| --- Log opened Fri Feb 15 00:00:46 2013 | ||
| -!- heiko [~heiko@027d8b72.bb.sky.com] has quit [Quit: Leaving.] | 00:02 | |
| n4nd0 | good night guys | 00:26 |
|---|---|---|
| -!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 00:26 | |
| -!- FSCV [~FSCV@204.45.110.163] has quit [Quit: Leaving] | 00:56 | |
| -!- blackburn [~blackburn@188.168.128.140] has quit [Quit: Leaving.] | 06:47 | |
| -!- pari [6eeaec2a@gateway/web/freenode/ip.110.234.236.42] has joined #shogun | 06:59 | |
| pari | Hi all, I am new to this community and am willing to start contributing to the toolbox. It would be very helpful if you could supply me with possible ideas/ areas to look into. Thanks in advance. | 07:06 |
| -!- blackburn [~lisitsin@mxs.kg.ru] has joined #shogun | 07:33 | |
| blackburn | pari: try to run compile and then try some examples, subscribe to the mailing list and check out ideas at github | 07:34 |
| blackburn | *remove 'run' word | 07:37 |
| -!- sumit [73f91219@gateway/web/freenode/ip.115.249.18.25] has joined #shogun | 07:55 | |
| @sonney2k | blackburn, what is wrong with the API? YOu could easily create SGMatrix etc from opencv matrices | 08:30 |
| @sonney2k | thanks to us using MALLOC & friends | 08:30 |
| @sonney2k | with zero overhead... | 08:30 |
| -!- pari [6eeaec2a@gateway/web/freenode/ip.110.234.236.42] has quit [Ping timeout: 245 seconds] | 08:31 | |
| blackburn | sonney2k: no, nothing wrong at all - I just have a feeling it could be better somehow | 08:31 |
| blackburn | sonney2k: although I have no idea how | 08:31 |
| @sonney2k | SGMatrix<float64_t>(opencvmat, m,n, false) | 08:31 |
| @sonney2k | then do whatever? | 08:31 |
| @sonney2k | how can it be more simple? | 08:32 |
| @sonney2k | only way I can think of is directly using their features | 08:32 |
| -!- hoijui [~hoijui@dslb-092-078-043-220.pools.arcor-ip.net] has joined #shogun | 08:38 | |
| blackburn | sonney2k: have you ever played to sid meier's alpha centauri? | 08:50 |
| sonne|work | no | 09:00 |
| -!- sumit [73f91219@gateway/web/freenode/ip.115.249.18.25] has quit [Ping timeout: 245 seconds] | 09:22 | |
| -!- n4nd0 [~nando@n145-p173.kthopen.kth.se] has joined #shogun | 09:45 | |
| sonne|work | n4nd0: do we have any news on the website? http://shogun-toolbox.org/page/Events/gsoc2013 looks awful now intermixed with the WS | 09:48 |
| n4nd0 | sonne|work: I started taking a look last week to the website code in github but nothing else | 09:49 |
| n4nd0 | I will focus on it from now on during my time shogun-ing | 09:50 |
| blackburn | sonne|work: what is ws? | 09:51 |
| n4nd0 | blackburn: workshop | 09:51 |
| blackburn | ah | 09:51 |
| -!- flx_ [~flx@fb.ml.tu-berlin.de] has joined #shogun | 10:37 | |
| -!- flx_ [~flx@fb.ml.tu-berlin.de] has left #shogun [] | 10:37 | |
| -!- flxb [~flx@fb.ml.tu-berlin.de] has joined #shogun | 10:40 | |
| -!- flxb [~flx@fb.ml.tu-berlin.de] has quit [Quit: leaving] | 10:47 | |
| -!- flxb [~flx@fb.ml.tu-berlin.de] has joined #shogun | 10:48 | |
| -!- flxb [~flx@fb.ml.tu-berlin.de] has quit [Ping timeout: 244 seconds] | 11:25 | |
| -!- flxb [~flx@fb.ml.tu-berlin.de] has joined #shogun | 12:03 | |
| -!- flxb [~flx@fb.ml.tu-berlin.de] has quit [Ping timeout: 244 seconds] | 12:34 | |
| -!- heiko [~heiko@nat-164-235.internal.eduroam.ucl.ac.uk] has joined #shogun | 12:56 | |
| -!- n4nd0 [~nando@n145-p173.kthopen.kth.se] has quit [Ping timeout: 252 seconds] | 13:15 | |
| -!- flxb [~flx@fb.ml.tu-berlin.de] has joined #shogun | 13:20 | |
| -!- n4nd0 [~nando@n145-p173.kthopen.kth.se] has joined #shogun | 13:30 | |
| -!- n4nd0 [~nando@n145-p173.kthopen.kth.se] has quit [Client Quit] | 13:30 | |
