Open in new window / Try shogun cloud
--- Log opened Tue Mar 06 00:00:19 2012
@sonney2kn4nd0, IIRC the regularization term just adds sth to the diagonal of the cov matrix00:06
@sonney2kso it penalizes non-zero weights in w00:08
@sonney2kn4nd0, I would say have a look at QDA in scikits learn and follow the reference above00:09
@sonney2kn4nd0, btw it would make a lot of sense to create one graphical example illustrating lda/qda/ lda on polyfeatures00:09
@sonney2kso we see how they differ :)00:10
@sonney2kat least I would love to see that00:10
@sonney2kso yes QDA is of interest especially when just running LDA on PolyFeatures00:11
@sonney2kfor comparison00:12
n4nd0sonney2k: I have been studying the code from scikits before, I think I got ir more or less00:14
n4nd0I have not seen the direct relation to the theory I read though00:15
n4nd0they use SVD to compute the covariances what I found surprising at first sight00:16
n4nd0would you like to have sth graphical like this?00:16
-!- blackburn [~qdrgsm@] has joined #shogun00:17
@sonney2kn4nd0, yes sth like this00:18
n4nd0sonney2k: I will work on that then :)00:19
@sonney2kn4nd0, do they use SVD for computing the inverse?00:19
@sonney2kpseudo inverse I mean00:20
n4nd0sonney2k: mmm I am not sure, by the name of their variables it looks like they compute the covariance, not the inverse00:20
blackburneigs of XX' are left singular vectors00:21
blackburneigvectors I mean00:22
blackburndamn are you crazy guys? pretty late00:22
n4nd0it is this fragment of code where they do it
n4nd0blackburn: haha my gf was sick today so I have been wandering around home almost the whole day and didn't work that much :P00:23
blackburnoh let she recover soon00:23
@sonney2kn4nd0, we compute pseudo inverse for LDA too via SVD00:24
@sonney2k/The pseudo inverse A+ can be constructed from the singular value00:24
@sonney2k/decomposition A = UDV^T , by  A^+ = V(D+)U^T.00:24
@sonney2k(fromt he code)00:24
@sonney2kwhat blackburn said00:24
n4nd0sonney2k: ok00:25
blackburnI had some article on pseudoinverse00:25
blackburnsonney2k: what have I said?00:25
@sonney2kblackburn, seen that ?00:25
n4nd0sonney2k: in any case I think that the whole point of computing the covariance and its inverse in QDA is to compute the Mahalanobis distance00:25
n4nd0sonney2k: so I thought that maybe we could just use MahalanobisDistance for that00:26
blackburnsonney2k: hah crazy00:26
@sonney2kblackburn, on #gsoc I had a chat with carol smith (gsoc organizer) and indeed the tutorial task seems to be out of scope for gsoc00:29
blackburnsonney2k: damn..00:30
@sonney2kI asked her if we can rescue this part somehow, i.e. by modifying it to only have code parts in there like: write/polish examples, enable test suite based on these examples, fix code to have fewer ASSERT's etc00:31
@sonney2klets see if that is possible00:31
n4nd0ups ... not good00:32
-!- blackburn [~qdrgsm@] has quit [Ping timeout: 265 seconds]00:47
-!- blackburn [d55780cc@gateway/web/freenode/ip.] has joined #shogun00:48
blackburnsonney2k: still there?00:48
blackburnn4nd0: still there?00:51
n4nd0blackburn: yes00:51
n4nd0blackburn: I drank coffee :P00:51
blackburntoo much?00:51
n4nd0blackburn: haha just a small cup, I am don't drink it usually so ...00:52
blackburnn4nd0: if you find qda too messy we may come up with some other idea?00:53
n4nd0n4nd0: I think it can be ok, I must learn to do new stuff so it is ok even if is a bit difficult to get it at the beginning00:55
n4nd0blackburn: or do you say because you would like it to take it yourself?00:55
blackburnn4nd0: no, I don't want to for sure :)00:55
blackburnI'm working on mc ocas this week probably ;)00:55
n4nd0blackburn: ok, then I will continue with it00:56
n4nd0blackburn: I have a couple of exams next week so I might work a bit slow but I won't leave it00:56
blackburnI think I can suggest you to implement some dim reduction algorithm00:56
