Open in new window / Try shogun cloud
--- Log opened Sat Jul 09 00:00:57 2011
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]00:18
@sonney2kf-x, yes?00:21
@sonney2kI am too sleepy now g'night.00:23
f-xokay, see ya!00:29
-!- f-x [~user@] has quit [Remote host closed the connection]02:12
-!- in3xes_ [~in3xes@] has joined #shogun05:22
-!- in3xes [~in3xes@] has quit [Read error: Operation timed out]05:24
-!- blackburn [~blackburn@] has joined #shogun10:44
-!- blackburn [~blackburn@] has quit [Read error: Connection reset by peer]11:58
-!- blackburn [~blackburn@] has joined #shogun12:03
-!- sploving1 [~sploving@2001:cc0:2020:2022:5eff:35ff:fe04:f091] has joined #shogun13:15
-!- sploving1 [~sploving@2001:cc0:2020:2022:5eff:35ff:fe04:f091] has quit [Ping timeout: 260 seconds]13:40
-!- f-x [~user@] has joined #shogun13:49
-!- f-x [~user@] has quit [Remote host closed the connection]13:59
-!- sploving1 [~sploving@] has joined #shogun14:07
-!- in3xes [~in3xes@] has joined #shogun14:10
-!- f-x [~user@] has joined #shogun14:40
-!- in3xes [~in3xes@] has quit [Quit: Ooops..!!]15:01
-!- in3xes [~in3xes@] has joined #shogun15:01
-!- in3xes [~in3xes@] has quit [Client Quit]15:02
-!- in3xes [~in3xes@] has joined #shogun15:05
-!- in3xes [~in3xes@] has quit [Remote host closed the connection]15:11
-!- in3xes [~in3xes@] has joined #shogun15:13
sploving1parameter_list.add(Arrays.asList(new Double(0.9), new Double(1e-3), new Integer(1)));15:27
sploving1why int num_threads = ((Integer)para.get(2)).intValue(); said: Exception in thread "main" java.lang.ArrayIndexOutOfBoundsException: 215:27
sploving1blackburn, do you know what is the matter?15:28
blackburnsploving1: no, looks strange15:30
sploving1blackburn, I got know. my mistake15:31
sploving1blackburn, svm.parallel.set_num_threads(num_threads);15:32
sploving1cannot find variable parallel15:32
blackburnno idea15:33
-!- in3xes [~in3xes@] has quit [Quit: Ooops..!!]15:45
sploving1what does this error mean? [ERROR] Not a two class labeling - no positively labeled examples found15:54
sploving1Exception in thread "main" java.lang.RuntimeException: [ERROR] Not a two class labeling - no positively labeled examples found15:54
sploving1blackburn, sonney2k,15:54
blackburnexactly what it says - you have olny one class labelling15:55
sploving1blackburn, svm.set_linear_term(new DoubleMatrix(new double[] {-1,-2,-3,-4,-5,-6,-7,-8,-7,-6}));15:55
sploving1how to fix the above statement?15:55
blackburncheck labels passed to svm15:56
sploving1blackburn, double label_train_dna[] = {-1,-1,-1,-1,-1,1,1,1,1,1};15:56
sploving1Labels labels = new Labels(new DoubleMatrix(label_train_dna));15:56
sploving1maybe this statement is wrong15:56
blackburnit should be ok15:56
sploving1so strange15:57
sploving1blackturn, i know. DoubleMatrix(new double[] {value1, value2, ...})Constructs a column vector.15:58
blackburnmake a row :)15:59
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]16:18
-!- blackburn [~blackburn@] has joined #shogun16:21
-!- f-x [~user@] has quit [Remote host closed the connection]16:25
-!- Guest587 [~mobile@] has joined #shogun16:33
-!- in3xes_ is now known as in3xes16:53
sploving1hello Math_init_random(17) in python, then in java, how to use it?17:13
-!- sploving1 [~sploving@] has left #shogun []17:38
-!- Guest587 [~mobile@] has quit [Ping timeout: 240 seconds]19:16
CIA-32shogun: Baozeng Ding master * re34a3b8 / (3 files): add remaining classifier java examples -
CIA-32shogun: Baozeng Ding master * r00b9a3c / (2 files): add clustering java examples -
CIA-32shogun: Baozeng Ding master * r53b6f91 / examples/undocumented/java_modular/ : add distribution_histogram_modular java example -
CIA-32shogun: Soeren Sonnenburg master * r08b175b / (6 files):20:35
CIA-32shogun: Merge pull request #184 from sploving/master20:35
CIA-32shogun: add remaining classifier java examples -
blackburnsonney2k: I came with new task for me: static memory relaxed fibonacci heap :(20:37
@sonney2kblackburn, you dont' like the easy things do you ?20:37
blackburnsonney2k: it could be the fastest isomap in shogun ;)20:38
blackburneven for really large datasets it could act well in theory20:39
blackburnsth like O(n^2 log 1/n)20:40
@sonney2kthat's the spirit20:40
blackburnfrom next week work on cover tree porting should be started20:42
@sonney2kall very welcome20:42
blackburnchris will be grandmentor (just like father becomes grandfather)20:42
blackburncause I will be some kind of mentor for guy that will try to port cover tree20:42
@sonney2kwill that turn me into grand-grand-mentor or grand-admin?20:43
blackburnSoeren the godfather I guess20:45
blackburnwhy not to play some mafia here :D20:45
@sonney2kblackburn, I prefer benevolent dictator - but hey I am no longer alone :)22:22
blackburnsonney2k: who are? :)22:22
@sonney2krumor has it - that some mad vodka drinking Russian is launching nuclear missiles too22:28
blackburnsonney2k: I'm sure you are ;)22:29
* sonney2k wonders if blackburn got replaced by some bot22:30
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]22:43
-!- blackburn [~blackburn@] has joined #shogun22:43
-!- f-x [~user@] has joined #shogun22:47
@sonney2kf-x, hi22:48
