Open in new window / Try shogun cloud
--- Log opened Mon May 09 00:00:50 2011
blackburnI see00:01
blackburnsonney2k: do you plan to change internals for SGVector, etc usage?00:15
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]01:29
-!- serialhex [] has joined #shogun05:09
-!- blackburn [~qdrgsm@] has joined #shogun10:14
@sonney2kblackburn, would love to10:31
@sonney2knot sure if I have time to do it now.10:31
blackburnme too10:31
blackburnbut I think we could do it during summer10:32
@sonney2kmaybe a first step is to put the struct definition in DataType.h10:32
blackburnI think we should make definition and then convert one algo to its usage10:32
@sonney2kthere is one problem though10:32
@sonney2ktypemaps need to be adjusted10:33
blackburnah, yeap10:33
@sonney2kso that means we have to add another round of typemaps10:33
@sonney2k(to not break anything inbetween)10:33
@sonney2kand then gradually remove the others10:33
blackburnso well when I will have holidays (mid of june - .. )10:33
blackburnmay be I will do it for GaussianNaiveBayes10:34
blackburnwith typemaps, etc10:34
blackburnif everything will work fine - we will convert others gradually10:34
blackburnsonney2k: can you recommend some BLAS/ATLAS tutorial?10:37
@sonney2kblackburn, it may actually be simpler to do the typemaps...10:37
@sonney2kwith SGVector etc10:38
@sonney2kblackburn, I only ever used this lapack / blas overview pdf that you find in the net10:38
@sonney2kthere is a book though10:38
@sonney2kblackburn, I am currently finishing the malloc / realloc / free -> SG_MALLOC etc transition10:43
blackburnsonney2k: I need to use atlas for solving systems of linear equations, right?10:51
blackburnand cblas for other routines, like dot product10:51
@sonney2kotherwise it will be tough10:51
blackburnsonney2k: do we have example of atlas usage in shogun? I got stucked with setting it up10:52
@sonney2kblackburn, PCACut.cpp or dot product in lib/Mathematics.h or in pr_loqo.cpp10:53
blackburnseems PCACut use only blas..10:54
blackburnsame thing in Mathematics.h10:56
@sonney2kcompute_eigenvectors in Math uses dsyev11:02
@sonney2kthat is eigenvectors/values for symmetric doubl ematrix11:02
blackburnthank you11:03
blackburnsonney2k: what is wrap_**?11:06
blackburnah, I see11:07
blackburnit is in lapack.h11:07
blackburnsonney2k: the problem was related to my bad atlas installation11:12
@sonney2kblackburn, yeah there are problems with when things are available in atlas or lapack only and different os / libs make a difference11:12
blackburnsonney2k: question now is should I add needed functions to lib/lapack.h or can I use it straight?11:12
@sonney2kblackburn, preferably add it to lapack.h (wrap_*) and then use it11:13
@sonney2kin case one has to change it later it will be easier to modify it there11:13
blackburnaha! I see11:14
blackburnthanks! you helped me very much11:14
-!- [1]warpy [] has joined #shogun12:50
-!- [1]warpy [] has quit [Ping timeout: 260 seconds]13:01
CIA-90shogun: Soeren Sonnenburg master * r7890ec2 / (10 files): remove roc eval functions from static interfaces -
CIA-90shogun: Soeren Sonnenburg master * r6e92985 / (59 files in 13 dirs): Convert all memory functions to use the SG_* ones. -
CIA-90shogun: Soeren Sonnenburg master * ra861f66 / examples/undocumented/matlab_and_octave/classifier_libsvmmulticlass.m : fix multiclass example -
CIA-90shogun: Soeren Sonnenburg master * r84e585e / (src/libshogun/lib/memory.cpp src/libshogun/lib/memory.h):13:07
CIA-90shogun: Introduce SG_MALLOC, SG_FREE, SG_REALLOC.13:07
CIA-90shogun: These replacements to the stdlib functions will throw exceptions in case13:07
CIA-90shogun: of errors (never return NULL). -
blackburnsonney2k: what is the purpose of SG_MALLOC, SG_FREE, SG_REALLOC?13:17
blackburnah I see13:18
blackburnin new code we should use SG_MALLOC, right?13:18
@sonney2kblackburn, usually new[] new()14:03
