--- Log opened Sun May 08 00:00:48 2011 | ||
blackburn | it is a sleep time :D | 00:37 |
---|---|---|
blackburn | serialhex: see you | 00:37 |
-!- blackburn [~qdrgsm@188.122.225.203] has quit [Quit: Leaving.] | 00:43 | |
-!- blackburn [~qdrgsm@188.122.225.203] has joined #shogun | 10:18 | |
@mlsec | Large matrices in LLE? The local covariance. Or distances | 10:22 |
@mlsec | Or the weight matrix? | 10:26 |
blackburn | mlsec: local covariance is dim*k | 10:31 |
@mlsec | ok | 10:32 |
blackburn | mlsec: pretty large anyway | 10:32 |
@mlsec | but sparse | 10:32 |
blackburn | yeap | 10:32 |
blackburn | hm.. can we work with sparse matrices effectively? | 10:33 |
@mlsec | why not | 10:33 |
@mlsec | what about the weight matrices? do you apply an eigen decomposion | 10:33 |
@mlsec | as suggested by roweis | 10:33 |
blackburn | mlsec: oh, I didn't started working on eigenproblem | 10:34 |
@mlsec | it could be that for high-dimensional data solving this eigenproblem can be very hard | 10:34 |
blackburn | I see | 10:35 |
@mlsec | i am not sure though. | 10:35 |
blackburn | as for sparse matrices, I don't see any class for that in shogun.. | 10:37 |
@mlsec | hmm. could be that sonneyk did not need a sparse matrix so far | 10:37 |
@mlsec | maybe we can build on some blas and atlas stuff, which hopefully can be linked to shogun (or are already linked) | 10:38 |
@mlsec | this would help in later doing decompositions (svd or eigen) | 10:38 |
blackburn | mlsec: we have SparseFeatures, I will take a look on it | 10:38 |
blackburn | anyway blas and atlas stuff sounds better | 10:39 |
@mlsec | sparse features is likely a sparse vector | 10:39 |
blackburn | at least I can implement it without sparseness and then speed it up | 10:41 |
@mlsec | yes. you could also start with a reference implementation in an interpreted language like python or matlab | 10:42 |
@mlsec | this could help to find the bottlenecks and pitfalls | 10:42 |
blackburn | I already have matlab implementation by roweis | 10:43 |
@mlsec | ok | 10:43 |
blackburn | python is more 'comfortable' for me, but anyway I got comprehension of matlab way to implement it | 10:44 |
blackburn | the only problem is that all the things done with matrices, even distance :) | 10:46 |
blackburn | mlsec: btw, do you know if LLE will work for other distances? | 10:47 |
blackburn | I don't sure it will for some distances | 10:47 |
@mlsec | blackburn: LLE cannot be applied to arbitrary distances | 11:30 |
blackburn | mlsec: only for euclidean? | 11:30 |
@mlsec | blackburn: as part of the embedding the distance between a vector and its reconstruction is phrased in terms of local covariance | 11:30 |
blackburn | yes, I understand it | 11:31 |
@mlsec | such a decomposition of a distance is only possible if the distance is related to a Euclidean distance | 11:31 |
blackburn | yeap, I see | 11:31 |
blackburn | thank you | 11:31 |
@mlsec | in other words, you can apply other distance functions, e.g. Manhattan, Kulback-Leibler, etc, if there is someway to express them as a Euclidean distance | 11:32 |
@mlsec | this is related to kernel functions and has the same underlying concept | 11:32 |
-!- blackburn [~qdrgsm@188.122.225.203] has quit [Quit: Leaving.] | 12:49 | |
-!- serialhex [~quassel@99-101-148-183.lightspeed.wepbfl.sbcglobal.net] has quit [Remote host closed the connection] | 13:12 | |
-!- blackburn [~qdrgsm@188.122.225.203] has joined #shogun | 13:21 | |
-!- blackburn [~qdrgsm@188.122.225.203] has quit [Quit: Leaving.] | 16:05 | |
-!- alesis-novik [~alesis@188.74.87.84] has quit [Quit: I'll be Bach] | 19:21 | |
-!- blackburn [~qdrgsm@188.122.225.203] has joined #shogun | 19:51 | |
-!- alesis-novik [~alesis@188.74.87.84] has joined #shogun | 20:42 | |
-!- warpy [~warpy@bzq-79-182-12-108.red.bezeqint.net] has joined #shogun | 22:38 | |
warpy | anyone awake at this late hour ? | 22:38 |
blackburn | yeap | 22:39 |
warpy | hey man | 22:39 |
warpy | whats your roll in shogun | 22:39 |
warpy | role* | 22:39 |
blackburn | I'm a GSoC student working on dimreduction algos | 22:40 |
warpy | which one are you working on ? | 22:40 |
blackburn | LLE, Hessian LLE, MDS, Isomap, SNE | 22:41 |
warpy | cant say i am familiar with those | 22:41 |
blackburn | me too | 22:42 |
blackburn | :D | 22:42 |
warpy | lol | 22:42 |
warpy | so why these ? | 22:42 |
blackburn | will get familiar gradually | 22:42 |
