Open in new window / Try shogun cloud
--- Log opened Tue Oct 09 00:00:17 2012
-!- Netsplit *.net <-> *.split quits: @sonney2k00:07
-!- Netsplit over, joins: sonney2k00:19
-!- cronor [] has joined #shogun00:33
-!- heiko [] has joined #shogun00:57
-!- heiko [] has left #shogun []00:57
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]01:45
-!- cronor_ [] has joined #shogun02:11
-!- cronor [] has quit [Read error: Connection reset by peer]02:11
-!- cronor_ is now known as cronor02:11
-!- cronor_ [] has joined #shogun02:51
-!- cronor [] has quit [Read error: Connection reset by peer]02:51
-!- cronor_ is now known as cronor02:51
-!- cronor_ [] has joined #shogun03:08
-!- cronor [] has quit [Read error: Connection reset by peer]03:08
-!- cronor_ is now known as cronor03:08
shogun-buildbotbuild #127 of nightly_default is complete: Failure [failed test]  Build details are at
-!- cronor [] has quit [Read error: Connection reset by peer]04:36
-!- cronor [] has joined #shogun04:37
-!- adoniscik [] has joined #shogun06:49
-!- cronor [] has quit [Quit: cronor]07:27
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has joined #shogun09:26
-!- ptizoom [~ptizoom@] has joined #shogun10:07
-!- adoniscik [] has quit [Quit: Leaving]10:12
-!- cronor [] has joined #shogun10:13
-!- cronor_ [] has joined #shogun10:29
-!- cronor [] has quit [Read error: Connection reset by peer]10:29
-!- cronor_ is now known as cronor10:29
-!- cronor [] has quit [Quit: cronor]11:04
wikingsonne|work: hey, btw: do you have download statistics of shogun 2.0 release? :)11:15
sonne|workwiking: no idea...11:19
wikingsonne|work: do u have the apache logs?:)11:19
sonne|workmaybe that
sonne|workor that
sonne|workwiking: no apache on that machine but fapws311:23
wikingbut it has logs right? :)11:24
wikingnone? :)11:25
wikinghow come :)11:25
sonne|workit sends me emails with error msg if sth goes wrong but no more11:26
wikingi see :S11:26
wikingbtw do we have something like a linked list structure in shogun?11:26
sonne|workCList ?11:27
blackburnsonne|work: copy(vec.begin(), vec.end(), inserter(v, v.begin());11:28
blackburnsonne|work: BOOST:SHARED_PTR11:28
blackburnfeel da pain? :)11:29
blackburnsonne|work: FUNCTORS!11:30
blackburnsonne|work: CHAR_TRAITS!11:30
blackburnsonne|work: abstract factory! singleton!11:33
blackburnsonne|work: MVC! dependency injection!11:33
blackburnwiking: CLEARCASE!11:33
wikingmvc is cool11:35
blackburnwiking: dependency injection as well11:36
blackburnwiking: and boost and stl and all the things sonne|work hate :D11:36
wikingyeah boost is cool11:36
wikingit's a bit bloated though11:36
wikingbut still cool11:37
blackburnI learned it a bit when doing test task for the company11:37
blackburnenjoyed to do some things11:37
sonne|workusing boost increases shogun compile time by factor 1011:40
blackburnsonne|work: without libraries we've got a lot or redundant code as well11:41
wikingsonne|work: but just compile time :)11:41
sonne|workblackburn: I don't mind libraries11:47
blackburnsonne|work: especially standard template one :)11:48
sonne| link against11:51
blackburnand the nobel prize goes to quantum guys11:53
wiking"applications exists on exceptions from SG_ERROR"12:07
wikingthis is the normal case :)12:07
wikingit's not a limitation12:07
wikinghahah who's up for this one: "MemoryMappedFile is not implemented (needs reimplementation based on Win API12:08
blackburnI don't think it is implemented even in linux :)12:10
sonne|workwiking: well no - in cmdline interface it should just return to command prompt12:12
sonne|workbut command line ansi characters can be disabled12:13
wikingsonne|work: what's with that elwms interface? they had it in the previous macports file12:50
wikingso i wonder how do we support it in 2.012:50
