackburn gave me that link16:23
n4nd0pluskid: I have been using this one too
pluskidblackburn: you are right, I should use size of Op(A)16:24
pluskidn4nd0: according to the doc, we should use the size *after* transposed :)16:25
shogun-buildbotbuild #633 of octave_static is complete: Failure [failed test_1]  Build details are at  blamelist: heiko.strathmann@gmail.com16:25
pluskidn4nd0: the API of GSL-BLAS seems slightly different from what we are using?16:25
blackburnyeah totally different16:25
pluskidit seems GSL is doing some more wrapping16:26
pluskidwhile the package we are using is straightforward fortran translation16:26
blackburnno we are tied to basic API16:27
blackburnit would stay the same for scalapck16:27
blackburnor cuda lapack16:27
blackburnor any other lapack16:27
blackburnit is standard api while gsl's is not16:27
pluskidOK, I see16:28
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Remote host closed the connection]16:29
shogun-buildbotbuild #620 of python_static is complete: Failure [failed test_1]  Build details are at  blamelist: heiko.strathmann@gmail.com16:29
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun16:29
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun16:32
shogun-buildbotbuild #633 of r_static is complete: Success [build successful]  Build details are at
shogun-buildbotbuild #655 of cmdline_static is complete: Success [build successful]  Build details are at
shogun-buildbotbuild #634 of octave_static is complete: Success [build successful]  Build details are at
shogun-buildbotbuild #621 of python_static is complete: Success [build successful]  Build details are at
shogun-buildbotbuild #485 of octave_modular is complete: Failure [failed test_1]  Build details are at  blamelist: heiko.strathmann@gmail.com17:01
n4nd0benchmarking cover tree with letter subset ...17:03
n4nd0in the paper on cover trees they say cover tree should be about ten times faster here17:04
blackburnn4nd0: btw I checked things17:04
blackburnand in my LLE implementation17:04
blackburncovertree is 3-4 times faster than quicksort17:04
blackburnon a range 4K - 8K 3d vectors17:04
n4nd0something fishy going on here then17:04
n4nd0it sucks17:05
n4nd0>>>>> Classification with quick sort took 51.003859043117:05
n4nd0>>>>> Cover tree construction took 99.723824977917:05
n4nd0>>>>> Classification with cover tree took 106.71281617:05
blackburnn4nd0: actually gsomix is on the way to integrate JL's17:05
n4nd0that is with 10000 vectors for training and classifcation17:05
n4nd0but I don't understand why is it faster with your tests17:06
n4nd0I am afraid I may have done something wrong :S17:06
blackburnno I don't think so17:07
blackburnpluskid measured distance evaluations and it is computed too often17:07
n4nd0it needs to compute distances quite often to apply queries too17:08
shogun-buildbotbuild #472 of csharp_modular is complete: Failure [failed test_1]  Build details are at  blamelist: heiko.strathmann@gmail.com17:08
pluskiddo you see JL's reply in the mailing list?17:17
blackburndid he reply??17:18
blackburncan't see anything17:18
n4nd0no, I don't17:18
pluskidhe said:17:18
pluskidCan you compute the expansion constant spectrum as in the original17:18
pluskidcovertree paper?17:18
blackburndid he answer directly?17:18
blackburnexpansion constant spectrum? wtf?17:18
pluskidhe replied to shogun-list17:18
n4nd0shouldn't it be here
pluskidI don't know17:19
pluskidmaybe the gmane archive has delay17:19
pluskidhe replied 3 hours ago17:19
pluskidif you subscribed to the mailing list17:19
n4nd0didn't get in my inbox either17:19
pluskidyou should receive one in your inbox17:19
n4nd0yeah, I am subscribed17:19