f-xsonney2k: hey!22:48
@sonney2kI am more alive today22:48
f-xI too just got back now, read your mail22:48
f-xhave you observed this kind of behaviour with that dataset before? i mean performance decreases with the last 100k or so examples22:49
f-xi'm pretty sure the online SGD and the original are essentially identical, except for lambda and epochs22:50
@sonney2kf-x, yeah but I might have made a mistake translating it22:50
@sonney2kf-x, so I would suggest to use liblinear22:50
@sonney2kfor liblinear we have guarantees22:50
f-xsonney2k: you mean a version of liblinear using the new framework?22:51
@sonney2kf-x, I meant just the batch version22:51
@sonney2kf-x, there is a script on the website to convert the data files into svmlight format22:52
@sonney2kwhich sparse features can directly read22:52
f-xsonney2k: ok thanks, i'll use that. but my computer can't take like more than 1 GB data at once22:53
@sonney2kf-x, yeah then just use a subset22:53
f-xso I think the examples will be in the 100k-200k range22:53
f-xand testing on say 10k examples which were not used for training?22:53
@sonney2kf-x, you should get similar results with liblinear22:54
@sonney2kf-x, I would test on 100k examples too22:54
f-xsonney2k: ah. okay.. btw how was testing done for the alpha dataset?22:54
f-xdid you have some extra labelled data with you?22:55
@sonney2kf-x, labels are non-public22:55
@sonney2kso use say first 100-200k examples for traiing22:55
@sonney2kand the next 100k examples for testing22:55
f-xwas just asking :) sure..22:56
@sonney2kf-x, but shouldn't make a difference22:56
f-xsonney2k: yeah.. i'll get on with that.. anything else you think i should be working on in the meantime?22:57
@sonney2kf-x, I mean the next steps would be to integrate liblinear/sgd-qn/vw22:58
@sonney2kof course getting this benchmark to work and speed optimizations22:58
@sonney2kf-x, sgd-qn should be trivial to do for you (very similar to sgd)22:59
f-xyes - i browsed through the code, and they looked pretty similar22:59
@sonney2kf-x, regarding liblinear I would only translate the CLibLinear::solve_l2r_l1l2_svc function23:00
-!- phrhbk [] has quit [Quit: Leaving]23:00
@sonney2kand this time provide an OnlineLibLinear in addition to the batch variant23:01
f-xsonney2k: yes.. just like we did for sgd23:01
@sonney2kf-x, for sgd we will only keep the online variant at some point23:02
@sonney2kbut liblinear has many solvers and the svm one is sufficient to us23:02
f-xokay.. for now i'll do the testing with liblinear and begin with OnlineLibLinear23:03
@sonney2kyeah these online learners won't be a lot of work - but it will be more interesting and the benchmarking will actually tell us whether things are correct and what we can improve23:05
blackburnanother day another wrapper23:10
blackburnsonney2k: grant me rights to execute CustomDistance ;)23:26
@sonney2kblackburn, ?23:26
blackburnsonney2k: I use it in landmark mds - I pass landmark distance matrix using it - but it creates a copy of matrix23:26
blackburnbecause it is only float32_t and I use float64_t23:26
@sonney2kblackburn, of course it is float32_t23:27
@sonney2kit tries to be efficient23:27
blackburnand produces necessarity of copy matrix..23:27
@sonney2kI mean when you have to store the matrix you should try to be23:27
@sonney2kwell then don't get the copy but the float32_t version23:27
blackburndistance always returns float64_t?23:28
blackburnbtw I guess I should use triangular matrices23:29
-!- f-x [~user@] has left #shogun ["ERC Version 5.3 (IRC client for Emacs)"]23:30
-!- f-x [~user@] has joined #shogun23:30
blackburnshit I've got the most beautiful embedding I ever seen23:31
@sonney2kblackburn, yes distcance returns float64_t  /scalars/23:40
@sonney2kbut the matrix itself is squeezed as float32...23:40
blackburndon't like it a bit23:40
@sonney2kI guess I will never understand23:41
@sonney2kwhat your problem with this is23:42
blackburnI get the whole matrix23:42
@sonney2kjust get the float32 one23:42
blackburnhow to get float32 one?23:42
@sonney2kblackburn, I see - it is not templated like in Kernel.h23:43
@sonney2kbasically, the distance class evolved from the kernel class23:44
@sonney2kthe kernel class has that function23:44
@sonney2kthe distance class has not23:44
@sonney2kand I see a new problem23:45
@sonney2kwe do return some SGMatrix<xx>23:45
blackburnI have found a way to slow down LLE 20x23:46
@sonney2kbut is it possbile to have two functions with same name and no input args but different output args?23:46
blackburnI guess :)23:46
blackburnhow can compiler know what do you want? :)23:47
@sonney2kblackburn, it could - I mean it knows the output type23:47
@sonney2kand that should match23:47
blackburnehh. if you call foo() how can compiler know you want int foo()?23:49
@sonney2kblackburn, well if you do int x = foo() it is clear23:54
blackburnsonney2k: I'm not sure it is clear for gcc ;)23:55
@sonney2kblackburn, the problem is that one can drop return arguments23:55
@sonney2ki.e. call a function ignoring the result23:55
blackburnget distance matrix and lost result23:56
blackburnhmm pretty way to get garbage23:56
--- Log closed Sun Jul 10 00:00:59 2011