blackburn | they was proposed by my mentor and I know they are good sometimes | 22:43 |
warpy | who is he ? | 22:43 |
blackburn | Christian Widmer, cwidmer | 22:43 |
blackburn | what about you? :) | 22:44 |
warpy | a guy i work with is trying to do the c# typemap | 22:44 |
warpy | this thing is complex and is driving us nuts | 22:44 |
blackburn | ah, you are mentor :) | 22:45 |
warpy | i am looking for some one to .... consult | 22:45 |
warpy | umm, dont think so | 22:45 |
blackburn | eh.. | 22:45 |
warpy | :) | 22:46 |
warpy | not officially thats for sure | 22:46 |
warpy | and i have zero experience with shogun/swig | 22:46 |
warpy | so ... cant take that honor yet :) | 22:46 |
blackburn | so you work with sploving, right? | 22:46 |
warpy | no | 22:46 |
warpy | Daniel | 22:46 |
blackburn | ahh | 22:46 |
blackburn | you are working not with GSoC student | 22:47 |
blackburn | I see | 22:47 |
warpy | right we missed the gsoc by a day | 22:47 |
warpy | doesnt matter, we need this functionality | 22:47 |
blackburn | I haven't had any experience with C# | 22:47 |
warpy | this work is all about swig typemaps, daniel is doing a good job decrypting the documentations | 22:49 |
blackburn | I have some issues with python swig | 22:49 |
warpy | what ? | 22:50 |
blackburn | had* | 22:50 |
warpy | ..explain | 22:50 |
blackburn | just little misunderstanding with some function returning numpy array | 22:50 |
warpy | so no bugs in it ? | 22:51 |
blackburn | but haven't had any serious experience, just know a little about current typemaps | 22:51 |
blackburn | I don't know, I just used already implemented typemaps, the problem was only with its usage | 22:52 |
warpy | cool | 22:54 |
* sonney2k is not aware of any bugs in python_modular and this is currently the interface I use most | 22:57 | |
warpy | hey sonney2k | 22:57 |
blackburn | wow sonney2k is here :D | 22:57 |
warpy | daniel is almost finished rewriting the java modular into c# | 22:57 |
warpy | without testing/debugging but | 22:57 |
warpy | is there someone who is an expert on this | 22:58 |
@sonney2k | blackburn, 50% here... | 22:58 |
warpy | that can provide help , if we need to consult | 22:58 |
@sonney2k | warpy, java_modular is still not really written... | 22:58 |
warpy | but its a start | 22:58 |
@sonney2k | so what did he port? | 22:58 |
@sonney2k | ahh ok | 22:58 |
warpy | most of the code in the java file | 22:59 |
@sonney2k | warpy, daniel should try to get vectors translated first | 22:59 |
warpy | single arrays were in there | 22:59 |
@sonney2k | warpy, did you get c# to work at all? | 22:59 |
warpy | the interface ? | 22:59 |
@sonney2k | I mean did you manage to start the resulting .exe? | 23:00 |
warpy | you should ask him, i know small amounts of detail | 23:00 |
warpy | not yet as far as i know | 23:00 |
@sonney2k | (the first step IMHO) | 23:00 |
@sonney2k | I am here sometimes but have to leave usually very unexpectedly (at least currently) | 23:01 |
warpy | do you have experience with swig, hands on stuff ? | 23:01 |
@sonney2k | warpy, I wrote all the interfaces + typemaps that are currently there. | 23:03 |
warpy | well, we would definitely can use some help. | 23:04 |
@sonney2k | well ask | 23:04 |
warpy | he is working on it. i am just following | 23:05 |
blackburn | sonney2k: have you any idea how I can work with sparse matrices in LLE? | 23:05 |
@sonney2k | warpy, then tell him to ask me :) | 23:06 |
warpy | i will let him know that he can contact you regarding this, i bet he have some unknowns. | 23:06 |
warpy | its a steep learning curve, complex stuff | 23:06 |
warpy | btw whe you said vectors | 23:07 |
@sonney2k | warpy, not so difficult if you have examples - and we have plenty in other languages | 23:07 |
@sonney2k | blackburn, use CSparseFeatures ... | 23:07 |
warpy | you were talking about arrays ? | 23:07 |
@sonney2k | for me this is the same thing | 23:07 |
warpy | well the jave thing have arrays | 23:07 |
blackburn | sonney2k: eigendecomposition..? | 23:07 |
warpy | so hopefully this will be a start | 23:08 |
warpy | also c# is not documented in swig | 23:08 |
warpy | which is a pain | 23:08 |
blackburn | in ass | 23:08 |
blackburn | :D | 23:08 |
@sonney2k | blackburn, you should check some other matrix packages like eigen or so - nothing like this is core shogun focus... | 23:09 |