| -!- n4nd0_ [~nando@n145-p173.kthopen.kth.se] has joined #shogun | 13:30 | |
| -!- n4nd0_ is now known as n4nd0 | 14:19 | |
| n4nd0 | blackburn: Georg has done a very nice job! | 14:20 |
| -!- hoijui [~hoijui@dslb-092-078-043-220.pools.arcor-ip.net] has quit [Quit: Leaving] | 14:23 | |
| blackburn | n4nd0: yeah | 14:26 |
| -!- tom___ [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has joined #shogun | 14:27 | |
| tom___ | hi | 14:27 |
| n4nd0 | blackburn: so what do you think of his main idea? the first one listed in the gdocs | 14:27 |
| n4nd0 | hey tom___ | 14:27 |
| tom___ | I would like some precision about output format in shogun (hdf5, json, ascii .. ) | 14:28 |
| tom___ | which one is more mature ? which one would you recommand , | 14:28 |
| n4nd0 | tom___: what do you want to do? | 14:28 |
| tom___ | saving/loading multiclass machine, features, labels .. | 14:28 |
| n4nd0 | I guess it depends a bit on your preferences, what you feel most comfortable with | 14:29 |
| tom___ | I am running with c++ api and would like to plot some results with python then save in c++ load in python | 14:29 |
| n4nd0 | if you are going to use large volumes of data, then you should go for hdf5 maybe | 14:29 |
| n4nd0 | AFAIK it should be the most efficient of those | 14:30 |
| tom___ | size of ascii files seems smaller than with hdf5 ? possible ? | 14:30 |
| n4nd0 | if you are not storing lot of data, it might be possible | 14:30 |
| tom___ | mmmh | 14:31 |
| tom___ | both format are compatible with the python api ? | 14:31 |
| n4nd0 | tom___: using shogun methods you mean? | 14:32 |
| tom___ | yes! | 14:32 |
| n4nd0 | yeah, they should be | 14:32 |
| n4nd0 | have you taken a look to python examples? | 14:32 |
| n4nd0 | let me see | 14:32 |
| tom___ | what format do you use yourself ? :) | 14:32 |
| n4nd0 | I have not really used them for large applications | 14:34 |
| n4nd0 | just small tests and IIRC I used ascii | 14:34 |
| n4nd0 | tom___: I see that features_io_modular.py | 14:35 |
| n4nd0 | under examples/undocumented/python/ | 14:35 |
| n4nd0 | uses the three formats | 14:35 |
| n4nd0 | ascii, binary and hdf5 | 14:35 |
| n4nd0 | maybe it can help you decide which one you want to select | 14:36 |
| tom___ | allright thanks for advices | 14:36 |
| tom___ | have a nice day | 14:36 |
| n4nd0 | tom___: thanks, you too | 14:37 |
| -!- tom___ [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has quit [Quit: Page closed] | 14:37 | |
| -!- FSCV [~FSCV@187.210.54.166] has joined #shogun | 14:42 | |
| -!- FSCV_ [~FSCV@204.45.110.163] has joined #shogun | 14:43 | |
| -!- FSCV [~FSCV@187.210.54.166] has quit [Ping timeout: 252 seconds] | 14:47 | |
| -!- blackburn [~lisitsin@mxs.kg.ru] has quit [Quit: Leaving.] | 14:59 | |
| -!- flxb [~flx@fb.ml.tu-berlin.de] has quit [Ping timeout: 264 seconds] | 15:26 | |
| -!- lambday [b6131049@gateway/web/freenode/ip.182.19.16.73] has joined #shogun | 15:48 | |
| -!- n4nd0 [~nando@n145-p173.kthopen.kth.se] has quit [Quit: leaving] | 16:01 | |
| -!- flxb [~flx@fb.ml.tu-berlin.de] has joined #shogun | 16:14 | |
| sonne|work | http://www.youtube.com/watch?feature=player_embedded&v=VFGxWVZ7tgo | 16:40 |
| sonne|work | ^ blackburn - such stuff can only happen in rossia :) | 16:41 |
| heiko | man, I hope blackburn is still alive :) | 16:52 |
| sonne|work | heiko: do you use eclipse and c++? | 17:03 |
| heiko | sonne|work, yes, why? | 17:03 |
| wiking | http://www.youtube.com/watch?feature=player_embedded&v=OaREmEdXZe4 | 17:03 |
| wiking | :D | 17:03 |
| heiko | wiking thats fake | 17:03 |
| wiking | nice photoshop :D | 17:04 |
| -!- flxb [~flx@fb.ml.tu-berlin.de] has quit [Ping timeout: 248 seconds] | 17:09 | |