blackburnyeah I have subscribed recently17:19
blackburnyear ago or so17:20
pluskidthat's what he said17:20
pluskidthough I don't understand17:20
pluskidI haven't read the paper17:20
blackburncan you paste 'To'?17:20
pluskidTo: Chiyuan Zhang <>17:21
pluskidCc: shogun-list@shogun-toolbox.org17:21
pluskidshogun-list doesn't accept CC?17:21
blackburnit does17:21
pluskidwho implemented our current cover tree? maybe we can ask him what is the expansion constant spectrum?17:22
n4nd0well, I think the cover tree is taken from somewhere else17:22
pluskidblackburn: you mentioned that gsomix  is merging JL's ?17:22
pluskidso JL has an official implementation now?17:22
pluskidmaybe we could wait to see what about that one17:23
pluskidsince none of us know the internal of how cover tree works17:23
n4nd0I just wanted to check whether the test data was affecting17:23
blackburnJL's one is better anyway17:23
pluskidis JL's implementation publicly available now?17:24
pluskidI think we should at least reply him in the mail17:25
blackburnyes could you please?17:25
blackburnbecause we can't see the email :D17:25
pluskidmaybe the mail arrived at my inbox because he also sent to me17:26
pluskidand the shogun-list is super slow17:26
blackburnyes could be17:26
n4nd0pluskid: in his web-page there is an implementation of isomap that includes cover tree17:26
pluskidI would say one of us is porting JL's implementation now, if we still have problem with his implementation, we will contact him to try to figure out the problem17:26
blackburnpluskid: yes thanks17:27
pluskidn4nd0: cool, he is still updating his Isomap code!17:27
n4nd0pluskid: I think is another's guy isomap17:27
pluskidn4nd0: ??17:28
pluskidJL is one of the authors of Isomap17:28
n4nd0pluskid: look under application17:28
n4nd0another's guy isomap's implementation :)17:28
n4nd0I don't really know if that sense makes sense using normal English17:28
n4nd0but I think you got the idea, or?17:29
pluskidOh, I see17:29
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]17:30
blackburnone great thing about us is we are able to write some meanless things17:30
blackburnwhile english is foreign for us :)17:30
pluskidenglish foreign for me, too17:31
n4nd0if some native speaker was here, he/she would probably get a headache :D17:31
pluskidn4nd0: JL's page provided some dataset for cover tree testing, maybe you could try out your KNN on that17:31
blackburnpluskid: yes we know :)17:32
blackburnwe have no native speakers there17:32
blackburnat all17:32
blackburnthe only active at the ML native speaker is Evan17:32
blackburnwe should consult him about language too as well as git :D17:33
pluskidhis IRC id is ?17:33
blackburnI've never seen him there17:33
n4nd0me neither17:33
blackburnpluskid: is it true there is no first/last name separation in china?17:34
pluskidnot ture17:34
pluskidbut the first/last is reversed17:34
blackburnso Zhang is a?17:35
pluskidactually, I'm always confused17:35
pluskidI'd rather prefer given name vs. family name17:35
pluskidZhang is family name, Chiyuan is given name17:35
blackburnaha! I see17:35
pluskidEnglish people write Chiyuan Zhang17:35
pluskidbut in China, we always use Zhang Chiyuan17:35
blackburnwe tend to reverse that stuff too17:35
blackburnin russian17:36
pluskidI don't know which one is first which is last...17:36
blackburnlast name is family name I think :)17:36
blackburnwe also make use of patronymes there17:36
blackburnthat's why "o" letter is in my email address17:37
blackburnpluskid: my brother works for huawei there17:37
pluskidhuawei is a big company in China17:37
pluskidbiggest local network device company according to my knowledge17:38