@sonney2kbut never malloc / realloc / free but the SG_* things instead14:03
blackburnI see14:04
@sonney2kit just has the advantage that you never need to check for NULL ptr's14:05
-!- serialhex [] has quit [Remote host closed the connection]14:48
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]15:35
-!- serialhex [] has joined #shogun16:35
-!- serialhex [] has quit [Remote host closed the connection]17:58
-!- serialhex [] has joined #shogun18:00
-!- serialhex [] has quit [Remote host closed the connection]18:01
-!- [1]warpy [] has joined #shogun18:15
-!- serialhex [] has joined #shogun18:41
-!- [1]warpy [] has quit [Ping timeout: 240 seconds]19:36
-!- serialhex [] has quit [Remote host closed the connection]20:14
-!- serialhex [] has joined #shogun20:16
-!- Daniel [50fa9238@gateway/web/freenode/ip.] has joined #shogun20:40
DanielHi Sonney20:40
DanielYou had told me to use ./configure --interfaces=libshogun,csharp  to compile the csharp typemap when it was ready20:41
Danielbut I got "Unknown interface: csharp"20:42
Danielthis was on the code where you had added the csharp folder20:43
DanielI emailed you20:48
Danieltake care20:48
-!- Daniel [50fa9238@gateway/web/freenode/ip.] has quit [Quit: Page closed]20:48
-!- blackburn [~qdrgsm@] has joined #shogun21:02
@sonney2kDaniel no I told you to use csharp_modular21:24
@sonney2ksigh too late21:29
blackburnwhat's up, sonney2k?21:29
@sonney2kblackburn, please have a look at the datatype21:30
CIA-90shogun: Soeren Sonnenburg master * ra996050 / src/libshogun/lib/DataType.h : draft definition for vector / matrix datatype -
@sonney2kblackburn, I think we should rename things to SGVector SGMatrix, SGSparseVector, SGSparseMatrix, SGString21:30
blackburnof course, I don't like T*21:30
@sonney2kblackburn, but have a look first21:31
blackburnsonney2k: what you think about implementing operations for it using blas, for example dot product21:32
blackburndo you*21:33
blackburnehhrr my english is becoming worse every day :D21:33
@sonney2kblackburn, makes sense - but we shouldn't over do it - there are many matrix/vector packages around that probably do a better job then us21:35
blackburnanyway it is better way21:36
blackburnbecause we can change all the things just changing implementation of sgvector21:36
@sonney2kon the other hand it is very convenient and more clean - we could just implement the functions that are needed anyways21:36
@sonney2kit will probably make some functions in *features and *math* obsolete21:37
blackburnsonney2k: what about +, * overloading? I think it is a bad practice21:38
@sonney2klets not do that now21:39
blackburnit looks undetermined21:39
blackburnsonney2k: do we really need template matrix type?21:53
blackburnI mean almost all of them are float21:54
@sonney2kblackburn, almost all yes - but not e.g. images22:02
blackburnmay be we could have some shorthand name22:02
blackburnSGFloatMatrix or so22:02
blackburnohhh overloading [] for matrices could be goooood22:06
blackburnz_transposed[dim*j+d] -= feature_vector[d];22:11
-!- serialhex [] has quit [Remote host closed the connection]22:30
-!- serialhex [] has joined #shogun22:32
-!- serialhex [] has quit [Remote host closed the connection]22:34
-!- serialhex [] has joined #shogun22:36
@sonney2kblackburn, overloading [] for 2d is not possible -22:38
blackburnah.. yeah, forgot22:38
@sonney2kI guess you want sth like set   / get functions22:38
blackburnyeap because that X*Y+Z is pretty unreadable if it is not your own code22:39
@sonney2kand it can probably be done more efficiently using some matrix lib like eigen anyways22:39
@sonney2kbtw having SGFloatmatrix etc is not really necessary - typemaps can have type specific implementations if needed22:40
blackburnI don't know which is most efficient, but we should use that one :)22:41
@sonney2kgoing to bed now22:47
blackburngood night22:48
@sonney2kthe mighty diaper master leaves the room22:48
--- Log closed Tue May 10 00:00:52 2011