| -!- sonne|work [~sonnenbu@sams-office-nat.tomtomgroup.com] has quit [Quit: Leaving.] | 17:26 | |
| -!- heiko [~heiko@nat-164-235.internal.eduroam.ucl.ac.uk] has quit [Quit: Leaving.] | 17:26 | |
| -!- heiko [~heiko@nat-164-235.internal.eduroam.ucl.ac.uk] has joined #shogun | 17:27 | |
| -!- KMcQuisten [d8338942@gateway/web/freenode/ip.216.51.137.66] has joined #shogun | 17:53 | |
| KMcQuisten | Is there a set way to make the shogun wrapper of LibSVR perform nu-SVR, or is it hardwired to do epsilon-SVR? | 18:12 |
| -!- alexlovesdata [~binder@2001:638:806:e001:eda1:1d04:97b:c9f] has joined #shogun | 18:17 | |
| -!- FSCV_ [~FSCV@204.45.110.163] has quit [Ping timeout: 245 seconds] | 18:25 | |
| -!- heiko [~heiko@nat-164-235.internal.eduroam.ucl.ac.uk] has quit [Quit: Leaving.] | 18:25 | |
| -!- FSCV [~FSCV@187.210.54.166] has joined #shogun | 18:27 | |
| KMcQuisten | Anyone alive, or are you all sleeping on the other side of the world? | 18:39 |
| -!- FSCV_ [~FSCV@173.254.212.46] has joined #shogun | 18:39 | |
| alexlovesdata | here | 18:39 |
| -!- FSCV [~FSCV@187.210.54.166] has quit [Ping timeout: 248 seconds] | 18:42 | |
| -!- KMcQuisten [d8338942@gateway/web/freenode/ip.216.51.137.66] has quit [Quit: Page closed] | 18:51 | |
| -!- heiko [~heiko@nat-164-235.internal.eduroam.ucl.ac.uk] has joined #shogun | 18:55 | |
| -!- blackburn [~blackburn@85.114.170.181] has joined #shogun | 19:07 | |
| heiko | blackburn, does it make sense to allow the tube epsilon for model selection? | 19:08 |
| heiko | and epsilon? | 19:08 |
| blackburn | heiko: I remember that question! | 19:08 |
| heiko | see email | 19:08 |
| blackburn | heiko: yeah | 19:09 |
| heiko | so? | 19:09 |
| heiko | just add this? | 19:09 |
| blackburn | heiko: okay it won't hurt right? | 19:10 |
| heiko | no, its just: does it make sense? | 19:10 |
| blackburn | heiko: I think it can be considered | 19:10 |
| blackburn | just some more flexibility | 19:10 |
| heiko | ok then, here we go | 19:11 |
| blackburn | wiking: about video - it is not fake but it was in tajikistan :D | 19:11 |
| blackburn | not that meteorite thing | 19:11 |
| blackburn | heiko: ah you said it is fake - it is not :) | 19:12 |
| heiko | blackburn, really? | 19:12 |
| heiko | this was said on a german news webiste | 19:12 |
| heiko | glad you still live :) | 19:12 |
| blackburn | heiko: it is some gas fire | 19:12 |
| heiko | we would be screwed without you | 19:12 |
| heiko | o | 19:12 |
| blackburn | heiko: fake is that it was caused by meteorite | 19:12 |
| heiko | yeah thats what I meant, not the meteor | 19:12 |
| blackburn | it happened before in other country | 19:12 |
| blackburn | but actually it happened :) | 19:13 |
| alexlovesdata | good the blackburn did not loose too much of his brain due to meteorite strike :D | 19:13 |
| -!- shogun-notifier- [~irker@7nn.de] has joined #shogun | 19:13 | |
| shogun-notifier- | shogun: Heiko Strathmann :master * 9f1daba / src/shogun/classifier/svm/SVM.cpp: https://github.com/shogun-toolbox/shogun/commit/9f1daba1d32cc786614f64f5fd7088bd2b88aa3b | 19:13 |
| shogun-notifier- | shogun: added epsilon and tube_epsilon for model selection | 19:13 |
| shogun-notifier- | shogun: Heiko Strathmann :master * 83d3fbe / src/shogun/classifier/svm/SVM.cpp: https://github.com/shogun-toolbox/shogun/commit/83d3fbed74dc4f9fd77f1a968b5f4cef4280e548 | 19:13 |
| shogun-notifier- | shogun: Merge pull request #877 from karlnapf/master | 19:13 |
| shogun-notifier- | shogun: | 19:13 |
| shogun-notifier- | shogun: added epsilon and tube_epsilon for model selection | 19:13 |
| blackburn | alexlovesdata: pretty far away from me | 19:13 |
| blackburn | you know, close is 1000 km for me | 19:13 |
