Open in new window / Try shogun cloud
--- Log opened Fri May 13 00:00:57 2011
-!- alesis-novik [~alesis@] has quit [Quit: I'll be Bach]00:15
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]00:18
-!- ameerkat [] has quit [Ping timeout: 240 seconds]01:29
-!- ameerkat [] has joined #shogun02:06
-!- alesis-novik [~alesis@] has joined #shogun03:37
-!- serialhex [] has quit [Remote host closed the connection]05:18
-!- serialhex [] has joined #shogun05:20
-!- f-x [b49531e3@gateway/web/freenode/ip.] has joined #shogun06:23
-!- f-x [b49531e3@gateway/web/freenode/ip.] has quit [Quit: Page closed]08:48
-!- warpy [] has joined #shogun08:51
CIA-31shogun: Soeren Sonnenburg master * r88994f7 / (71 files in 17 dirs):10:36
CIA-31shogun: Introduce SGVector, SGMatrix, SGString, SGSparseMatrix data types.10:36
CIA-31shogun: And add a get_labels() function for SGVector testing. -
CIA-31shogun: Heiko Strathmann master * r764d28d / src/libshogun/base/SGObject.cpp : remove redundant save/load for model selection parameters -
CIA-31shogun: Heiko Strathmann master * rffeca12 / examples/undocumented/libshogun/parameter_set_from_parameters.cpp : added copyright -
CIA-31shogun: Heiko Strathmann master * r5bc70da / examples/undocumented/libshogun/Makefile : added parameter iteration example -
CIA-31shogun: Soeren Sonnenburg master * ra7c1e48 / src/libshogun/features/Labels.h : introduce set_labels based on SGVector<float64_t> data type -
CIA-31shogun: Soeren Sonnenburg master * rec32295 / (6 files in 3 dirs): Merge branch 'master' of git:// -
@sonney2kwarpy, you wanted to ask sth?12:00
-!- ameerkat [] has quit [Ping timeout: 240 seconds]14:00
CIA-31shogun: Soeren Sonnenburg master * r29dfea9 / (3 files in 3 dirs): add get/set vector in/out typemap for python -
-!- serialhex [] has quit [Remote host closed the connection]17:41
-!- blackburn [~qdrgsm@] has joined #shogun17:55
-!- serialhex [] has joined #shogun18:02
-!- ameerkat [] has joined #shogun20:37
@sonney2kwarpy, ?20:52
@sonney2kblackburn, I did the SGVector thing for one examples - CLabels20:53
blackburnsonney2k: aha, I have seen20:54
@sonney2kblackburn, it seems to work from python already20:55
@sonney2kI realized that we need this fro ndim arrays too20:56
blackburnsonney2k: what do you think about implementing CLLE this way?20:56
@sonney2kehh how is this related?20:57
blackburnI could use SGMatrix instead of float64_t :)20:58
@sonney2kyes but algorithmwise there is no difference...21:01
blackburnit could be more readable21:01
@sonney2khmmhh not much21:02
@sonney2kanyways I will try to do the python typemaps for matrix...21:03
warpywhat about c# ? :P21:03
blackburnbut the problem is if we will use *only* SGVector - we have to change internals for its usage21:04
blackburni.e. if get_feature_vector will return SGVector we would rather use it21:04
blackburnI'm not sure I understand concept21:05
blackburnare SG* structures purposed for interfaces?21:05
@sonney2kwarpy, well I won't do more except the draft - I was hoping you continue work on this...21:07
@sonney2kblackburn, we have to change serialization code to only work with SGVector/Matrix etc and indeed all classes to internally use these things21:08
warpyi was talking about the code daniel did21:10
warpydidnt he send it to you21:10
warpythe alpha version21:10
warpyto put inside the git repo21:10
@sonney2kwarpy, did he finally get the example to run?21:10
warpyi think so.21:10
blackburnsonney2k: so would you like if I will change algos I implement(ed) for its usage?21:10
@sonney2kwarpy, I answered but he didn' treply21:10
warpyhe told me we can start using more than just the simple basic example21:11
warpybecause it compiled and returned "1" or something, correct ?21:11
