Open in new window / Try shogun cloud
--- Log opened Thu Mar 08 00:00:06 2012
--- Day changed Thu Mar 08 2012
n4nd0I am really glad I discovered it, I think it can become very useful once I start working on my thesis :)00:00
n4nd0sonney2k: anyway, I will check again what can I apply for in gsoc and maybe we can bring again the topic of the documentation after summer if everything goes well00:01
@sonney2kn4nd0, of course no one hinders you to work on this - to the contrary :)00:02
n4nd0sonney2k: what is your opinion about making more than one gsoc application for different projects within shogun?00:04
@sonney2kI guess more than 2 doesn't make sense00:05
@sonney2kn4nd0, of course you can propose anything you'd like to do - this list is not exhaustive00:11
n4nd0sonney2k: the more I am discovering about SO, the more I like it :)00:12
n4nd0sonney2k: I will give a try to it, and maybe something easier like the GPs just in case00:13
n4nd0but SO looks pretty cool00:14
n4nd0there is nothing in shogun implemented yet, right?00:14
@sonney2kn4nd0, I would say easiest task is the newton svm one - just port a few lines of matlab code00:19
@sonney2kSO - it is about the framework - so indeed nothing like that exists though we have the argmax already in there00:19
n4nd0sonney2k: btw, how does the new documentation get included in the official doxygen in the web?00:30
n4nd0I have seen that there is nothing about Mahalanobis yet and I got curious00:30
@sonney2kI guess I should automate this - whenever all tests go well, run make doc and put it live00:31
@sonney2kall of that is prepared but not done :(00:32
n4nd0so there is a testing framework that checks if the things that are about to enter in the project work well?00:32
@sonney2kn4nd0, there is the examples/undocumented/python_modular directory (for example)00:37
@sonney2kall things in there are examples illustrating things for the user but they are also used as tests00:37
@sonney2kso there is a running these examples and dumping the output they produce00:38
@sonney2kand there is a that checks if results are the same00:38
n4nd0but what are the results compared to?00:38
@sonney2kproblem is that this is not 100% enabled for everything00:38
n4nd0so for example, right now for the Mahalanobis distance stuff00:39
@sonney2kn4nd0, nothing - it is not activated00:39
n4nd0aha ok00:39
@sonney2kyou would have to run generator.py00:39
@sonney2kthat would then store the output of mahalanobis distance stuff00:39
@sonney2kand so tester could later compare it00:39
@sonney2kI think this would also be an excellent gsoc task - but again I don't know about googles official policy about such endeavors00:40
@sonney2kin any case writing just 1 example and getting regression tests for free with it is the most painless way I can think of for tests :)00:41
n4nd0I was thinking that maybe the tester compared the results with the ones obtained by other software or sth like that00:42
@sonney2kn4nd0, no - that is something we do of course when implementing $METHOD00:43
@sonney2kbut from then on it is not needed00:43
@sonney2kanyways bed time00:44
n4nd0good night00:45
n4nd0wiking: hey!00:47
n4nd0wiking: are you thinking also of applying for gsoc?00:47
n4nd0wiking: if you don't mind that I ask of course ;)00:47
-!- n4nd0 [] has quit [Ping timeout: 246 seconds]02:24
-!- wiking_ [~wiking@2002:4e17:bfc9:1:fc01:a93b:44d8:ae33] has joined #shogun02:34
-!- wiking_ [~wiking@2002:4e17:bfc9:1:fc01:a93b:44d8:ae33] has quit [Changing host]02:34
-!- wiking_ [~wiking@huwico/staff/wiking] has joined #shogun02:34
-!- wiking [~wiking@huwico/staff/wiking] has quit [Ping timeout: 245 seconds]02:35
-!- wiking_ is now known as wiking02:35
-!- in3xes [~in3xes@] has joined #shogun02:46
-!- in3xes [~in3xes@] has quit [Quit: Leaving]03:09
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]03:51
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun07:31
-!- n4nd0 [] has joined #shogun07:40
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]08:14
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun08:16
wikingn4nd0: hey mate! sorry yesterday i was already sleeping....10:34
wikingn4nd0: yeah i'm planning to apply for gsoc10:34
wikingit's not only a matter of which course to take... but most probably it's going to be the latent-svm implementation within shogun10:34