wikingi've never used it....12:50
wikingblackburn: have u managed clang_complete to complete for stl?13:04
blackburnwiking: it was working IIRC13:19
blackburnwiking: I wish clang_complete showed doc13:19
blackburnwiking: do you know whether it is possible?13:19
blackburnmaybe it is not that hard to patch it for that, I don't know13:20
sonne|workwiking: probably not too important... it is one .so file that can be used as R extension, python extension, octave & matlab extensiuon at the same time while even enabling to transfer matrices etc across languages (e.g. one can run python from octave etc)13:48
sonne|workwiking: I used that to slowly transition away from matlab - but I wanted to call my newly written python code13:50
sonne|workand e.g. use matplotlib13:50
wikingsonne|work: oh13:59
wikingsonne|work: i think i'll remove it then :D13:59
wikingsonne|work: none the less the port file is now in a good shape13:59
wikingwonder how long it'll take to apply it to the repo13:59
wikingsomebody already tested it and gave some feedback so i've chnaged some minor stuff13:59
sonne|workwiking: probably faster than getting into debian14:13
blackburnsonne|work: 6 minutes support reaction time! faster than all that enterprise :D14:34
blackburn(see mailing list)14:34
sonne|workbut your answer doesn't help14:35
sonne|workhe used exactly that right?14:36
sonne|workI think he is missing lapack/blas/atlas14:36
blackburnsonne|work: no, he used wrong names14:37
blackburnonly one is right but yes you are right he must be missing blas14:37
sonne|workLIBLINEAR_L2R_L1LOSS_SVC_DUAL                 else if (strcmp(name,"LIBLINEAR_L2R_L1LOSS_SVC_DUAL")==0)14:38
blackburnsonne|work: we should have a QA14:45
sonne|workblackburn: we have :)14:46
blackburnsonne|work: users?14:46
blackburnthey could give up too early14:46
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has quit [Quit: Page closed]15:05
-!- blackburn [~blackburn@] has joined #shogun17:25
-!- zxtx [] has quit [Ping timeout: 260 seconds]19:45
-!- heiko [] has joined #shogun21:00
blackburnhey heiko21:11
heikohey blackburn21:11
blackburnheiko: got you message, where should I answer? :)21:12
heikoemail best, I have food on the oven and have to go there now :)21:12
heikoI just found the pyramid match kernel that might be worth a try21:12
blackburnheiko: is it 1/0 image?21:13
blackburnor black/white means no color?21:13
blackburnheiko: pyramid thing seems to be good idea but I'd recommend to think about features construction21:13
heiko8bit image21:15
heikoi am back21:16
blackburnheiko: back? then lets chat :)21:16
blackburnheiko: any assumptions on the 2d relations?21:16
heikoenergy fields21:16
heikoso spatial relationship or at least distance is important21:16
blackburnanything about neighborhood?21:17
blackburncan you say neighborhood of one 'pixel' radius matter21:17
blackburnor more?21:17
heikono that should be fine21:18
heikothe radius21:18
blackburnfirst idea came to me is LBP21:18
heikowhat is that?21:18
blackburnit is a method of constructing features of images21:18
blackburnuseful for textures and faces21:18
heikodont know21:18
heikoI mean, these fields are not really images21:18
blackburnlet me briefly describe21:19
heikobut ok21:19
blackburnfor each 'pixel'21:19
blackburnwe compare it with its neighbors21:19
blackburnso we've got result of 8 compares21:19
blackburnlike left pixel is greater than center21:19
blackburnright pixel is less than center21:19
blackburnthen we compute histograms of such results21:20
blackburnlike in block B we've got mostly left greater than right21:20
blackburnso the histogram describes pattern in the block21:20
blackburngot what I mean?21:21
heikoyes get it21:21
blackburngood thing for you is that we have it in shogun :D21:22
