--- Log opened Tue Aug 02 00:00:23 2011 | ||
alesis-novik | blackburn, you there? | 01:22 |
---|---|---|
blackburn | alesis-novik: yes | 01:22 |
alesis-novik | blackburn, what does the k argument in LLE mean? | 01:30 |
blackburn | alesis-novik: the number of neighbors used for linear reconstruction | 01:31 |
alesis-novik | thanks | 01:34 |
blackburn | alesis-novik: why are you interested in lle? ;) | 01:35 |
alesis-novik | I'm getting ready for my PhD and my supervisor asked me to do some stuff. Essentially I'm hitting stock profit data from multiple strategies as dimentions with different nldr techniques :D | 01:36 |
blackburn | hmm let me know if you will get some success with it | 01:37 |
blackburn | I'm pretty interested in manifold learning | 01:37 |
blackburn | alesis-novik: how large is data? | 01:41 |
alesis-novik | blackburn, huge, but I'm using subsets. | 01:42 |
blackburn | and dimension? | 01:42 |
blackburn | alesis-novik: I would suggest you to use LTSA first, it is faster >4x than LLE | 01:43 |
blackburn | and sometimes it produces good results | 01:43 |
alesis-novik | it has 161 dimensions | 01:43 |
blackburn | I've applied LTSA to 200x200 images (40000d->2d) and it was pretty well | 01:44 |
blackburn | alesis-novik: http://dl.dropbox.com/u/10139213/shogun/LocalTangentSpaceAlignment%28k%3D40%29.png | 01:46 |
alesis-novik | Cool | 01:46 |
blackburn | alesis-novik: the underlying dimension is angle and brightness or so | 01:46 |
blackburn | dimensions* | 01:46 |
alesis-novik | thanks blackburn, this should be useful | 01:46 |
alesis-novik | essentially I'm trying to find if there is some underlying structure that the strategies are using to get profit | 01:47 |
blackburn | alesis-novik: are you using shogun techniques? | 01:47 |
blackburn | alesis-novik: one 'advice' I would say - better use larger data than smaller, sometimes manifold is ruined with bad sampling | 01:48 |
alesis-novik | blackburn, how many datapoints would be realistic with 161 dimensions? | 01:49 |
blackburn | LTSA have something like O(N^2 + N*k^2) or so | 01:50 |
blackburn | alesis-novik: it depends on sample size more than on data dimension, I guess time for 3k-5k examples will be under one minute | 01:51 |
alesis-novik | I'm not worried that much about time, more about memory | 01:52 |
blackburn | now it uses NxN matrix | 01:52 |
blackburn | in fact it should be sparse but we lack some sparse matrix routines.. | 01:53 |
blackburn | alesis-novik: >6k will take >2g memory | 01:54 |
alesis-novik | cool | 01:54 |
alesis-novik | one more thingy - any idea what could cause this in Classic MDS: RuntimeWarning: [WARN] Embedding is not consistent: features 3-4 are wrong | 01:55 |
blackburn | alesis-novik: it is because eigenvalues are <0 | 01:55 |
alesis-novik | ah, ok. Thanks | 01:55 |
blackburn | alesis-novik: just try to decrease target dimension, it means mds can't construct embedding to 4d space | 01:57 |
blackburn | and usually number of positive eigenvalues shows the underlying dimension | 01:58 |
blackburn | e.g. for swissroll dataset it is 2 just as manifold dimension | 01:58 |
blackburn | alesis-novik: you could use LandmarkMDS for large datasets with number of landmarks > dimensionality | 01:59 |
blackburn | usually it is pretty accurate and much faster than Classic one | 02:00 |
alesis-novik | I'll try that | 02:00 |
alesis-novik | landmarks > target or original dim? | 02:00 |
blackburn | original | 02:00 |
blackburn | it uses triangulation like GPS use | 02:00 |
