Open in new window / Try shogun cloud
--- Log opened Sun Oct 23 00:00:34 2011
-!- blackburn [~blackburn@] has joined #shogun11:32
blackburnsonney2k: how it is going? :)17:05
CIA-3shogun: Sergey Lisitsyn  * re74284b / (16 files in 5 dirs): Improved arpack wrapper, introduced LLTSA -
CIA-3shogun: Sergey Lisitsyn  * r1ca6b60 / (3 files): Added references to LLTSA and NPE, updated NPE impl -
CIA-3shogun: Sergey Lisitsyn  * r0358e2e / data : Updated data -
CIA-3shogun: Sergey Lisitsyn  * rf236e25 / applications/faces/ : Added faces dimreduction and clustering application -
CIA-3shogun: Sergey Lisitsyn  * re7b34f1 / src/shogun/converter/DiffusionMaps.cpp : Removed normalization from DiffusionMaps -
CIA-3shogun: Sergey Lisitsyn  * rbf0bf48 / (2 files): Added warning of LLTSA and NPE are hardly applicable to very high-dimensional data -
@sonney2kblackburn, fine20:32
blackburnsonney2k: I did a lot of things today :)20:32
@sonney2kthe orange people will implemt the python buffer protocol20:32
@sonney2kand maybe will add some wrappers such that all of shogun's machine's can be used in their framework20:33
@sonney2kthe other way round is difficult (their code is python...)20:33
@sonney2kso the buffer protocol is what will help us to e.g. use their data loading etc routines20:33
@sonney2kblackburn, then I talked to some llvm/clang guy20:34
@sonney2kand we discussed how to reduce memory requirements in compilation20:34
@sonney2kclang requires only 1/2 of what gcc needed btw20:34
@sonney2kso in the end I now have a plan and will try to split up the file20:35
CIA-3shogun: Soeren Sonnenburg  * r89c91c4 / (src/shogun/io/InputParser.h src/shogun/io/SimpleFile.h): fix compiler errors occurring with clang -
@sonney2kblackburn, I also talked to the opencv guy (gary bradski)20:36
@sonney2kthey have some new input features they call 'orb'20:36
@sonney2kso it makes sense to add some example based on these (or the more general concept)20:37
blackburnsonney2k: orb? what is it?20:37
@sonney2kblackburn, then we long discussions with all the science guys20:37
@sonney2kabout basically everything20:37
@sonney2kwas great20:37
@sonney2kblackburn, sth like sift20:37
blackburnI see :)20:37
blackburnsonney2k: have you kicked chris?20:37
@sonney2kin what respect20:38
blackburnin ass20:38
@sonney2kI told him that you want to talk to him20:38
@sonney2khe said no time :)20:38
@sonney2kwell he said in november again20:38
@sonney2khe will fly to NY from here20:39
@sonney2kand so is pretty busy...20:39
blackburnyeah I know20:39
blackburnokay will try to finish paper by myself :D20:40
blackburnsonney2k: how soon will you fly to berlin again?20:42
blackburnsonney2k: I have integrated pics:
blackburnsonney2k: you should like it :D20:49
@sonney2knetwork is slow :(20:52
-!- in3xes [~in3xes@] has joined #shogun20:55
blackburnsonney2k: it is embedding by diffusion maps and clustering by kmeans21:02
@sonney2kI have the figure now21:06
@sonney2kthe example is not perfect but mostly works21:06
@sonney2kI mean she is smiling in the upper left21:06
@sonney2kbut this should be on the right side21:06
@sonney2kmaybe you can make it more simple ?21:06
@sonney2kjust same expression in face and then turn head to the left ..middle...right?21:06
blackburnsonney2k: it depends on method applied21:08
blackburnsonney2k: I have no time to take more photos more precisely :)21:08
blackburnsonney2k: treat it as demo21:08
@sonney2kbut certainly nice :)21:09
blackburnsonney2k: two new methods available: LLTSA and NPE21:14
blackburnunfortunately they involves crazy complex task21:14
blackburne.g. for pics 200x200 it requires to solve 40000x40000 *generalized* eigenproblem21:14
-!- in3xes_ [~in3xes@] has joined #shogun21:27
-!- in3xes [~in3xes@] has quit [Ping timeout: 240 seconds]21:30
-!- in3xes__ [~in3xes@] has joined #shogun21:44
-!- in3xes_ [~in3xes@] has quit [Ping timeout: 260 seconds]21:47
-!- in3xes__ [~in3xes@] has quit [Quit: Leaving]21:57
blackburnsonney2k: still around?23:14
blackburnsonney2k: I would need kernels between images23:14
blackburnessentially I would need to have kernels between say 20x20 and 40x30 images23:15
blackburnI hope to come up with any idea of that kernel23:15
blackburnwe need ImageFeatures or so23:15
@sonney2kno you could use gaussian kernel on bytefeatures23:17
blackburnsonney2k: eh?? how to store images of different sizes in bytefeatures?23:17
@sonney2kwhy not scale them to some standard size alwyas?23:18
blackburnsonney2k: do you think it is always possible?23:18
@sonney2kwhy not?23:18
blackburnsonney2k: it leads to losing information23:19
blackburne.g. if we have 200x150 pic23:19
blackburnand 20x1523:19
blackburnif we resize 200->20 we will lose too much23:19
blackburnbut I'm pretty sure there is a way to define kernel between it23:20
@sonney2kthen scale them all to some reasonable size - say 50x50 or so23:20
blackburnsonney2k: so do you think it is the only way?23:22
blackburnsonney2k: are you against image features?23:36
@sonney2kblackburn, you mean 2d features?23:36
@sonney2kthat have arbitrary sizes?23:36
blackburnsonney2k: yeah generally yes23:36
@sonney2klike stringfeatures?23:36
@sonney2kwell it will just explode the wrapped code even more :)23:37
blackburnehhh how?23:37
@sonney2kanother feature class for arbitrary types...23:37
@sonney2kI got to reboot23:37
blackburnsonney2k: ok23:37
blackburnsonney2k: I'm pretty sure it would have ~4 types23:50
blackburnI don't think anybody wants to have float64_t 2d features23:50
@sonney2kblackburn, well I don't mind23:54
blackburnsonney2k: yokay23:54
blackburnsonney2k: I don't want to resize images because I've got to make some classifying on roadsigns dataset23:55
blackburnand pictures are pretty bad23:56
blackburnresizing could make it even worse23:56
blackburnsonney2k: is it true "ich bin ein berliner" is a right german?23:59
--- Log closed Mon Oct 24 00:00:40 2011