heikoany experience with that? :)21:22
blackburnheiko: I know it works good on faces, not that good on road signs (but ok)21:22
blackburnheiko: face detectors works very good with it21:22
heikoI see21:23
heikowork on the histograms?21:23
heikowith svm for example21:23
blackburnheiko: for histograms best way is to map they explicitly using HomogeneousKernelMap21:24
blackburnand work like in typical linear space21:24
blackburnwith no kernel21:24
blackburnfast and nearly no accuracy loss21:24
heikoHomogeneousKernelMap just takes a bunch of histograms and produces a sparse representation of them?21:25
blackburnheiko: no, unfortunately not sparse and of 3x size21:25
blackburnbut linear21:25
blackburnsay you have kernel21:25
blackburn<x,y> = \sum_i min(x_i, y_i)21:25
blackburnthen HKM is a map M21:25
blackburn<Mx,My> = k(x,y)21:26
blackburnflaw is that Mx is have to be at least 3 times bigger21:27
heikoI see21:27
blackburnit comes from fourier21:27
blackburnso sin, cos and 121:27
heikowell that sounds like a good approach21:27
heikodo you have any lbp code?21:27
blackburnheiko: shogun/features/LBPPyrDotFeatures.cpp21:27
blackburnyou could try that first21:28
heikono documentation :(21:28
heikobut dense_dot  is implemented21:29
heikoblackburn, looks like the code does what you said21:30
blackburnheiko: yeah I know :D21:30
blackburnheiko: did you plot these fields?21:31
heikono, just reading the code21:31
blackburnheiko: I mean ones you are classifying21:31
heikooh, yes I did21:31
heikovery sparse21:32
heikoand I cannot find a pattern21:32
heikowith my eye and brain21:32
blackburnwhat are the classes?21:32
heikosvm with RBF kernel gives about 60% accuracy21:32
heikobut using vectors of the images is not good21:32
blackburnyes bad idea always21:33
heikoonly 1D local structure21:33
heikohow to do this best with shogun21:33
heikothere is only this dense_dot method21:33
blackburnheiko: so looks like noise for you?21:33
heikoI have to create a matrix of all the dot products and then use as custom kernel matrix right?21:34
blackburnheiko: why? add is here to21:34
heikoblackburn, more or less, yes21:34
blackburnyeah line 28621:34
heikowhat can I do with add? :)21:34
blackburnit is COFFIN features in general21:34
blackburnadd and dense dot is the best a cat can get21:35
heikohow to use then?21:35
blackburnwith liblinear or ocas21:35
blackburnjust like dense features21:35
heikoso just pass these features to liblinear and thats it?21:35
blackburnyeah if there is no bug there yes21:35
blackburnother way is to employ opencv but I'd recommend shogun first21:36
blackburnheiko: did you try dimreduction?21:36
heikono I didnt but this should be pointless here21:37
heikosvm are designed to work in high dimensions21:37
blackburnheh no, not for features construction21:37
blackburnfor visualization21:37
heikoi see21:37
heikohow would you do that?21:38
blackburnheiko: is data big?21:38
heikonever tried this21:38
heikoindividual images are not21:38
heikobut I have 20000 of them :)21:38
blackburndo you have 4-5 gb of free ram then?21:38
heikoI have21:39
heikobut what would you do?21:39
blackburnwell I'd try MDS21:39
blackburnjust to see whether some compactness is here21:40
blackburnheiko: create MultidimensionalScaling()21:40
blackburncall set_landmark(true)21:40
blackburnuhh I am bad21:41
blackburnit would compute all 20000x20000 distances21:41
blackburnstill if you don't mind call set_target_dim(2)21:42
blackburnand call embed on your features21:42
heikowhat do I get?21:42
blackburnit will output 2d embedding (dense features)21:42
heikowhat does it do?21:42
blackburnscatter plot it to see how data is distributed according to distance21:42
heikoI see21:42
blackburnit attempts to save distances21:43
heikoone thing I know is:21:43
