Open in new window / Try shogun cloud
--- Log opened Wed Jan 30 00:00:38 2013
-!- cronor [~cronor@] has quit [Quit: cronor]00:59
-!- FSCV [~FSCV@] has quit [Quit: Leaving]01:31
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.] has quit [Quit: Page closed]01:55
shogun-buildbotbuild #266 of nightly_default is complete: Failure [failed test]  Build details are at
-!- ptizoom [] has joined #shogun07:44
-!- cronor [] has joined #shogun09:22
-!- sumit_ [ca4eaca2@gateway/web/freenode/ip.] has joined #shogun10:30
sonne|workptizoom !10:38
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.] has joined #shogun10:41
sonne|workn4nd0: I realize that an extra page won't work - we would somehow need another nesting level10:56
n4nd0sonne|work: mmm not sure why it won't work10:56
n4nd0tell me10:56
sonne|workn4nd0: It would be good to have one page with announcements/news about the workshop, one with the program and one for registration and later one for user contributed slides/code or so10:57
n4nd0what is your idea? how do you want it to look like?10:57
sonne|workbut this will only work for 201310:57
sonne|workif we have another WS in 2014 what then?10:57
n4nd0ok I see your point now10:57
n4nd0the idea would be to go to events10:58
n4nd0later to workshop 201310:58
n4nd0and then find those pages10:58
n4nd0right now workshop 2013 and workshop program are at the same level, which it is not that ^ idea10:58
-!- sumit_ [ca4eaca2@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]11:24
n4nd0sonne|work: I am talking to Bernard about it11:46
n4nd0we will find a solution11:46
sonne|workcool thx11:47
n4nd0sonne|work: btw I still have my SO-SVM question :)11:48
n4nd0maybe you have a minute for it?11:49
sonne|workwhich is?11:49
n4nd0does it handle predictions over continuous structured data?11:49
n4nd0for what I have read so far, the framework presented in Tsochantaridis which is what I know for SO-SVM doesn't do it11:50
n4nd0some papers about kernel density estimation, and joint kernel estimation handle that; but those are other things11:50
sonne|workyou mean e.g. a task like learn where roads end, e.g. split a road after x meters11:51
sonne|workso non-discrete segmentation11:51
n4nd0think of the HM-SVM too for example, the SO-SVM predicts sequence of states there11:53
n4nd0and each state is identified with an integer11:53
n4nd0what about sequences of states where each state is a real number?11:53
n4nd0it sounds a bit like regression I think, I have seen something called structured output regression too11:54
sonne|workn4nd0: I think something like this is around for at least hidden markov models11:54
sonne|worksomething with continuous state space hmms11:55
n4nd0I have been looking around quite a lot but didn't find anything promising to tell the truth11:55
sonne|workI just had a look11:56
sonne|workwikipedia names it also like this11:56
sonne|workno idea about concrete papers though11:57
sonne|workdidn't need it yet11:57
n4nd0sonne|work: I looke up in wikipedia, they called it linear dynamical system11:58
blackburnI am still convinced linear dynamical systems are related to math :D11:59
sonne|worksounds like differential equation11:59
n4nd0blackburn: isn't machine learning related to math? :)11:59
blackburnI consider machine learning as a tool to *estimate* dependencies not known before12:00
n4nd0state spaces are represented via differential equations12:00
n4nd0blackburn: don't you find that correspondence in this problem?12:00
n4nd0estimate the dependencies between observations and states that have produced those observations12:01
blackburnwell differential equation is known dependency with unknown parameters12:01
n4nd0in addition to dependencies between states in a chain themselves12:01
blackburnokay, I just can't imagine so-svm find something for that :)12:02
n4nd0blackburn: why not?12:02
n4nd0blackburn: does it seem natural for you if the states are discrete? i.e. in a more common HMM12:03
blackburnI don't know - a little more12:04
sonne|workblackburn: well my example above  a task like learn where roads end, e.g. split a road after x meters would be continuous state space12:05
blackburnsonne|work: and estimate that from say video?12:06
blackburnthat sounds pretty complex to handle with anything linear12:07
sonne|workthough video is descrete :)12:07
blackburnwell I have seen SO used for stereo disparity estimation12:07
blackburnn4nd0: is that the case?12:07
n4nd0blackburn: yes, it is an example12:08
blackburnthen I am more convinced it is possible12:08
n4nd0people use SO for human pose estimation all the time12:08