alesis-novik | I don't really have experience with DR other than the linear ones :D | 02:00 |
blackburn | as for me I don't have any experience with linear ones :D | 02:00 |
blackburn | alesis-novik: landmark mds chooses some landmarks and constructs embedding as classic do for these landmarks | 02:01 |
blackburn | for other points it just compute triangulation | 02:02 |
alesis-novik | thanks for the tips, this will be useful | 02:04 |
blackburn | alesis-novik: I just tried to embed 7000 examples with landmark mds and it took 3.03s | 02:05 |
blackburn | not very useful but fast | 02:05 |
blackburn | alesis-novik: and btw, laplacian eigenmaps will be implemented soon | 02:07 |
serialhex | you and your floating face images blackburn!!! | 02:57 |
blackburn | serialhex: hehe | 02:58 |
serialhex | i just looked at the image you posted a bit ago... and almost all of your images are like that :P | 02:59 |
serialhex | so how have you been? | 02:59 |
blackburn | well that's how I can compare different methods ;) | 02:59 |
blackburn | fine, but I don't understand why I'm not sleeping hehe | 02:59 |
serialhex | yeah, it's like 9pm here, i can imagine how late it is there :P | 03:01 |
blackburn | 5am hehe | 03:02 |
serialhex | oh man! | 03:02 |
serialhex | yeah, i'm thinking of getting up extra early so i can talk to soeren & settle some Q's i had about how the configure stuff works | 03:02 |
blackburn | boom | 03:07 |
CIA-87 | shogun: Sergey Lisitsyn master * ra2be4dd / (4 files): Added fast neighbor search for LLE, HLLE, LTSA with Fibonacci heap - https://github.com/shogun-toolbox/shogun/commit/a2be4dde5f464aa0e202e3d5ef1abb68209a09d5 | 03:07 |
serialhex | so what? you have write access to the repo or somethin blackburn? | 03:09 |
serialhex | and whats a fibonacci heap?? | 03:09 |
serialhex | sounds like it has something to do with the fib seq | 03:10 |
blackburn | serialhex: yes I have, since june or so ;) | 03:12 |
serialhex | cool cool | 03:12 |
blackburn | serialhex: fibonacci heap is a data structure like binomial heap | 03:12 |
blackburn | indeed, it is related to fib seq | 03:13 |
serialhex | cool | 03:14 |
blackburn | it is good for dijkstra algo | 03:14 |
blackburn | because extract-min have O(lg n) time complexity | 03:14 |
serialhex | nice | 03:14 |
blackburn | actually it can't be faster hehe | 03:15 |
blackburn | because the fastest sort is O(n log n) | 03:15 |
serialhex | unfortunatey i only know enough to know that i don't know enough :P i need to learn more! | 03:15 |
serialhex | niiice!! | 03:15 |
* serialhex is reading the wikipedia article, and should have done that in the first place :P | 03:17 | |
blackburn | I just heard us senate got some convention, right? | 03:18 |
serialhex | yeah, something like that... i should really pay more attention but i don't really | 03:18 |
blackburn | I pay attention to it because usd/rub rate depends on that | 03:18 |
blackburn | and I want higher rate hehe | 03:19 |
@bettyboo | he | 03:19 |
serialhex | yeah, our $$ is going down unfortunately... because our polittons are all crooks! | 03:19 |
serialhex | that's a big reason why i dont pay much attention: it's too depressing!! | 03:20 |
blackburn | haha what would you say if you was in russia | 03:22 |
serialhex | i don't know, but probably about the same :D they're all crooks, wherever they are in the world! | 03:23 |
blackburn | at least yours obey laws :) | 03:23 |