| blackburn | it is somewhat 1500 km far away I think :D | 19:13 |
| heiko | haha | 19:13 |
| blackburn | alexlovesdata: I have lost my brain due to vodka though! | 19:14 |
| blackburn | alexlovesdata: http://cs14112.userapi.com/c7005/v7005319/3d3/xD8UN_by11E.jpg see mr. poutine riding | 19:16 |
| blackburn | http://s.pikabu.ru/post_img/2013/02/15/8/1360930135_841226929.jpg that's how russian tv looks like | 19:18 |
| alexlovesdata | uhh | 19:18 |
| alexlovesdata | WW3 welcome :| | 19:19 |
| alexlovesdata | do we have some hidden long range cannon under the north polar ice cap? | 19:20 |
| alexlovesdata | do the russians operate devices like this?? | 19:20 |
| blackburn | alexlovesdata: russians exchanged everything to vodka | 19:20 |
| blackburn | :D | 19:20 |
| blackburn | alexlovesdata: ahh I was searching for this pic to show to you | 19:22 |
| blackburn | http://s.pikabu.ru/post_img/2013/02/15/5/1360911381_1738019492.jpg | 19:22 |
| blackburn | alexlovesdata: translating: chelyabinsk (the place hit by that stone) before stone shower | 19:22 |
| blackburn | and chelyabinsk after stone shower | 19:23 |
| blackburn | heiko: ha! have you seen that mail? local alignment kernel as gsoc project | 19:24 |
| heiko | what? | 19:24 |
| heiko | blackburn, good idea | 19:25 |
| heiko | although most of them should be there | 19:25 |
| heiko | its just a mess to use | 19:25 |
| blackburn | heiko: most of what? | 19:25 |
| heiko | would be good to cleanup and have examples and proper documentation | 19:25 |
| heiko | string kernels | 19:25 |
| blackburn | ah yes | 19:25 |
| blackburn | but not only one kernel | 19:25 |
| blackburn | heiko: who can mentor that? | 19:26 |
| heiko | s?ren :D | 19:26 |
| blackburn | heiko: anyone else? :D | 19:27 |
| heiko | maybe oli stegle? | 19:27 |
| heiko | but he is busy in Cambridge now :=) | 19:27 |
| heiko | dont know | 19:27 |
| heiko | but would be good to have | 19:27 |
| heiko | since many biologists out there | 19:28 |
| heiko | and its an easy project | 19:28 |
| blackburn | the more ideas we have the more difficult it would be to choose | 19:28 |
| heiko | so far most of the projects are more advanced | 19:28 |
| blackburn | but the final choice should get better too | 19:28 |
| blackburn | yes I want to improve biological part of shogun | 19:28 |
| blackburn | as I don't solve any biological problems I do not understand anything | 19:29 |
| heiko | I know a little bit about these kernels so I could help out a bit if e.g. the main mentor is too busy | 19:30 |
| heiko | but would be better to have some expert | 19:30 |
| blackburn | yes | 19:30 |
| blackburn | may be I should ask chris | 19:30 |
| blackburn | he should know something :D | 19:30 |
| heiko | yes, I mean they are close to this | 19:31 |
| alexlovesdata | I wonder a bit ... | 19:33 |
| blackburn | heiko: have you seen they moved gsoc thingy a month forward? | 19:33 |
| heiko | not seen | 19:33 |
| blackburn | it ends sept. 27 | 19:33 |
| heiko | what means forward? | 19:33 |
| heiko | to later | 19:33 |
| alexlovesdata | can we optimize large scale primal svms when the features do not fit into memory? | 19:33 |
| heiko | uh | 19:33 |
| alexlovesdata | I know Cstreamingfeatures anf vowpal wabbit | 19:33 |
| blackburn | heiko: and it starts a bit later | 19:33 |
| blackburn | heiko: and I won't receive t-shirt :D | 19:33 |
| alexlovesdata | I was thinking: what strategies do we have for that mega large scale case? | 19:34 |
| blackburn | due to rules | 19:34 |
| blackburn | alexlovesdata: ultra mega scale | 19:34 |
| blackburn | alexlovesdata: good q | 19:34 |
| blackburn | alexlovesdata: we have outdated vowpal wabbit though | 19:34 |