blackburnno problem just keep in touch00:57
blackburnso you won't have any exams in may or later?00:57
n4nd0blackburn: sure, tell me about that dim reduction if you have thought of it already00:57
n4nd0blackburn: yes I will :S00:57
blackburnmay be t-SNE00:57
n4nd0I have to take a couple of exams from Spain in late April early00:58
blackburnI will have some in may as well00:58
blackburnand bachelor's work defence (?)00:58
n4nd0I did a bit of mess with my studies00:58
n4nd0I came last year in Erasmus to KTH, in my 4th year of university00:58
n4nd0in Spain the studies are normally 5 years long so I had still one year left after my Erasmus00:59
blackburnI see00:59
n4nd0but I got the opportunity of staying for a master at KTH ... and I didn't want to refuse it00:59
n4nd0so here I am, doing the master and taking courses from my degree in Spain00:59
blackburnso you are a msc student?01:00
n4nd0but I don't have a real bachelor01:00
blackburnI did not get even bachelor yet ;)01:00
n4nd0in Spain it is 5 years so it is like if we get bachelor + msc when we graduate01:00
blackburnso you are probably older than me :)01:00
n4nd0I say it is like because from my point of view it is not like that01:00
n4nd0I am 2201:01
blackburnyeah, I'm 20 currently ;)01:01
n4nd0I am from November 198901:01
n4nd0you go pretty well actually01:01
blackburnbtw we have had similar system for a while01:02
blackburnabout 5 year01:02
n4nd0I don't like ... it sucks01:02
blackburnbut now we have bachelors too01:02
blackburnand I'm going to be one of01:02
n4nd0it if 5 years with really broad studies ... we don't really learn to do useful stuff of anything01:02
n4nd0we learn a bit of everything and master nothing :(01:03
blackburnis it better in swe?01:03
n4nd0I like it more yes01:03
n4nd0from my point of view in Spain the hard part is to pass the exams01:04
n4nd0but the labs are really easy normally01:04
n4nd0here the exams are quite ok, they test that you get the point of the course, basic exercises01:04
n4nd0but the labs require lot of work01:04
blackburnI have two algos to suggest to implement to you01:05
n4nd0tell me01:05
blackburnmaximum variance unfolding and t-SNE01:05
blackburnfirst one is simple01:05
n4nd0do you have a good reference in a paper or a book for them?01:06
blackburnn4nd0: in fact we would even create some idea for dimensionality reduction..01:06
blackburnmatlab code is provided in the matlab dimensionality reduction toolbox01:06
blackburnand some paper may be01:07
blackburnhmm I need to talk with soeren :)01:07
blackburnI guess it can be an idea and I could help smb with mentoring01:08
blackburnn4nd0: broad choice for you :)01:08
-!- cronor [] has joined #shogun01:09
blackburn4.09 already and I have to be on my job on 10.00 damnation!01:10
n4nd0I think you should get some sleep01:10
blackburncronor: sorry is it your pull req?01:11
blackburnif so - we need some time for this issue :)01:11
cronorblackburn: huh?01:11
blackburncronor: then not your :)01:12
cronorblackburn: probably yes01:12
blackburnn4nd0: okay see you01:12
n4nd0blackburn: good night01:12
blackburngood night01:13
-!- blackburn [d55780cc@gateway/web/freenode/ip.] has quit [Quit: Page closed]01:13
-!- cronor [] has left #shogun []01:44
-!- n4nd0 [] has quit [Quit: leaving]02:43
-!- vikram360 [~vikram360@] has quit [Read error: Connection reset by peer]02:53
-!- vikram360 [~vikram360@] has joined #shogun02:54
-!- vikram360 [~vikram360@] has quit [Ping timeout: 276 seconds]03:17
-!- Zoe [a66f43f2@gateway/web/freenode/ip.] has joined #shogun04:10
-!- Zoe [a66f43f2@gateway/web/freenode/ip.] has left #shogun []04:12
-!- cronor [] has joined #shogun05:16
-!- vikram360 [~vikram360@] has joined #shogun05:27
-!- vikram360 [~vikram360@] has quit [Ping timeout: 272 seconds]05:33
-!- vikram360 [~vikram360@] has joined #shogun05:33
-!- vikram360 [~vikram360@] has quit [Ping timeout: 244 seconds]05:57
-!- vikram360 [~vikram360@] has joined #shogun05:57