@sonney2kblackburn, yes :)21:12
blackburnsonney2k: okay21:13
warpywas that yes for me ?21:17
warpyah yes i wanted to ask you21:17
warpyhow does swig/shogun deal with sparse matrices21:17
warpyas far as we see the java code that we trnaslated to c# only deals with arrays and matrices.21:18
warpyis there something special that we need to do to support sparse matrices ?21:19
@sonney2kwarpy, yes21:22
@sonney2kwarpy, is there any sparse matrix library in c#?21:24
@sonney2kblackburn, the good news is that typemaps are drastically simplified21:52
@sonney2kwe now just need in and out typemaps21:53
@sonney2kand no more %apply stuff21:53
blackburnaha, I see, moreover it is more comprehensive21:53
@sonney2kas soon as someone returns a SGVector or SGMatirx etc these things are swig wrapped21:53
@sonney2ksame for inputting21:53
@sonney2kthe only problem I have is that it is hard to do without breaking things at the moment21:54
@sonney2kI think one would need to a) write all the new typemaps for all languages21:54
@sonney2kb) add set/get functions to all relevant classes21:55
@sonney2k(that use SGVector & friends)21:55
@sonney2kc) add functions to do the serialization with SGVector etc21:55
blackburnwho it could be? :D21:55
@sonney2kd) translate all classes to use SGVector internally21:55
@sonney2kconverting typemaps is piece of cake (I just did matrix for python)21:56
@sonney2kadding set/get functions too21:56
@sonney2kserialization is something heiko could help21:57
blackburnaha, d) is for me :)21:57
@sonney2kd) is potentially more demanding21:57
blackburnanyway we don't need to rush d), right?21:58
blackburnsonney2k: LLE for the moment works ~20s on 1000021:58
blackburnwith no eigenproblem solving though21:58
@sonney2kthere is no rush for anything - but it will give us much more readable code22:00
blackburnI mean I could do it for sure, but I haven't much time right now :(22:01
@sonney2kblackburn, well you cannot start before a-c are done22:02
@sonney2kand believe me - you have much more time than I have currently...22:02
blackburnsonney2k: hrr, looks like you think I'm relaxing :)22:05
@sonney2kblackburn, no I don't think you are.22:06
-!- serialhex [] has quit [Remote host closed the connection]22:06
blackburnsonney2k: SystemError: Out of memory error, tried to allocate 800000000 bytes using new[].22:07
@sonney2kWe should postpone this discussion until you too have 2 babies at home :)22:07
* sonney2k counts the zeros22:07
blackburnfloat64_t 10000*10000 matrix22:07
blackburnI really need sparse matrix :D22:08
-!- serialhex [] has joined #shogun22:08
@sonney2kif the matrix is sparse then yes :)22:09
blackburnsonney2k: 5.690s for 5000 examples22:09
blackburnnot so bad this time22:09
blackburnbut memory complexity is terrifying22:09
blackburnin each row of that NxN matrix only k<<N values aren't zero22:10
* blackburn is in the crape, russia is beaten in 1/222:11
@sonney2kok then sparse could work22:17
@sonney2kbut no idea if all the algos you need are available for sparse matrices22:17
blackburnsonney2k: only eigenvalues22:18
@sonney2kI never did that fro sparse matrices...22:18
blackburnme too but I don't know any other way22:19
blackburnsonney2k: nevermind, on last stage it is not sparse22:20
@sonney2kanyways I wouldn't optimize before you have a set of baseline algos22:21
blackburnW is sparse matrix, then (I-W)'*(I-W) is computed and eigenvectors of it are feature vectors22:21
blackburn(I-W)'*(I-W) is exactly don't sparse :)22:22
@sonney2kgood night!22:34
blackburngood night..22:38
blackburnLLE is almost working now22:38
-!- ameerkat [] has quit [Ping timeout: 260 seconds]23:32
-!- ameerkat [] has joined #shogun23:59
--- Log closed Sat May 14 00:00:58 2011