n4nd0there are a bunch of papers about it12:08
n4nd0that is continuous state space12:08
blackburnthe main problem is features construction I guess12:08
n4nd0I would guess for SIFT and SURF type of features12:09
blackburnwith keypoints?12:10
blackburnor dense?12:10
ptizoomhi... I am lost....13:05
n4nd0hi ptizoom13:09
n4nd0what is it?13:09
ptizoomI spend a lot of time trying to link different idea on auto-learning instead of trying to finishing off this swig-perl with sparse matrix...13:09
n4nd0you probably learnt a lot on auto-learning :)13:10
ptizoomalso still looking for a smart way to translate 'python' tests into 'perl' tests!13:11
n4nd0I see13:11
ptizoomplease direct me into a nice group who does this kind of stuff "learning and matching parsing relations"...13:13
n4nd0I have never used that13:14
ptizoomyep, parsers are nice candidates to be auto-learned! I thought using hidden markov approx (like every one), but really how to picking up local features an work relevant relations in an  orthogonal (kernelized) fashion...13:21
ptizoomeg. does someone know how to create a stemmer ? I believe HMM algo is not enough.13:23
ptizoomeven more it seems shogun can only use a 2-gram HMM.....13:25
ptizoomho well... may be see you in Berlin!13:28
sonne|workptizoom: yeah come to berlin :)13:52
n4nd0ptizoom: regarding the stemmer13:55
n4nd0I suggest you to take a look to
n4nd0they have very good material in there, slides, book, etc13:56
-!- FSCV [~FSCV@] has joined #shogun14:16
ptizoomthanks n4nd0,(I like this name), I have been rumbling my soul at indeed. I even got up to taste the madness of Wolfram...and his amazing cellular automatons... that shall be the way to program (he said), I ended up in looking into the F4 algo to reduce critical pairs...yes, one to select features needs to make some choices....15:38
n4nd0it sounds like you had lot of fun :)15:40
ptizoomyes, but I still do not get it.15:40
ptizoomit is something beautiful , about ways of rewriting equations, involving the golden ratio....!15:45
-!- sumit [ca4eaca2@gateway/web/freenode/ip.] has joined #shogun15:46
ptizoomanyway, yes will do! Thanks for the tip.15:50
-!- sumit [ca4eaca2@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]16:09
n4nd0you are welcome16:10
-!- kakashi_ [~kakashi_@nltk/kakashi] has joined #shogun16:14
kakashi_Hi, does shogun have an SVD implementation? If so, is there a performance analysis of the same?16:14
sonne|workblackburn: ^16:17
kakashi_well, I was just browsing through the code and foundout a rough jacobi implementation?16:18
kakashi_or it seems to use eigen++, hmmm16:19
n4nd0kakashi_: we mainly use eigen or arpack for linear algebra16:22
kakashi_n4nd0: aha, can you give me any links? also, can I use sparse matrices?16:26
n4nd0kakashi_: where do you want to use the sparse matrices?16:26
n4nd0we have some nice sparse features :)16:26
n4nd0links to where? examples of where svd is used in shogun?16:27
kakashi_yes, well, liknks to some examples16:27
kakashi_at present I have a sparse matrix in scipy.sparse or pysparse format (and ofcourse their combinations including a (row, column, data) as a csv is also doable)16:29
kakashi_dimensions are approx (10000x200000)16:29
n4nd0I will be back in a few minutes16:30
kakashi_take your time16:31
n4nd0kakashi_: in multiclass/QDA.cpp SVD is used IIRC16:38
n4nd0yeah exactly16:38
n4nd0wrap_dgesvd sorry16:39
kakashi_I see, is there any python wrapper, by anychance?16:41
kakashi_hmm, btw, this is SVD of a dense matrix, right?16:44
n4nd0kakashi_: I think that example probably is16:46
n4nd0what do you want to do? svd for a sparse matrix from python?16:46
n4nd0shouldn't there be something in scipy for that?16:47
kakashi_python would be great (numpy/scipy/sparsesvd are pretty slow)16:47
kakashi_scipy.sparse uses arapack implementation, a single svd of the 10000x200000 takes approx 2-3 days16:48
-!- heiko [] has joined #shogun16:51
n4nd0hehe that's lot of time16:51
n4nd0I am not aware if there are faster implementations16:51
n4nd0maybe there are some approximate methods out there16:51
kakashi_hmm, thanks anyways!16:54
sonne|workkakashi_: might be that blackburn knows though...16:56
kakashi_aha! sure, I will keep an eye on it ...16:57
kakashi_I have not seen anything that is as good as mahout's implementation, but the pity is all my code is in java and JPype may obfuscate the code :(16:58
-!- cavalier- [~cavalier-@] has quit [Ping timeout: 260 seconds]17:25
-!- cavalier- [~cavalier-@] has joined #shogun17:26
n4nd0kakashi_: do you happen to know what kind of implementation do the use in mahout?17:35