blackburn | 5-25 am and somebody is jumping, the whole house is joggling :D | 03:25 |
blackburn | uh I have to sleep | 03:49 |
blackburn | serialhex: see ya | 03:49 |
serialhex | nite blackburn | 03:49 |
-!- blackburn [~blackburn@188.122.224.26] has quit [Quit: Leaving.] | 03:50 | |
-!- alesis-novik [~alesis@cpat001.wlan.net.ed.ac.uk] has quit [Quit: Leaving] | 04:04 | |
-!- in3xes [~in3xes@180.149.49.227] has joined #shogun | 06:51 | |
-!- f-x [~user@117.192.221.58] has joined #shogun | 07:20 | |
-!- f-x [~user@117.192.221.58] has quit [Ping timeout: 260 seconds] | 07:34 | |
-!- f-x [~user@117.192.193.154] has joined #shogun | 07:46 | |
-!- in3xes [~in3xes@180.149.49.227] has quit [Ping timeout: 240 seconds] | 08:28 | |
-!- in3xes1 [~in3xes@180.149.49.227] has joined #shogun | 08:41 | |
-!- in3xes_ [~in3xes@210.212.58.111] has quit [Ping timeout: 258 seconds] | 08:45 | |
-!- in3xes [~in3xes@180.149.49.227] has joined #shogun | 08:52 | |
-!- f-x [~user@117.192.193.154] has quit [Read error: Connection reset by peer] | 08:56 | |
-!- f-x [~user@117.192.193.154] has joined #shogun | 09:11 | |
-!- in3xes_ [~in3xes@180.149.49.227] has joined #shogun | 09:33 | |
-!- in3xes1 [~in3xes@180.149.49.227] has quit [Ping timeout: 240 seconds] | 09:37 | |
-!- in3xes__ [~in3xes@59.163.196.121] has joined #shogun | 11:15 | |
-!- in3xes1 [~in3xes@59.163.196.121] has joined #shogun | 11:15 | |
-!- in3xes [~in3xes@180.149.49.227] has quit [Ping timeout: 240 seconds] | 11:17 | |
-!- in3xes_ [~in3xes@180.149.49.227] has quit [Ping timeout: 240 seconds] | 11:17 | |
-!- heiko [~heiko@134.91.54.169] has joined #shogun | 12:25 | |
heiko | sonney2k, around? | 12:36 |
-!- in3xes_ [~in3xes@180.149.49.227] has joined #shogun | 12:37 | |
-!- in3xes1 [~in3xes@59.163.196.121] has quit [Ping timeout: 240 seconds] | 12:41 | |
-!- in3xes__ [~in3xes@59.163.196.121] has quit [Ping timeout: 276 seconds] | 12:41 | |
-!- in3xes__ [~in3xes@180.149.49.227] has joined #shogun | 12:44 | |
-!- in3xes_ [~in3xes@180.149.49.227] has quit [Ping timeout: 240 seconds] | 12:48 | |
-!- in3xes_ [~in3xes@180.149.49.227] has joined #shogun | 12:49 | |
-!- in3xes__ [~in3xes@180.149.49.227] has quit [Ping timeout: 276 seconds] | 12:53 | |
-!- in3xes_ [~in3xes@180.149.49.227] has quit [Remote host closed the connection] | 12:54 | |
-!- blackburn [~blackburn@109.226.69.219] has joined #shogun | 13:07 | |
-!- heiko [~heiko@134.91.54.169] has quit [Ping timeout: 258 seconds] | 13:54 | |
-!- heiko [~heiko@134.91.54.169] has joined #shogun | 14:39 | |
@sonney2k | heiko, today only at >20hrs | 14:49 |
heiko | sonney2k, ok | 14:49 |
-!- gsomix [~gsomix@178.45.94.247] has joined #shogun | 14:52 | |
-!- blackburn [~blackburn@109.226.69.219] has quit [Ping timeout: 255 seconds] | 14:52 | |
-!- gsomix [~gsomix@178.45.94.247] has quit [Quit: Ухожу я от вас (xchat 2.4.5 или старше)] | 14:58 | |
-!- in3xes_ [~in3xes@180.149.49.227] has joined #shogun | 15:22 | |
-!- in3xes_ is now known as in3xes | 15:22 | |
-!- in3xes_ [~in3xes@180.149.49.227] has joined #shogun | 15:43 | |
-!- in3xes [~in3xes@180.149.49.227] has quit [Ping timeout: 240 seconds] | 15:46 | |
-!- in3xes__ [~in3xes@180.149.49.227] has joined #shogun | 16:06 | |
-!- in3xes__ is now known as in3xes | 16:09 | |
-!- in3xes_ [~in3xes@180.149.49.227] has quit [Ping timeout: 240 seconds] | 16:10 | |
-!- sploving1 [~sploving@210.77.14.135] has joined #shogun | 16:19 | |
sploving1 | sonney2k, are you around? | 16:19 |
-!- sploving1 [~sploving@210.77.14.135] has left #shogun [] | 16:34 | |