heikoI can successfully do two-sample tests on the data21:43
heikoso there is a difference21:43
heikoand this already happens when using image vectors21:43
heikomaybe I should dig a bit in there first ...21:43
blackburnMDS tries to find 2d embedding w.r.t. distances21:43
heikodistances of what?21:44
blackburnso far vectors are far and close are close21:44
blackburnbetween each other21:44
heikoimage columns?21:44
blackburnno, vectors21:44
blackburnbetween images in general21:44
heikothese are vectors21:44
blackburnthen yes21:44
heikook, thanks for the hints21:44
heikoIll let you know what I got21:44
blackburnheiko: okay, will wait for  that21:45
heikoprobably will take a while - you know I am slow :D21:45
blackburnheiko: did you try to find anything about energy fields?21:45
heikono, but I also dont want to invest too much time now21:45
heikofirst want to get some usable results21:45
heikoand then refine etc21:45
heikoso imaging stuff might be a good start21:45
blackburnI think MRF could be a good idea but I am a bit lame in graphical models21:46
heikoyeah same here21:46
heikoactually, there is this new kernel belief propagation which can do inference in pairwise MRF21:46
heikothis could be wortha try21:46
blackburnheiko: Ising model was created for something like that21:47
heikobut I havent yet understood it :)21:47
blackburnin 1934 :D21:47
heikolol :)21:47
blackburnheiko: belief propagation kernel? interessant to see21:47
heikoI am planning to implement that into graphlab btw :)21:47
blackburndarf ich ansehen?21:47
heikoIf I find some time21:47
blackburnwhat is graphlab?21:48
heikoan open-source distriuted graph-based ML software21:48
heikocan do stuff like pagerank in a distributed fashion21:48
blackburnI see21:48
heikoand graphical models fit in there very well21:48
blackburnheiko: did I say that right in german? :D21:49
heikonein :)21:49
heikowhat did you want to say exactly?21:49
blackburnwhat is wrong?21:49
heikoDarf ich mal sehen - but that is really slangy :)21:49
blackburnwell I tried to change typical 'darf ich vorstellen' (I learned that from a book :D)21:50
heikoKannst du mir das zeigen21:50
heikowould be better21:50
blackburncan you show me?21:50
blackburnI know du and kannst and das21:50
heikothere is the subject missing thats also wrong in english :)21:50
blackburnheiko: where?21:51
heikocan you show me what?21:51
blackburnah well not that wrong I thought21:51
blackburnI mean when we are talking and you got what I mean :)21:52
heikoyes, its just a language thing21:52
heikoyou always have a subject in any roman language21:52
blackburnyeah I know21:52
blackburnbut it is lost sometimes still, right?21:53
heikoespecially in chatrooms21:53
blackburnheiko: I am now developing that test task thing - got totally crazy with design21:55
blackburnEntityManager + EntityFactory + EntityLoader + EntityKeeper21:55
heikolol :D21:55
blackburnI should stop describing new entities of entities21:56
blackburnor I will end with a metamodel21:56
blackburnof a metamodel21:56
blackburnheiko: may be it is good that sonney2k is here to stop my crazy intentions :D21:58
blackburnheiko: if I get that job I am going to become commercial counterpart of alex binder :D21:59
blackburnmostly the same that he does22:00
heikocool thing22:00
heikoyou two then can join each other to create a super ML monster :LD22:00
blackburnwe should all meet and constuct a MLdzilla22:01
blackburnheiko: I recently noticed that not only shogun slowed down, scikits is being developed by other people now22:05
heikoshogun slowed down?22:06
blackburnwell two commits per week22:06
blackburnor so22:06
heikoI see22:07
heikobut we got the gsoc stuff22:07
heikothat was massive this time22:07
