IRC logs of #shogun for Monday, 2011-10-10

« Sunday, 2011-10-09 Index Tuesday, 2011-10-11 »
--- Log opened Mon Oct 10 00:00:28 2011
-!- blackburn [~blackburn@85.114.187.69] has quit [Read error: Connection reset by peer]00:19
-!- blackburn [~blackburn@85.114.187.69] has joined #shogun00:19
-!- blackburn [~blackburn@85.114.187.69] has quit [Quit: Leaving.]00:24
-!- av3ngr [96cbdf9c@gateway/web/freenode/ip.150.203.223.156] has quit [Ping timeout: 252 seconds]04:30
-!- alesis-novik [~alesis@dhcp-90-089.inf.ed.ac.uk] has joined #shogun17:54
-!- blackburn [~blackburn@188.168.4.72] has joined #shogun20:28
-!- blackburn [~blackburn@188.168.4.72] has quit [Ping timeout: 258 seconds]20:33
-!- blackburn [~blackburn@188.168.4.100] has joined #shogun20:34
blackburnsonney2k: what's up?21:36
@sonney2kmaking progress on the debian package...21:36
blackburnoh that's pretty nice21:36
blackburnwould you mind making .exe installation?21:36
blackburnwell windows is a piece of shit but21:37
@sonney2kimpossilbe21:37
blackburnпочему?21:38
blackburnsonney2k: okay, rpm?21:39
@sonney2kit will be some cygwin stuff21:39
blackburnsonney2k: I have one problem still to be solved - I can't use dimensionality reduction on say SparseFeatures although it is possible to do21:46
blackburnit makes me mad21:47
blackburnI can't sleep21:47
blackburnI eat wood21:47
blackburnok joke21:47
@sonney2kwell that's life21:53
@sonney2kyou cannot have it all21:53
blackburnsonney2k: I see one solution but I guess you won't like it21:55
blackburnsonney2k: DimensionReductionPreprocessor as proxy of DimensionReductionMachine, but not derived from CMachine21:57
blackburnwith methods apply(CFeatures*) returning CFeatures*21:57
blackburnhow it is for you>?21:59
blackburnmoreover I could even return SimpleFeatures all the time21:59
blackburnimagine how useful it would be to input StringFeatures and get SimpleFeatures representing it22:00
blackburnain't it?22:00
@sonney2kyes but your solution is not nice :(22:04
@sonney2kI think we need the concept of 'converters' for that22:05
@sonney2kthat take as input feature type X and are allowed to return Y22:05
blackburnthink about it please, I can't see any better solution22:06
blackburnsonney2k: I'm now using svn at my job heh22:13
blackburnugly thing, don't know how you had used it before gsoc22:13
@sonney2kblackburn, why not have a converter class that has an apply method and gets features in and returns features22:15
blackburnsonney2k: it is the same I propose22:16
@sonney2kno much better - more general and more clear22:16
blackburnhow could we call it?22:16
@sonney2kConverter22:17
blackburnhah22:17
blackburnso, DimensionReductionConverter?22:17
blackburnI like it22:18
@sonney2kwhat I don't understan though is how you proxy w/ preprocs22:18
blackburneasy22:18
blackburnDimensionReductionPreprocessor(LLE)22:18
blackburnor22:18
blackburnpreproc.set_technique(LLE)22:19
blackburnpreproc.set_technique(BLABLAE)22:19
blackburnis it bad for you?22:19
blackburnit just delegate responsibilities to converter22:20
blackburnI have no in-place algorithms22:20
blackburnso proxy will just take and replace feature matrix22:20
blackburnafter converting22:20
blackburnwhat I would say with some confidence - NLDR techniques are used as converters more often than as preprocessors22:22
blackburnwith proxy we will still able to built even chains of these techniques (if it is needed, it is a little crazy thought)22:23
@sonney2kblackburn, couldn't you add an inplace modify option?22:24
@sonney2kIt certainly is not supported to dealloc / alloc the features in the preproc22:25
blackburnsonney2k: in most cases it is impossible22:25
blackburnI do it in every algorithm22:25
blackburnI change feature matrix22:25
blackburnsonney2k: very difficult challenge to replace it in-place22:27
blackburnand well22:27
blackburnit is not that bad to have 2 matrices of N*dim while I have 1-2 matrices of N*N, right?22:28
blackburnsonney2k: or do you mean any other cons of replacing feature matrices?22:29
@sonney2kblackburn, well why not create a tmp copy then and do the mdoification later inplace22:29
@sonney2k(if there is no easy other way)22:30
@sonney2kcertainly ok for most of the applications22:30
blackburnwhat do you mean?22:30
blackburntmp copy of ?22:30
@sonney2kother preprocessors didnt' need any helper array22:30
@sonney2kdimred feature matrix22:30
blackburnI need a LOT of arrays22:30
blackburnand almost 2x memory for 2 threads22:31
blackburnetc22:31
blackburnso what do you suggest? sometimes it is a little more possible but I would have to make transpose of non-square matrix in-place and etc22:32
blackburnI have no time now for such crazy algos22:32
blackburnmy algos are not that easy :)22:34
blackburnyou know I guess22:34
blackburn*valgrind became my favourite tool*22:40
@sonney2kblackburn, well when you have a copy you can just copy it over to the original matrix22:42
@sonney2kso no problem22:42
blackburnit is just technical issues22:43
blackburnnot to use set_feature_matrix but overwrite gotten feature matrix pointer22:43