n4nd0wiking: no problem, it was already late actually :)10:47
n4nd0wiking: ok, I have never heard of them apart from your conversations around here with blackburn10:48
n4nd0completely ignorant in that topic :S10:48
wikingheheh no worries :P10:54
-!- n4nd0 [] has quit [Ping timeout: 245 seconds]11:01
-!- conor [] has joined #shogun11:29
conorhow can i compile shogun with an old compiler? there is this problem that matlab interfaces must be compile with g++-4.3 and i can't find where to change it in the configure or Makefile.11:30
sonne|work./configure --help :)11:37
-!- n4nd0 [] has joined #shogun11:42
-!- vikram360 [~vikram360@] has quit [Read error: Connection reset by peer]12:51
conorsonne|work: thanks that was helpful :)13:31
-!- vikram360 [~vikram360@] has joined #shogun13:34
conori have a weird problem. i'm used an older shogun version with the LIBLINEAR_L1R_L2LOSS_SVC classifier. there everything worked fine, but now i use the recent shogun version and use LIBLINEAR_L1R_L2LOSS_SVC again. But now i have to transpose the features and the results for w are all NaN. Does anyone know what this could be? I searched the commit messages for LIBLINEAR_L1R_L2LOSS_SVC but couldn't find anything so i don't know what changed there.13:35
sonne|workconor: you could help figuring out when this problem happened by doing a git bisect and determining the commit which caused this13:43
conorok i will do this13:43
sonne|workbtw you can do git log shogun/classifier/svm/SVM_linear.cpp13:44
sonne|workor shogun/classifier/svm/LibLinear.cpp to see the changes for these files only13:44
sonne|work(I guess I broke it ...)13:45
-!- conor [] has quit [Read error: Operation timed out]13:46
-!- blackburn [~qdrgsm@] has joined #shogun13:57
CIA-64shogun: Sergey Lisitsyn master * rdb96f7c / (3 files): Added bias usage capability for crammer-singer liblinear -
-!- vikram360 [~vikram360@] has quit [Ping timeout: 245 seconds]14:31
-!- vikram360 [~vikram360@] has joined #shogun14:31
-!- cronor [~cronor@] has joined #shogun14:31
cronorsonne|work: i looked into that but is there a way to access the pre-May2011 logs?14:41
sonne|workall logs since 2006 are in there14:42
sonne|workahh but I probably moved files around14:43
sonne|workbefore that data the dir shogun/ did probably not exist14:44
n4nd0sonne|work: hey! one question about dynamic memory allocation, if I want to create an array of SGVector14:55
n4nd0SG_MALLOC(SGVector< float64_t >, nvecs)14:56
n4nd0and later loop for each vector and do14:56
n4nd0SG_MALLOC(float64_t, nelems)14:56
n4nd0is that the right way to do it?14:56
n4nd0I have doubt about the second MALLOC, to allocate space for each of the vectors14:57
blackburnn4nd0: you may use array_of_vectors[i] = SGVector<float64_t>(nelems)14:58
n4nd0n4nd0: hey! I didn't see you are around :)14:59
n4nd0ups I talked to me15:00
n4nd0blackburn: thank you very much :)15:00
blackburnsonne|work: +0.01% of accuracy for adding bias LOL15:00
blackburn(I've added bias support for crammer-singer liblinear)15:00
wikingblackburn: yoyo15:01
blackburnwiking: hey15:01
wikingmmm i wonder where's alex :)15:01
wikingi supposed to have a Skype conf with him now :P15:01
blackburnlets shout ALEX and wait him to come ;)15:01
wikingmaybe he won't be able to decode the irc protocoll15:01
n4nd0blackburn: what about later to free that stuff, would SG_FREE(array_of_vectors) do the trick without having to call free_vector for every one of them?15:03
blackburnn4nd0: it won't, you would have to free it15:03
n4nd0n4nd0: ook, thank you again :)15:04
blackburnblackburn: no problem15:04
CIA-64shogun: Sergey Lisitsyn master * r0c71dba / src/shogun/classifier/svm/MulticlassLibLinear.cpp : Added bias usage capability for crammer-singer liblinear -
blackburnwiking: did you try multiclassliblinear already?15:06
wikinghaven't had time yet :(15:06
wikingsorry man15:06
blackburnno need to sorry, I mean you were waiting for it :)15:06
-!- cronor [~cronor@] has quit [Quit: cronor]15:11
wikingheheh he has shown up \o/15:19
wikingsonney2k: here?15:37
sonne|workwiking: yes15:47
sonne|workwhats up?15:47
vikram360sonney2k : There seems to be no information about how or why C5.0 is better than C4.5. Is there any paper in particular or any resource that you know about that elicits the superiority?16:42
-!- n4nd0 [] has quit [Ping timeout: 244 seconds]16:53