-!- in3xes_ [~in3xes@180.149.49.227] has joined #shogun | 17:03 | |
-!- in3xes [~in3xes@180.149.49.227] has quit [Ping timeout: 240 seconds] | 17:06 | |
-!- in3xes__ [~in3xes@180.149.49.227] has joined #shogun | 17:21 | |
-!- in3xes_ [~in3xes@180.149.49.227] has quit [Ping timeout: 240 seconds] | 17:25 | |
-!- in3xes__ is now known as in3xes | 17:45 | |
-!- blackburn [~blackburn@109.226.69.194] has joined #shogun | 18:14 | |
-!- heiko [~heiko@134.91.54.169] has quit [Ping timeout: 258 seconds] | 18:21 | |
-!- f-x [~user@117.192.193.154] has quit [Ping timeout: 260 seconds] | 19:25 | |
-!- in3xes_ [~in3xes@180.149.49.227] has joined #shogun | 19:44 | |
-!- in3xes [~in3xes@180.149.49.227] has quit [Ping timeout: 240 seconds] | 19:48 | |
-!- in3xes__ [~in3xes@180.149.49.227] has joined #shogun | 19:50 | |
-!- in3xes_ [~in3xes@180.149.49.227] has quit [Ping timeout: 276 seconds] | 19:54 | |
-!- in3xes_ [~in3xes@180.149.49.227] has joined #shogun | 20:00 | |
-!- in3xes__ [~in3xes@180.149.49.227] has quit [Ping timeout: 276 seconds] | 20:04 | |
CIA-87 | shogun: Soeren Sonnenburg master * r6387718 / (29 files in 9 dirs): | 21:05 |
CIA-87 | shogun: Merge pull request #258 from karlnapf/master | 21:05 |
CIA-87 | shogun: model selection now really works on model_selection_parameters (+24 more commits...) - https://github.com/shogun-toolbox/shogun/commit/6387718cb3a8b19bb0ff3ff603cfefd0714056db | 21:05 |
-!- in3xes__ [~in3xes@180.149.49.227] has joined #shogun | 21:28 | |
-!- in3xes_ [~in3xes@180.149.49.227] has quit [Ping timeout: 276 seconds] | 21:32 | |
-!- in3xes__ is now known as in3xes | 21:39 | |
-!- in3xes [~in3xes@180.149.49.227] has quit [Quit: Leaving] | 21:39 | |
blackburn | sonney2k: around? | 21:44 |
@sonney2k | y | 21:44 |
@sonney2k | blackburn, you too ;-) | 21:46 |
@sonney2k | ? | 21:46 |
blackburn | what me? ;) | 21:46 |
@sonney2k | around? | 21:46 |
blackburn | sonney2k: sagen sie mal, have you ever worked with mutual k-nearest neighbor graphs? | 21:46 |
@sonney2k | blackburn, nitschewo net | 21:46 |
blackburn | sonney2k: ploho | 21:47 |
blackburn | sonney2k: so, howdy? | 21:47 |
@sonney2k | I finished plumbing my 'server' back together | 21:48 |
@sonney2k | now it is time to do some buildbot stuff ... I really hope I have that going by the end of the week | 21:48 |
blackburn | how do you mind we can use it? | 21:49 |
@sonney2k | blackburn, I could install buildbot - then we get nice pictures like this http://trac.buildbot.net/wiki/ScreenShots | 21:51 |
blackburn | sonney2k: so every change will be checked one more time on buildbot and we will fast notice if something goes bad? | 21:52 |
@sonney2k | blackburn, or detect a bug in the buildbot yes :) | 21:53 |
blackburn | I see | 21:53 |
blackburn | it is pre-release thing, right? | 21:53 |
blackburn | sonney2k: ah! and you plan to do some VM so we can check compilation under windoze/osx, right? | 21:54 |
blackburn | or not? | 21:54 |
@sonney2k | blackburn, it is a debian package for some time | 21:54 |
@sonney2k | blackburn, if I manage osx and some cygwin too | 21:54 |
@sonney2k | blackburn, then I put sth like this on my desk http://mackeblog.blogspot.com/search/label/lavalamp | 21:55 |
blackburn | ahaha | 21:56 |
blackburn | you will wake up not only with baby's cry but with shogun's cry too | 21:56 |
blackburn | and robovoice should say ZZZZZZZ COMPILATION FAILED | 21:56 |
@sonney2k | blackburn, I actually have an alarm light in my office | 22:01 |