| -!- sumit [73f91219@gateway/web/freenode/ip.115.249.18.25] has joined #shogun | 20:30 | |
| -!- lambday [b6131049@gateway/web/freenode/ip.182.19.16.73] has quit [Ping timeout: 245 seconds] | 20:51 | |
| -!- sumit [73f91219@gateway/web/freenode/ip.115.249.18.25] has quit [Ping timeout: 245 seconds] | 20:51 | |
| shogun-notifier- | shogun: Heiko Strathmann :master * 2e37cd5 / / (11 files): https://github.com/shogun-toolbox/shogun/commit/2e37cd5beea2fc7f827611e4155b8025fc11bba2 | 20:58 |
| shogun-notifier- | shogun: Made new kernel selection methods for linear time MMD available under modular interfaces. | 20:58 |
| shogun-notifier- | shogun: Includes some other minor changes | 20:58 |
| shogun-notifier- | shogun: Heiko Strathmann :master * 17d6ba9 / / (11 files): https://github.com/shogun-toolbox/shogun/commit/17d6ba9eb5fea2d38673b03ef28c72f81ffe425f | 20:58 |
| shogun-notifier- | shogun: Merge pull request #878 from karlnapf/master | 20:58 |
| shogun-notifier- | shogun: | 20:58 |
| shogun-notifier- | shogun: kernel selection for linear time mmd for modular interfaces | 20:58 |
| -!- heiko [~heiko@nat-164-235.internal.eduroam.ucl.ac.uk] has left #shogun [] | 20:58 | |
| @sonney2k | alexlovesdata, hey there | 22:23 |
| alexlovesdata | hey sonney2k | 22:23 |
| alexlovesdata | I am thinking of an MKL extension on cross-validation data | 22:24 |
| alexlovesdata | so that it works beter in practice | 22:24 |
| alexlovesdata | on Sunday I consider to write some KDA for shogun ... once I got behing ther git monster | 22:25 |
| @sonney2k | alexlovesdata, about the after workshop happening | 22:26 |
| alexlovesdata | yep ... what will be the input to Andrea? | 22:26 |
| @sonney2k | could you or someone from the TU people take care of the crowd? | 22:26 |
| alexlovesdata | yes I could do | 22:26 |
| @sonney2k | like AZ | 22:26 |
| alexlovesdata | andreas Z ? | 22:27 |
| @sonney2k | or if chris is around (he is officially TU right?) | 22:27 |
| @sonney2k | yes | 22:27 |
| @sonney2k | I am away sadly :/ | 22:27 |
| alexlovesdata | I can sk AZ whether he likes to join in helping me ... but what is the motivation of AZ to help me? | 22:27 |
| alexlovesdata | Chris will help because he is involved in it clearly | 22:27 |
| @sonney2k | alexlovesdata, he will mentor ICA for shogun anyways | 22:27 |
| alexlovesdata | ahhh , ok | 22:27 |
| @sonney2k | so it might get him connected to the crowd? | 22:28 |
| @sonney2k | he might like it | 22:28 |
| alexlovesdata | ye ... I can ask him with the pretext that he will anyway mentor it | 22:28 |
| alexlovesdata | any special expectations about taking care of the crowd? | 22:29 |
| @sonney2k | I guess you only need some official who is around from beginning to end on the weekend | 22:29 |
| @sonney2k | you could probably split this up | 22:29 |
| alexlovesdata | should be doable | 22:29 |
| @sonney2k | I would have organized that thing more early but we wanted it to be as close as possible to ISMB | 22:30 |
| @sonney2k | (and me around) might not be close enough but nevertheless | 22:30 |
| @sonney2k | alexlovesdata, for ultra mega large scale we use mean(x) > threshold | 22:31 |
| @sonney2k | btw ;) | 22:31 |
| alexlovesdata | how does mean(X) | 22:31 |
| alexlovesdata | > threshold help in that? | 22:31 |
| @sonney2k | alexlovesdata, ultra mega largescale boils down to computing *simple* statistics | 22:32 |
| @sonney2k | and simple rules | 22:32 |
| @sonney2k | no more | 22:32 |
| @sonney2k | everything else is beyond ... | 22:32 |
| @sonney2k | but yes we don't have anything distributed | 22:32 |
| @sonney2k | as in map-reduce & friends | 22:32 |