blackburnnot only local afaik17:38
blackburneven backbones are constructed with huawei equipment17:38
blackburnactually they have beaten cisco here I think17:39
blackburnbiggest mobile operator, megafon, uses mainly huawei stuff17:40
pluskidhere ? means where?17:40
pluskidI didn't know they have made such achievements17:41
n4nd0they have two very big countries then17:41
pluskidcan't find "patronymes" in my dictionary17:41
blackburnpluskid: hmm we include father's name17:42
n4nd0I think they say patronymic17:42
blackburnprobably my bad17:42
pluskidOh, I see now17:43
blackburnn4nd0: are your middle names related to your family's members?17:43
pluskidso where are you from n4nd0 ?17:43
n4nd0pluskid: Spain, resident in Sweden17:43
n4nd0blackburn: they are, but it is just a choice17:44
n4nd0I mean, some people say like I will call my son with my father's name17:44
blackburnn4nd0: so Jos? is your father's name?17:44
n4nd0but we don't use patronyms or matronyms17:44
n4nd0blackburn: no, Jos? is the masculin for my grandmother's name17:44
n4nd0blackburn: Fernando is my godfather's17:45
pluskidhaha, I heard that in some restaurant at Spain, people can use some kind of Chinese dialect to order dishes :D17:45
blackburnn4nd0: what about third one then?17:45
n4nd0pluskid: probably is they way in which we would speak Chinese :P17:45
n4nd0blackburn: well Iglesias and Garc?a are just surnames, my father's family name and my mother's17:46
pluskidbut that kind of dialect is very different from our official language, actually, I totally can't understand17:46
pluskidcomplicated name system!17:46
n4nd0pluskid: aham, it sounds something more serious then what you are talking about17:46
pluskidIRC IDs would be enough :p17:47
pluskidthere are many many dialects in Southern China17:47
blackburnpluskid: from which part of chinae you are?17:48
pluskidaround Western17:48
pluskidpreviously I thought many here would be Germany :)17:49
pluskidas shogun is originally developed from MPI, am I right?17:49
blackburnpluskid: well yes17:50
blackburnSoeren is from Berlin17:50
pluskidblackburn: if you are in Russia, you must be in some similar timezone as me17:52
blackburnpluskid: wow, why?17:52
blackburnwe have 9 timezones :D17:52
pluskidwe are using a unified timezone in china17:53
pluskideverywhere is Beijin time17:53
blackburnwestern one is kaliningrad - GMT+117:53
pluskidbig difference17:53
blackburnand eastern is some kamchatka GMT+9 probably17:53
blackburnlet me check17:53
shogun-buildbotbuild #486 of octave_modular is complete: Success [build successful]  Build details are at
pluskidnever mind17:54
pluskidso what's your local time now?17:54
pluskidit's 23:54 here17:55
blackburnI'm in GMT+3 timezone17:55
blackburnor +4?17:55
pluskidhmm, quite different17:55
n4nd017:55 here17:55
pluskidso you use a dot to separate time in Russia?17:56
pluskidn4nd0: even earlier17:56
blackburnnot really17:56
blackburnboth are ok I think :)17:57
blackburn: or .17:57
blackburnand kaliningrad is +317:57
blackburnI was wrong17:57
blackburnokay anadyr is +1217:57
blackburn8 timezones still17:57
blackburnwould be bad to have night at say 14:30 and day at 3:30 :)17:58
shogun-buildbotbuild #473 of csharp_modular is complete: Success [build successful]  Build details are at
-!- pluskid [~chatzilla@] has quit [Read error: Connection reset by peer]18:01
-!- wiking [] has joined #shogun18:06
-!- wiking [] has quit [Changing host]18:06
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun18:06
-!- harshit_ [~harshit@] has quit [Remote host closed the connection]18:07
-!- pluskid [~chatzilla@] has joined #shogun18:08
n4nd0good night18:08
-!- pluskid [~chatzilla@] has quit [Client Quit]18:09