@sonney2kblackburn, how did you add the bias term to multiclass liblinear?17:02
blackburnsonney2k: I'd say virtually17:03
blackburnw = [w, b] , x = [x, 1]17:03
@sonney2kso you added some constant?17:03
@sonney2kwithout modifying features?17:03
blackburnattached b to w and 1 to x17:03
blackburnsonney2k: just replaced dotproducts in MCSVM solver17:04
@sonney2kthat is one of the problems that they modify data from time to time17:04
blackburnpretty useless though17:04
@sonney2kIIRC L1 svm had that problem17:04
@sonney2kdepends on the problem17:05
blackburnsonney2k: how can it be true that bias is so useless?17:05
@sonney2kI have to leave the train...17:05
blackburnI never thought it is not really required for high-dim tasks17:05
-!- cronor [~cronor@] has joined #shogun18:11
-!- vikram360 [~vikram360@] has quit [Read error: Connection reset by peer]18:18
-!- vikram360 [~vikram360@] has joined #shogun18:28
@sonney2kblackburn, well it is a *regularized* bias so not really the the one you have in standard SVM18:41
@sonney2kso one more dim more doesn't really matter much18:41
blackburnsonney2k: I guess it is the one vojtech did in his 2002 paper18:41
blackburnfor kernel bsvm18:41
@sonney2kI haven't read his paper18:41
@sonney2kbut almost everyone does this 'hack'18:42
@sonney2kocas / liblinear ...18:42
blackburnsonney2k: min (||w||^2+b^2) + C \sum \xi18:42
blackburnsonney2k: I'm nearly finished with mc ocas!18:42
blackburntook a little while18:42
@sonney2kvery good18:43
@sonney2know we only need to get good old ocas back to work :)18:43
blackburnsonney2k: yeah probably we need to18:44
blackburnsonney2k: what do you think about that idea you suggested?18:44
@sonney2kthere is certainly enough other stuff too... ohh well18:44
@sonney2kwell I would say we don't have to rush18:45
blackburndeadline is coming, but ideas are not concerned I guess?18:45
@sonney2khave to leave train now18:45
blackburndamn do you move in train whole day long? ;)18:46
blackburndo you work in siberia? :)18:46
@sonney2kwell I think one can even do spontaneous things18:51
@sonney2kso if some student has some nice idea that he wants to persue - that woudl be possible as well18:52
@sonney2kbut I certainly would love tighther python integration18:52
@sonney2kin the end that is the interface I use mostly18:53
blackburnsonney2k: others are pretty useless for me as well18:59
@sonney2knot useless - I but I just use them very rarely19:00
blackburndo you use it? :)19:01
blackburnI have never used neither java nor ruby :)19:01
blackburnsonney2k: btw java integration is pretty interesting thing as well19:02
-!- n4nd0 [] has joined #shogun19:02
blackburnsonney2k: directory arrangement is still crappy unclear for me :(19:11
-!- cronor [~cronor@] has quit [Quit: cronor]19:30
-!- cronor [~cronor@] has joined #shogun19:32
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]19:41
-!- cronor [~cronor@] has quit [Quit: cronor]20:03
blackburnsonney2k: done with mc ocas - got 21.6% accuracy :D20:05
blackburngoing to diff libocas20:05
-!- cronor [~cronor@] has joined #shogun20:18
-!- cronor [~cronor@] has quit [Quit: cronor]20:24
blackburnah I see, maxtraintime :)20:54
blackburnsonney2k: btw I checked libocas and libqp - all looks ok20:54
blackburnI do not understand why you ->ed it to int32_t/float64_t, but everything else is ok20:55
-!- n4nd0 [] has quit [Ping timeout: 246 seconds]21:05
-!- cronor [] has joined #shogun21:21
-!- n4nd0 [] has joined #shogun21:39
CIA-64shogun: Sergey Lisitsyn master * rec86ec8 / (2 files): Updated multiclass machine base class -
CIA-64shogun: Sergey Lisitsyn master * r229943e / src/shogun/classifier/svm/libocas.cpp : Fixes for MC OCAS -
CIA-64shogun: Sergey Lisitsyn master * r7e1d354 / (5 files in 3 dirs): Introduced multiclass OCAS -
-!- n4nd0 [] has quit [Read error: Operation timed out]22:55
-!- wiking [] has joined #shogun23:07
-!- wiking [] has quit [Changing host]23:07
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun23:07
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]23:15
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun23:15
-!- wiking [~wiking@huwico/staff/wiking] has quit [Ping timeout: 260 seconds]23:20
-!- blackburn [~qdrgsm@] has quit [Ping timeout: 276 seconds]23:22
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun23:36
--- Log closed Fri Mar 09 00:00:19 2012