blackburn | sonney2k: I will need two 'special' things with sparses - dot product (it is in SparseFeatures) and eigenproblem solving | 23:10 |
blackburn | may be just do it without any sparseness could be a good idea? | 23:10 |
blackburn | then latter add it | 23:10 |
warpy | night :) | 23:11 |
@sonney2k | blackburn, I think so... | 23:11 |
@sonney2k | warpy, cu | 23:11 |
@bettyboo | bye sonney2k | 23:11 |
@sonney2k | bettyboo, :P | 23:11 |
@bettyboo | sonney2k: not exactly perfect but intelligible :P | 23:11 |
-!- warpy [~warpy@bzq-79-182-12-108.red.bezeqint.net] has quit [Ping timeout: 252 seconds] | 23:15 | |
@sonney2k | blackburn, I woudl really do the simplest possbile things first | 23:20 |
@sonney2k | then more advanced stuff | 23:20 |
blackburn | yeah I think the same | 23:21 |
@sonney2k | otherwise you can easily become stuck and get nowhere | 23:21 |
@sonney2k | blackburn, you were asking sth yesterday but I had to leave/got disconnected - so didn't see the question... | 23:32 |
blackburn | ah | 23:32 |
blackburn | sonney2k: I think these messages | 23:33 |
blackburn | blackburnsonney2k: we discussed possibility of writing a paper about dimreduction with shogun to some journal21:26blackburnsonney2k: he said it is possible with "Open Research Computation"21:27blackburnbut while I was reading FAQ, I became surprised a lot with "We charge a flat fee of £995/$1660/€1120 for each accepted manuscript."21:28blackburnsonney2k: is it a common practice?21:28blackburnI'm a bit newbie with these issues | 23:33 |
blackburn | jh | 23:33 |
blackburn | eeeeeeee | 23:33 |
blackburn | reh | 23:33 |
blackburn | blackburnsonney2k: we discussed possibility of writing a paper about dimreduction with shogun to some journal21:26 | 23:33 |
blackburn | blackburnsonney2k: he said it is possible with "Open Research Computation"21:27 | 23:33 |
blackburn | blackburnbut while I was reading FAQ, I became surprised a lot with "We charge a flat fee of £995/$1660/€1120 for each accepted manuscript."21:28 | 23:33 |
blackburn | blackburnsonney2k: is it a common practice?21:28 | 23:33 |
blackburn | blackburnI'm a bit newbie with these issues | 23:33 |
@sonney2k | blackburn, common practise in open access journals... | 23:34 |
blackburn | why they charge a fee? | 23:35 |
@sonney2k | usually when you apply for a grant it is covered in there... | 23:35 |
blackburn | ah | 23:35 |
@sonney2k | blackburn, to cover their costs - it is free for all afterwards under some CC-by* license | 23:35 |
blackburn | I see | 23:36 |
blackburn | so there is no way to have a publication about shogun dimreduction without fees, right? | 23:36 |
@sonney2k | blackburn, you didn't fix the calcroc issue in GUIMath.cpp right? | 23:36 |
@sonney2k | blackburn, there could be - jmlr for example | 23:37 |
blackburn | sonney2k: eh.. I asked you to comment lines, I didn't wanted to create a new pull request for that | 23:37 |
@sonney2k | but standards are set very high ... | 23:37 |
@sonney2k | blackburn, I thought you had some idea how to fix things. anyway deleting the offending lines now. | 23:38 |
blackburn | I don't know, there are many things I don't understand how to fix | 23:38 |
@sonney2k | deleted | 23:39 |
@sonney2k | blackburn, which others? | 23:40 |
blackburn | I mean that function in GUIMath doing some things I don't understand :) | 23:40 |
blackburn | you said that nobody uses libshogunui so I decided to not think about it :) | 23:41 |
blackburn | JMLR is pretty serious.. I don't sure I can do anything for it | 23:43 |
blackburn | sonney2k: well may be I should fix it for ROCEvaluation usage? | 23:47 |
@sonney2k | blackburn, not nobody... just not myself | 23:47 |
@sonney2k | (others complained already that git trunk doesn't compile) | 23:47 |
blackburn | I have involuntary rejection of libshogunui :D | 23:49 |
@sonney2k | blackburn, I wish this could die ... | 23:51 |
@sonney2k | I mean all the static interfaces - they are just *pff* | 23:52 |
blackburn | I feel the same :) | 23:52 |
@sonney2k | it is only that we have no matlab support in swig... | 23:56 |
blackburn | is it possible? | 23:57 |
blackburn | * to have one | 23:57 |
@sonney2k | yes - but I don't know about swig's internals nor do I have the time to write it...... | 23:59 |
--- Log closed Mon May 09 00:00:50 2011 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!