-!- gsomix [~gsomix@] has quit [Quit: ????? ? ?? ??? (xchat 2.4.5 ??? ??????)]18:11
-!- gsomix [~gsomix@] has joined #shogun18:11
gsomixhi all18:11
n4nd0gsomix: hey!18:12
blackburnhmm am I here?18:19
-!- blackburn_ [5bde8018@gateway/web/freenode/ip.] has joined #shogun18:21
n4nd0I will see you later guys18:22
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]18:23
wikingblackburn_: r u here? :)18:24
gsomixblackburn_, ???18:24
wikinganyhow i have to test now mkl :>18:24
wikingblackburn_: any suggestions which mkl to use for multi class case?18:25
blackburn_mkl? whoa that's nice :D18:25
blackburn_I do not know anything about mkl :D18:26
wikingah ok18:26
wikingi'll have to try how it works for me18:26
wikingi have 3 different sets of features18:26
wikingand seems like a good idea18:27
wikingto use different kernels for different feature sets18:27
blackburn_yeah may be18:27
wikingsince i've tried it with only concatenating all the features18:27
wikingbut i wonder if it really makes a big difference18:27
wikingif i use mkl18:27
wikingi guess it should :P18:28
blackburn_I think it depends whether they have rather different structure18:28
wikingyeah they doo18:28
wikingone of it is a very sparse feature set... that is coming from text processing18:29
wikingwhile the other is like a histogram feature from images18:29
wikinguntil now what i've done is to concatenate them l2-norm them and then use polykernel18:29
wikingand as we know using intersection or jensen shannon kernel for those histogram features are giving me better results...18:30
blackburn_in my experience it is not very good to use any dot product on histograms18:30
wikingso i'm thinking about now the mkl18:30
wikingbut i guess the only way to see the outcome is to actually try it :>>>18:31
wikingjust quite lazy18:31
blackburn_what about structured output latent model multiple kernel learning svm?18:31
wikingthat'd be amazng18:32
wikingi wonder where's alex18:32
wikingi guess he is taking his easter holidays...18:32
blackburn_wiking: afaik it uses gmnp18:32
wikingblackburn_: yeah i've just checked the code18:33
wikingi mean the headers...18:33
blackburn_while gmnp is very good it should be nice too18:33
wikingwell i guess now it's time for me to do the coding18:34
wikingthe reading in the various features18:34
wikingand then do the setting up of the mkl18:34
wikingand see what i get18:34
wikingok i need to switch to vpn18:34
-!- wiking [~wiking@huwico/staff/wiking] has quit [Remote host closed the connection]18:35
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun18:35
wikingheheh it's funny to see the amount of members now being decreased of the #shogun channel.... almost back to it's normal 8 members state before the gsoc fever :)18:36
blackburn_however it will increase tomorrow I think18:40
wikingwhy? :)18:41
blackburn_well it is sunday today18:41
wikinghahahah your are an optimist :>18:43
wikingblackburn_: remember i've mentioned about my patch for mahout a couple of months ago18:47
wikingthey've managed to merge it now like 2 days ago....18:47
blackburn_oooh fast18:47
wikingtook them like 2-2.5 months18:47
wikingand it was like a 5 liner patch... nothing serious ... :(18:48
blackburn_yeah they are serious guys18:48
* gsomix hates photometrics stuff.18:55
@sonney2kahh nice and quiet :)19:00
gsomixsonney2k, at home? in chat?19:01
@sonney2kat home19:02
@sonney2kkids are asleep and I am going to shogun mode now :)19:03
@sonney2kgsomix, I see that you patch got merged already? so did you have time to think about what to do next?19:07
@sonney2kI mean your preferences?19:07
gsomixsonney2k, I want to work on SG_Vector.19:08