heikoI am still working quite a bit, but the streaming framework is so broken that I cannot commit :)22:07
heikospent almost all yesterday trying to fix the stuff22:07
blackburnI will become active once I get or not get the job22:07
heikobut hope to be done soon22:07
blackburnnow I am thinking how to design that thing not to look stupid22:08
blackburnhowever I have to hurry up, I said it will be ready by 1122:08
blackburnof oct22:08
heikothen gogogog :)22:09
blackburnit works already but they asked multithreading22:09
blackburnthinking how to avoid redundant locks..22:09
heikolots of things22:10
blackburnand JSON22:11
blackburnfor loading22:11
blackburnthey didn't ask for JSON but it seems to be easier for me22:11
blackburnthan handcrafting my own format22:11
blackburnahh and UML22:12
heikocool !: :)22:12
heikoalthough I hate UML22:13
blackburnclass diagram is ok22:13
blackburnI hate sequence diagram22:13
blackburnI've never seen a process that is simplified by a sequence diagram22:14
blackburnwasting paper maybe22:14
blackburnor spending time to avoid doing things at job :)22:14
heikoI also never saw this22:15
heikowhen the processes are easy, it works fine, but then you dont need a diagram22:15
heikowhen the processes are complicated, its a mess22:16
heikobetter use plain blackboard :)22:16
blackburnheiko: the only thing is that there are some big companies22:18
blackburnbut they should rather disappear22:18
blackburnoracle especially22:18
heikowhat do you mean?22:20
blackburnthey have to work with all that IDEF0, UML, anything22:20
blackburnbecause they are so big that person means nothing22:21
blackburnit should work even if robots come22:21
heikohehe you got a quite bad image of these guys dont you? :)22:30
blackburnheiko: do SAP and oracle deserve respect? :)22:36
heikolol SAP22:37
heikoI heard a talk of this statistical analysis software22:37
heikohow is it called again?22:37
heikothe guy was in our uni early this year22:37
heikodude, this software is ridiculus22:37
blackburnyes, SAS22:37
blackburnheiko: no idea about that, what does it contain? kmeans and linear regression?22:38
heikoand programming in comic sans :)22:38
blackburnoracle data mining is the same thing22:38
blackburnit has kmeans, svm and apriori22:38
blackburnnot even fp-tree, just apriori22:38
blackburnhey 70s22:39
heikobut this will change22:40
heikofor all the new methods, you really need to know what you are doing22:40
blackburnyeah once we get super semi svm they will get MKL22:40
heikoyeah maybe22:41
heikoI mean these things require quite some knowledge22:41
heikobut KMeans does not22:41
blackburnsure but I don't think they aim data mining products for people who doesn't know what data mining is22:42
heikoMKL, man, this is so specialised knowledge22:42
blackburnI understand, just some example22:42
heikoI prefer open-source stuff anyway22:42
heikobut companies need reliable things that are known to work22:42
heikoand these products offer this22:43
blackburnno they need money :)22:43
heikothey also can generate power-point presentations from data :)22:43
blackburnyeah that's true business thing22:43
heikobtw just running two-sample tests on the vectorised images with gaussian kernel22:44
heikoand I get 0% type 2 error22:44
heikostrange that the SVM fails22:44
heikoat least now I know which kernel to use to maximise differences between the classes22:44
blackburnheh yeah statistical approach would work too I think22:45
blackburnhowever you are the expert here22:45
heikothis stuff is related22:45
heikothere is a formal connection between linear hard margin SVM and MMD test22:45
heikoso if two-sample test works good, svm should work well too22:46
blackburnthere are connections between svm and everything :)22:46
heikoindeed :)22:47
--- Log closed Wed Oct 10 00:00:13 2012