blackburnsonney2k: got an idea22:44
blackburnwhat if check sizes at set_feature_matrix22:44
blackburnwhy to delete available feature matrix?22:45
blackburnif it have the same sizes22:45
@sonney2kblackburn, others might use the ptr22:45
@sonney2kso I don't like the idea of deleting it...22:46
@sonney2kbetter adjust shape (shrink)22:46
@sonney2kand waste a littley memory22:46
blackburnsonney2k: currently in set_feature_matrix we free it22:46
blackburnwhy not to realloc it if necessary22:46
blackburnor just copy content if have the same size22:46
blackburn?22:47
blackburnit will make my preprocessors work in-place22:47
blackburnat least22:47
blackburnnot in-place but with realloc22:47
blackburnI see no reason not to do it22:51
@sonney2kblackburn, one reason would be someone having a ptr to e.g. a SimpleFeature object22:53
blackburnand?22:53
@sonney2kthen changing the data ptr might make things crash22:53
blackburnsonney2k: it is being changed now!!122:54
@sonney2kblackburn, you do this in your prepro*s?22:54
blackburnI just call set_feature_matrix22:54
blackburnof simplefeatures22:54
blackburnjust take a look on SimpleFeatures.h:27322:55
blackburnI suggest not to free it but realloc!22:55
@sonney2kohh well we can give it a try...22:55
blackburnand if it was 400*3 and now should be 200*622:56
blackburnwe even don't need to realloc it22:56
CIA-3shogun: Sergey Lisitsyn  * r00b7950 / (19 files): Cleaned DR preprocessors, removed effective target dim computation - http://git.io/QsWUaw23:12
-!- alesis-novik [~alesis@dhcp-90-089.inf.ed.ac.uk] has quit [Ping timeout: 276 seconds]23:13
blackburnnono23:14
blackburnI don't know what is better :(23:15
blackburncopying is slower than freeing23:15
CIA-3shogun: Soeren Sonnenburg  * r77efd4c / src/interfaces/elwms_static/elwms.cpp : Fix include paths for elwms interface. - http://git.io/W2YdOQ23:18
blackburnsonney2k: it's an achievement there is no memory leaks with my preprocs :)23:20
blackburnbtw I got an idea of memory check test23:20
blackburnwe should run every test with valgrind --leak-check=full and search for 'shogun'23:21
blackburnsonney2k: TU decided not to give you buildbots?23:21
@sonney2kblackburn, it takes time23:21
blackburnbtw I don't understand why my code have copyright of TU23:22
blackburn:D23:22
blackburnsonney2k: why is it a TU property?23:22
@sonney2kblackburn, all the new code you write (outside GSoC) can have some other code23:22
@sonney2klicense23:22
blackburnGSoC code -> TU copyright?23:23
blackburnI'm asking because I can't see TU care any about us23:23
@sonney2kargh, I still have to update all copyright stuff for the debian package23:24
@sonney2kblackburn, patience young padawan23:24
blackburnI'm not about buildbots only23:24
blackburnwhy GSoC code is a TU property while TU doesn't care about it anyhow? :)23:24
@sonney2kit is MPI / TU23:25
blackburnwhat is MPI/TU?23:26
-!- alesis-novik [~alesis@cpat001.wlan.net.ed.ac.uk] has joined #shogun23:27
blackburnI'm pretty interested if I can't change it some day23:27
blackburnI can't, right?23:27
blackburnsonney2k: why my code is licensed over orgs I'm not a member of? :)23:28
@sonney2kblackburn, indeed you cannot23:29
@sonney2kbecause you took part in gsoc and agreed to the gsoc terms23:29
@sonney2kall new code you wrote after gsoc is 100% yours though :)23:30
blackburnsonney2k: impossible to determine what I have wrote after gsoc23:30
@sonney2kgit log23:30
blackburnwrote = new file?23:30
blackburnthen I could remove that shit from KLLE and KLTSA23:31
blackburnah nevermind but that's strange23:32
blackburnyou do all the work and receive no support neither from MPI nor TU23:32
@sonney2kwell we receive support - not as much as you would like but we do23:34
blackburnwhat is the support23:34
blackburn?23:34
blackburnokay I have removed copyrights from new algos ;P23:36
CIA-3shogun: Sergey Lisitsyn  * rf6d08c2 / (4 files): Removed wrong copyrights from preprocessors developed after GSoC - http://git.io/s9V8TA23:36
blackburnfuck I did it in gsoc23:38
@sonney2kblackburn, you should put Copyright (C) 2011 then instead of the writting stuff23:38
blackburnrevert23:38
-!- alesis-novik [~alesis@cpat001.wlan.net.ed.ac.uk] has quit [Ping timeout: 258 seconds]23:38
blackburnyeah that's odd I'll better sleep23:39
CIA-3shogun: Sergey Lisitsyn  * ra92e6b4 / (4 files):23:39
CIA-3shogun: Revert "Removed wrong copyrights from preprocessors developed after GSoC"23:39
CIA-3shogun: This reverts commit f6d08c2226fd1b297744a4c7379fc7f180d6d60e. - http://git.io/YnvpnQ23:39
@sonney2kok23:39
blackburnI still see no link shogun <-> TU23:40
blackburnexcept of you was there23:40
@sonney2kblackburn, we will get 3 buildbots just be patient23:40
blackburnhah23:41
blackburnit would be better to suggest students to take participance23:42
blackburnI will never forget methods I have implemented23:42
-!- alesis-novik [~alesis@cpat001.wlan.net.ed.ac.uk] has joined #shogun23:47
--- Log closed Tue Oct 11 00:00:32 2011
« Sunday, 2011-10-09 Index Tuesday, 2011-10-11 »

Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!