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun06:11
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]06:32
-!- cronor [] has quit [Quit: cronor]08:01
-!- vikram360 [~vikram360@] has quit [Ping timeout: 260 seconds]08:02
-!- vikram360 [~vikram360@] has joined #shogun08:02
-!- vikram360 [~vikram360@] has quit [Ping timeout: 252 seconds]08:33
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun08:48
-!- n4nd0 [] has joined #shogun09:12
-!- n4nd0 [] has quit [Ping timeout: 265 seconds]09:54
-!- _dr [] has joined #shogun10:06
-!- n4nd0 [] has joined #shogun10:33
CIA-64shogun: Soeren Sonnenburg master * rcdd0218 / src/shogun/distance/MahalanobisDistance.h : fix latex error in documentation (x_i not \x_i) -
n4nd0sonne|work: ups sorry for that typo :(11:28
n4nd0sonne|work: one question about QDA, where does it make more sense in the class hierarchy?11:29
sonne|workn4nd0: no that was my typo!11:30
sonne|workno need to be sorry for the mistakes I make11:30
n4nd0sonne|work: ah ok! :)11:30
n4nd0sonne|work: I think it should go directly under CMachine11:30
n4nd0sonne|work: I have not seen other quadratic classifiers in the toolbox11:31
sonne|workindeed ...11:32
n4nd0so it is ok doing CQDA directly under CMachine or something in between like CQuadraticMachine analogous to CLinearMachine?11:33
sonne|workjust do it directly for now11:38
sonne|workas long as we don't have any other quadratic machine it doesn't make sense to be overly general11:39
n4nd0I see, ok then :)11:40
-!- Netsplit *.net <-> *.split quits: naywhayare13:20
-!- Netsplit over, joins: naywhayare13:23
-!- vikram360 [~vikram360@] has joined #shogun13:36
-!- wiking [~wiking@huwico/staff/wiking] has quit [Read error: Connection reset by peer]15:23
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun15:23
-!- n4nd0 [] has quit [Ping timeout: 252 seconds]15:23
-!- n4nd0 [] has joined #shogun15:41
-!- n4nd0 [] has quit [Ping timeout: 252 seconds]16:19
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds]17:39
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun17:45
-!- puneetgoyal [~puneet@] has joined #shogun18:41
vikram360What is the Software Engineering approach that Shogun developers usually tend to take when building a new feature into shogun? Is a prototype first built, tested and then integrated18:43
vikram360Is there any strict methodology that all developers have to follow when building a feature18:44
-!- n4nd0 [] has joined #shogun18:44
-!- n4nd0 [] has quit [Ping timeout: 252 seconds]18:50
-!- n4nd0 [] has joined #shogun18:51
-!- blackburn [~qdrgsm@] has joined #shogun18:52
-!- n4nd0 [] has quit [Ping timeout: 244 seconds]18:56
-!- n4nd0 [] has joined #shogun19:08
-!- blackburn [~qdrgsm@] has quit [Ping timeout: 255 seconds]19:16
-!- blackburn [~qdrgsm@] has joined #shogun19:28
blackburnsonney2k: around?19:40
-!- in3xes [~in3xes@] has joined #shogun20:20
-!- in3xes [~in3xes@] has quit [Ping timeout: 260 seconds]20:25
-!- in3xes [~in3xes@] has joined #shogun20:25
-!- in3xes [~in3xes@] has quit [Ping timeout: 252 seconds]20:33
-!- in3xes [~in3xes@] has joined #shogun20:49
-!- n4nd0 [] has quit [Read error: Operation timed out]20:50
-!- n4nd0 [] has joined #shogun20:52
-!- blackburn [~qdrgsm@] has quit [Ping timeout: 245 seconds]21:31
-!- in3xes [~in3xes@] has quit [Ping timeout: 240 seconds]21:34
-!- puneetgoyal [~puneet@] has quit [Quit: Leaving]22:12
-!- n4nd0 [] has quit [Ping timeout: 255 seconds]22:27
@sonney2kvikram360, really depends - you could write an external class + some main method and just link to libshogun from c++ side and then when you feel comfortable submit your extension to core shogun providing some examples and interface integration - since we  use these examples also as regression tests everything will be fine then.22:49
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking]23:33
--- Log closed Wed Mar 07 00:00:19 2012