@sonney2kgsomix, ok - so then first thing to do is to convert everything from returning copies of SGVector to references19:10
@sonney2ki.e. functions doing SGVector foo() should be SGVector& foo()19:10
@sonney2ksame for input arguments19:11
@sonney2knext step is then to just copy & paste the ref/unref code from SGObject19:11
@sonney2kHIGHLY INTRUSIVE19:11
@sonney2kblackburn_, that is why we should do this ASAP19:11
blackburn_sonney2k: ah that is kind of painful19:12
blackburn_penetration lol19:12
@sonney2kblackburn_, it is not as bad as last years double* -> SGVector transition though19:12
blackburn_sonney2k: what about typemaps?19:13
@sonney2kgsomix, then the next step would be to remove destroy_vector / etc19:13
blackburn_any changes are required?19:13
@sonney2kblackburn_, yes sure add a &19:13
blackburn_nothing more?19:13
@sonney2kgsomix, and to go through all the code and use the SG_REF / UNREF macro on sgvectors19:14
blackburn_gsomix: what about covertree? :)19:14
@sonney2kthat is sth we probably all have to help19:14
@sonney2kblackburn_, isn't pluskid working on this?19:14
blackburn_sonney2k: no19:14
gsomixblackburn_, in spare time.19:14
blackburn_gsomix: ok19:14
n4nd0sonney2k: he's focusing on LARS I think19:14
n4nd0I can work on cover tree if it is ok19:15
blackburn_why not19:15
n4nd0I don't think I am doing so much with the testsuite19:15
blackburn_gsomix: can you introduce your pretty ready work to n4nd0 ?19:15
gsomixblackburn_, yep. But it's a little ugly. ^_^''19:17
blackburn_gsomix: it doesn't really matter if it is faster19:18
-!- blackburn_ [5bde8018@gateway/web/freenode/ip.] has left #shogun []19:18
n4nd0gsomix: what is the state of it?19:18
-!- blackburn_ [5bde8018@gateway/web/freenode/ip.] has joined #shogun19:18
gsomixn4nd0, many functions and global variables. And some stuff for v_array.19:19
n4nd0gsomix: what is the implementation based on? JL's paper or another implementation?19:20
gsomixn4nd0, JL code.19:20
n4nd0the one here right?
n4nd0first link after the enumerate with 3 points19:21
gsomixn4nd0, yep. templated version.19:21
n4nd0gsomix: ok19:21
gsomixn4nd0, I even think it's better to start anew.19:22
n4nd0gsomix: oh, do you think so?19:22
gsomixit's seems my code contains errors.19:22
blackburn_gsomix: can you remember any pitfalls you have struggled with?19:22
gsomixblackburn_, yep.19:23
gsomixit's related with differences in v_arrays, our ans JC's.19:23
gsomixwait a sec19:24
blackburn_ans sounds a little french ;)19:24
gsomixhuh :)19:25
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has joined #shogun19:26
n4nd0gsomix: cool, that's your implementation so far right?19:26
gsomixn4nd0, yep.19:26
n4nd0gsomix: what did you do more or less? take the methods in JL's code and add template for the point object19:27
n4nd0more changes?19:27
-!- blackburn_ [5bde8018@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]19:28
gsomixn4nd0, and formatting. :)19:29
gsomixn4nd0, there are little changes to our v_array support.19:30
n4nd0gsomix: and what's the problem of this implementation?19:30
gsomixn4nd0, memory leaks.19:31
gsomixin push and alloc19:31
n4nd0gsomix: ok19:31
n4nd0but for the rest, is it functional?19:31
n4nd0I mean, is it complete?19:31
gsomixn4nd0, iirc, yes.19:32
n4nd0gsomix: cool then19:32
@sonney2kwell ask John Langford about the leaks then (with some small example code)19:33
gsomixn4nd0, maybe I'm wrong. Just try to run.19:34
n4nd0gsomix: ok, I will take a look to understand a bit how it works and later run19:34
n4nd0sonney2k: is it v_array better than std::vector too? :)19:35
@sonney2kI am not a big fan of any std:: stuff .... simply because std::vector is not even specified to be something contiguous etc19:36