@sonney2k | would be pretty impressive to use it for that task :D | 22:01 |
blackburn | fire alarm, terrorist alarm, shogun alarm | 22:02 |
@sonney2k | nahh shogun alarm should be rank #1 | 22:02 |
blackburn | sonney2k: how expensive to use DynArray for storing integers? | 22:03 |
@sonney2k | it is like int[] | 22:03 |
@sonney2k | no error checks.. | 22:03 |
blackburn | hmm why is it Dyn? | 22:03 |
blackburn | hmm I need a Set | 22:04 |
blackburn | sonney2k: we don't have Set, right | 22:04 |
blackburn | ? | 22:04 |
@sonney2k | blackburn, we have a poor mans set | 22:05 |
@sonney2k | CSet | 22:05 |
blackburn | how poor it is? | 22:05 |
@sonney2k | it is a Q&D hack with O(n) insert/dedup costs | 22:05 |
@sonney2k | I wrote this only for the memory tracing stuff | 22:06 |
blackburn | ah I see | 22:06 |
@sonney2k | blackburn, so sth great for gsomix I would say :) | 22:06 |
@sonney2k | He seems to be data structures affine | 22:06 |
blackburn | I've suggested him to parallelize distance and kernels | 22:06 |
blackburn | sonney2k: can I use std for now? | 22:07 |
@sonney2k | well if you have too.... | 22:08 |
blackburn | sonney2k: we already have some std in mkl or so | 22:08 |
blackburn | sonney2k: why are you against it btw? | 22:08 |
@sonney2k | I am not a big fan of using C++ libraries for standard tasks... I guess nowadays historical esoteric reasons :-| | 22:11 |
blackburn | well I don't like std-like code with iterators, much templates and blablabla | 22:12 |
blackburn | but the set is example where we are going to duplicate efforts or so | 22:12 |
blackburn | sonney2k: is it ok to just make a front-end to std:set? | 22:14 |
@sonney2k | blackburn, yeah but where do you draw the line? How good is the 'set' implementation in std? What overhead does it involve? | 22:14 |
blackburn | I guess it should be very good | 22:14 |
blackburn | hm stl not std | 22:15 |
@sonney2k | blackburn, I had the issue that using qsort() from c library was 10 times slower than my stupid qsort() and it was ruining a whole algorithm because suddenly we had an unexpected bottleneck | 22:18 |
blackburn | uh | 22:18 |
@sonney2k | so I am saying check carefully ... | 22:18 |
blackburn | sonney2k: btw fibonacci heap makes possible to do kNN finding in O(klg n), I've already adapted my LLE* algos to do that faster | 22:20 |
@sonney2k | nice... does john langford's covertree improve this further? | 22:20 |
blackburn | yes, sure | 22:21 |
blackburn | for large dimensions | 22:21 |
blackburn | but in fact it is not a bottleneck | 22:21 |
blackburn | spectral methods involve large-scale eigenvalues problems and it is the most time-demanding thing | 22:22 |
blackburn | so covertree will improve only kNN | 22:23 |
@sonney2k | blackburn, or we use that here | 22:25 |
@sonney2k | http://bitten.edgewall.org/build/trunk | 22:25 |
@sonney2k | 'bitten' | 22:25 |
@sonney2k | nahh buildbot looks better... | 22:26 |
@sonney2k | blackburn, buildbot seems to even have an IRC bot :) | 22:30 |
blackburn | hehe | 22:30 |
@sonney2k | so this channel will finally be taken over by robots :D | 22:30 |
@bettyboo | rotfl | 22:30 |
blackburn | it's funny to watch bettyboo and buildyboo will have a conversation | 22:32 |
@bettyboo | blackburn: oh, those 14 billions... i don't watch much news so i dont *know* but tyhere are soem people taking it sereously, and others probably aren't | 22:32 |
--- Log closed Wed Aug 03 00:00:34 2011 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!