-!- cronor [] has quit [Quit: cronor]18:00
-!- cavalier- [~cavalier-@] has quit [Ping timeout: 276 seconds]18:09
-!- cavalier- [~cavalier-@] has joined #shogun18:09
kakashi_I think they use jacobi implementation18:24
kakashi_n4nd0: i will be back in a while18:27
blackburnkakashi_: we have some code ported from redsvd18:39
blackburnmay be it is of interest for you18:40
-!- cronor [] has joined #shogun18:43
blackburnkakashi_: in my experience it doesn't work that well for sparse matrices due to randomization but I think that's quite ok still18:44
-!- heiko [] has quit [Quit: Leaving.]18:49
-!- heiko [] has joined #shogun18:50
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]19:26
-!- heiko [] has left #shogun []20:26
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.] has joined #shogun20:37
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.] has quit [Quit: Page closed]21:31
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.] has joined #shogun21:49
@sonney2kn4nd0, what did bernard say?23:34
n4nd0sonney2k: that we have to change some css and python code23:36
n4nd0he said he is a bit busy now23:36
n4nd0but I will try to do it with his mentorship :D23:36
blackburnsonney2k: so does that mean I should answer you?23:41
@sonney2kblackburn, ?23:41
blackburnyou submitted the form23:41
@sonney2kn4nd0, ok23:41
@sonney2kblackburn, yes I am trying to automate the stuff23:42
@sonney2klike that the mailinglist gets the form23:42
blackburnsonney2k okay - am I subscribed to that list btw?23:43
@sonney2kA user made changes from 1/31/13 2:40 AM to 2:40 AM (Azerbaijan Standard Time) ?23:43
@sonney2kblackburn, IDK23:43
blackburnsonney2k: yeah so you live in azerbaijan23:43
blackburnit is not the best notification we could see23:44
@sonney2knot very helpful23:44
blackburnyeah we can view the lnie but still23:44
@sonney2kblackburn, would even say that this is useless23:44
@sonney2kso I think it would be better you reply everyone submitting manually23:45
blackburnsonney2k: yeah I see no other way so far23:46
@sonney2kit is no use to create any kind of automagic notification23:46
blackburnsonney2k: some js could help23:46
@sonney2kso about the mailinglist23:47
blackburnbut nobody wants to do that23:47
@sonney2khow should we do it?23:47
@sonney2ksubscribe people to it?23:47
blackburnhmm can you let me send mails from that list?23:47
@sonney2kI originally wanted to use it for notifications23:47
@sonney2ksuch that heiko/you/me etc get an email23:48
@sonney2kbut now idk23:48
@sonney2kblackburn, you can use whatever address you want in from:23:48
blackburnwell probably the best way could be - I subscribe for these notices23:49
blackburnonce somebody submits the form I write him a mail from @shogun-toolbox.org23:49
blackburnand CC it to workshop@shogun-toolbox.org23:49
blackburnwhich has you, me, .... in subscribers23:50
blackburnsleep time23:52
@sonney2kblackburn, yup!23:53
@sonney2kblackburn, or alternatively you subscribe everyone registering to that mailinglist and we can use it for updates later23:53
blackburnsonney2k: we need something private then23:53
@sonney2kn4nd0, hmmhh I tried to embed the form23:53
blackburnto BCC23:53
@sonney2kblackburn, ?23:54
@sonney2kblackburn, mailinglist is not giving out email addresses of others...23:54
blackburnI mean if we subscribe they will receive answers purposed for other people23:54
@sonney2kn4nd0, but it does not like me23:54
@sonney2kblackburn, ahh I see yes sure23:55
blackburnlets just use it for our purposes23:55
@sonney2kblackburn, might be enough to just send around these emails among us few (all in CC)23:55
blackburnhmm yes23:55
blackburnthen mailing list is not needed for all23:55
n4nd0sonney2k: what can it be?23:55
@sonney2kn4nd0, it converts the iframe to some text23:55
@sonney2kthat is the code we need to insert to have it directly on the page <iframe src="" width="760" height="1629" frameborder="0" marginheight="0" marginwidth="0">Loading...</iframe>23:56
@sonney2kn4nd0, but it removes iframes23:56
blackburnisn't iframe something from 1992?23:56
@sonney2kblackburn, only for people that attend23:56
blackburnhmm I am probably wrong23:57
@sonney2kblackburn, click 'embed on website' in the spreadsheet23:57
@sonney2kblackburn, I guess you meant frames23:57
blackburnsonney2k: yeah I mix it up with something else23:57
blackburnframes are left behind before your first kiss23:57
blackburnsonney2k: do you mean integrating it to the page?23:59
blackburnisn't better to just put a url?23:59
--- Log closed Thu Jan 31 00:00:37 2013