@sonney2kso *everything* is better :D19:36
gsomixsonney2k, I will start to work tomorrow.19:36
gsomixAt now it's time to optics. :)19:37
@sonney2kgsomix, you mean on shogun or some job :)19:37
n4nd0sonney2k: I am a big fan of STL but you are starting to make me doubt about it :O19:37
gsomixsonney2k, shogun, of course.19:37
@sonney2kgsomix, ok19:37
@sonney2kn4nd0, maybe I am just not knowledgeable enough about it19:38
n4nd0sonney2k: I have never analyzed its efficiency to tell the truth, I think that's the part you critize about it19:40
blackburnactually it is as well efficient afaik19:41
@sonney2kblackburn, where is it specified that vector is contiguous in memory?19:45
blackburnit is not probably however that does not implies it is very uncontiguous :)19:47
n4nd0blackburn: why wouldn't it be contiguous?19:47
blackburnn4nd0: because of resize19:48
blackburnhmm however some says it is19:49
n4nd0blackburn: this v_array class we have in shogun, have you worked with it?19:53
blackburnn4nd0: yeah had used it somewhere.. any problem?19:54
n4nd0blackburn: yeah, for example the method pop19:55
n4nd0inline T pop() { return *(--end); }19:55
n4nd0shouldn't be a check around there?19:55
blackburnit isnot safe :)19:55
n4nd0because it is meant to be so?19:55
gsomixsonney2k, can I ask you? :)19:56
@sonney2kgsomix, please always ask - not ask if you can ask!19:57
@sonney2kn4nd0, well speed trade off19:58
@sonney2kchecks cost quite a bit of performance when you do them *often*19:58
gsomixsonney2k, it's offtop. :) what have you studied in your minor/major courses at the university?19:58
blackburnn4nd0: yeah it tries to be as fast as it could be19:59
n4nd0blackburn: ok19:59
@sonney2kgsomix, well major: computer science, minor: physics19:59
blackburnI never knew you studied physics :)20:01
gsomixsonney2k, thanks. :) do you like physics?20:02
@sonney2kgsomix, back then yes - now I like machine learning and traffic :D20:02
gsomixI have major in applied math and physics. It's long.20:08
gsomixI cut up to "physics" in my proposal :)20:08
gsomixbecause blackburn is more a mathematician than I. :320:13
@sonney2kalright guess I've requested 8 slots for gsoc (I think we really have 8 very very good students that applied) and 9 if there was an inf amount of slots available20:13
@sonney2kspare time for me now!20:14
* sonney2k goes watching the revision demo party live stream20:14
blackburndemoscene memories? ;)20:15
@sonney2kI am too old :D20:15
@sonney2kI mean I went to 'scene' parties since 1993 or so20:15
blackburnsonney2k: really? I was 2 :D20:16
n4nd0blackburn: I doubled your age at that time :P20:17
gsomixnow it is clear how the CPU Fire Applet appeared. demoscene! :)20:17
blackburnyeah kind of20:17
blackburnlorenz stuff as well I guess20:18
n4nd0is any of you guys fan of topcoder or programming contests in general?20:20
n4nd0just curious20:20
blackburnoh I do not like that stuff pretty much :)20:20
gsomixn4nd0, just codeforces (russian analog). but I'm not very good in sports programming.20:21
n4nd0I think it is quite fun :)20:22
n4nd0but kind of difficult, lot of good people around there20:22
@sonney2kyeah I wrote lorenz at a demoscene party :)20:23
blackburnn4nd0: my main point is that these guys can hardly fit to real programming20:23
@sonney2kgsomix, yes...20:24
gsomixn4nd0, this system differs from topcoder. five tasks, two hours.20:24
n4nd0gsomix: I have participated in codeforces ;)20:24
@sonney2kblackburn, it is also the reason I know the uint and int operations have same costs20:24
gsomixn4nd0, ah :)20:24
blackburnsonney2k: yeah I know that20:24
@sonney2kand all that low level assembly programming stuff20:25
@sonney2kanyways have a nice evening20:25
gsomixn4nd0, what is your nickname?20:25
blackburnsonney2k: you too20:25
* sonney2k party time *yay*20:25
n4nd0gsomix: fj.iglesias20:26
blackburnwow it seems you are finally happy with spare time :)20:26
gsomixn4nd0, thanks.20:26
blackburn(that you have that background)20:26
n4nd0gsomix: I suck at sports programming, but I think it is fun :P20:27
PhilTilletHello everybodyyyy20:30
PhilTilletwhat was the name of that gay video yesterday ?20:33
PhilTilletaaah yes20:34
PhilTilletthis is seriously completely wtf music20:36
blackburnmusic is not as wtf20:40
blackburnbut video..20:40
PhilTilletGolden WTF Awards.20:43
gsomixwtf music is: :)20:45
PhilTilletJaaaaapaaaaan = WTF20:45
PhilTilletOk this music is completely f*cked up. XD20:46
PhilTilletPerfect background for coding.20:46
gsomixjapan and breakcore :]20:46
PhilTillet wtf music+video20:46
gsomixOne my friend loves this music20:47
PhilTilletme too, I can actually dance it ...20:47
PhilTilletwell I could20:48
PhilTilletdidn't have time to practice last months..20:48
gsomixI can not imagine how you can dance to this music :)20:51
PhilTilletit's the epilepsy dance20:52
gsomixou, I prefer something a little more slowly20:55
PhilTilletwhere are you from gsomix ?20:56
PhilTilletRussia is so big20:57
PhilTilletwestern Russia?20:58
gsomixPhilTillet, yep. Samara city20:58
PhilTilletI see :)21:03
shogun-buildbotbuild #203 of nightly_none is complete: Success [build successful]  Build details are at
gsomixshogun-buildbot, thanks. :)21:16
PhilTilletI love this buildbot21:17
PhilTilletblackburn, i'm pushing the nearest centroid in a few minutes I think :p21:21
shogun-buildbotbuild #202 of nightly_all is complete: Success [build successful]  Build details are at
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]21:27
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has joined #shogun21:51
PhilTilletblackburn, i've just sent the pull request for nearest centroids21:57
blackburnokay nice21:57
PhilTilletas far as i've tested it, it worked21:57
blackburnI guess wrong branch21:58
PhilTilletah true, I had some problem with git, i forgot to put it in another branch sorry :/21:58
blackburnit is included to opencl21:59
PhilTilletaaaaaah true21:59
blackburnyeah better add new branch :)21:59
PhilTilletyou are right, i'll fix that22:01
PhilTilletbut it might take some time :D22:02
PhilTilleti wish i was a git expert22:02
PhilTilletokay, done :)22:07
PhilTilletnow i must find how to cancel the one on the opencl branch22:10
PhilTilletI think I can't, could you delete that commit in the wrong branch?22:11
blackburnno I can't22:13
blackburnyou can22:13
blackburnyou can revert it actually22:14
PhilTilleti've just figured it out22:14
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]22:24
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has joined #shogun22:41
PhilTilletblackburn, what did you mean with "May be it would be better to use matrix there and then construct features" ?22:45
blackburnPhilTillet: you construct features from matrix and then enqueue its feature vectors22:46
blackburnwhy not to construct matrix and then construct features?22:46
PhilTilletOkay, I'll do that22:49
-!- genix [~gsomix@] has joined #shogun22:56
-!- gsomix [~gsomix@] has quit [Ping timeout: 264 seconds]22:56
-!- blackburn [5bde8018@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]23:03
genixgood night, guys23:24
PhilTilletgood night :)23:24
n4nd0good night23:29
-!- n4nd0 [] has quit [Ping timeout: 272 seconds]23:46
-!- blackburn [~qdrgsm@] has joined #shogun23:57
--- Log closed Mon Apr 09 00:00:19 2012