| alexlovesdata | hmmm ... that could be something for GSOC ;) | 22:32 |
| @sonney2k | so we can only realisticly handle a billion of examples | 22:32 |
| alexlovesdata | where does the limit come from? | 22:33 |
| @sonney2k | actually there is no limit | 22:33 |
| @sonney2k | it is just that you need a fast harddisk to stream data | 22:33 |
| alexlovesdata | svm optimization could be parallelized in principle | 22:34 |
| @sonney2k | with streaming features you can just ask: 'give me the next example' | 22:34 |
| @sonney2k | alexlovesdata, yeah but that is not the bottleneck | 22:34 |
| @sonney2k | bottleneck is reading from disk | 22:34 |
| alexlovesdata | well ... fast hard disks are no problem since the arrival of ... ;) | 22:34 |
| @sonney2k | no | 22:34 |
| @sonney2k | they are dog slow | 22:34 |
| @sonney2k | you can do thousands of svm iterations while reading the next example from disk | 22:34 |
| alexlovesdata | SSDs?? | 22:35 |
| @sonney2k | so mini-batch algorithms are fast | 22:35 |
| @sonney2k | yes even ssds are dog slow | 22:35 |
| alexlovesdata | but one could work on distrobuted batches | 22:35 |
| @sonney2k | it is like <300MB/s vs. 10GB/s main memory speed vs. 1000GB/s cache speed | 22:35 |
| alexlovesdata | the problem is an interface that is flexible to a lot of ways of parallelization | 22:36 |
| @sonney2k | that would just give you cache misses | 22:36 |
| -!- sumit [75e10b98@gateway/web/freenode/ip.117.225.11.152] has joined #shogun | 22:36 | |
| alexlovesdata | I would not dare to comentor GPU svms | 22:36 |
| @sonney2k | but I agree there is lots of stuff one could do | 22:36 |
| @sonney2k | GPU svms are overhyped | 22:36 |
| @sonney2k | they can be *only fast* if all data fits in GPU memory | 22:36 |
| alexlovesdata | yep ... | 22:37 |
| @sonney2k | so lets say you have 2 G of GPU mem | 22:37 |
| alexlovesdata | that expeerience I had with clustering on GPUs already | 22:37 |
| @sonney2k | that woudl be a tiny problem | 22:37 |
| @sonney2k | so it is useless | 22:37 |
| alexlovesdata | but distributed SVMs is a thing we can try | 22:37 |
| @sonney2k | considering that I was solving 60G problems with SVMs 5 years ago | 22:37 |
| @sonney2k | I would say linear SVMs with piece wise linear functions is the thing you should use | 22:38 |
| alexlovesdata | piece wise linear on what? | 22:38 |
| alexlovesdata | you mean locally linear SVMs ? | 22:38 |
| @sonney2k | with shogun these piece wise linear functions can be implicilty used so you can have a svm-normal vector of a few GB but only say 100 dimensional input data | 22:38 |
| @sonney2k | no linear SVMs but you use binary features: | 22:39 |
| @sonney2k | look at http://www.shogun-toolbox.org/doc/en/2.0.1/classshogun_1_1CBinnedDotFeatures.html | 22:39 |
| @sonney2k | got to leave the train | 22:40 |
| alexlovesdata | why losing information with binarization? | 22:40 |
| @sonney2k | you hardly do | 22:40 |
| @sonney2k | l8r | 22:40 |
| alexlovesdata | is there any paper on that to udnerstand it? | 22:41 |
| alexlovesdata | ok I understand it | 22:42 |
| alexlovesdata | roughly ... not fully | 22:43 |
| -!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 22:49 | |
| @sonney2k | alexlovesdata, I got some pretty good results with that and extremely fast... it is more expressive than linear - just depending on the number of bins | 22:57 |
| @sonney2k | alexlovesdata, what I would hope is that fastfood (smolas paper) would give us sth even better w/ kernels | 22:59 |
| alexlovesdata | I have to read the fast food paper | 23:00 |
| alexlovesdata | well ... for the most common kernel ... linear | 23:00 |
| alexlovesdata | RF features always allow to use primal solveras | 23:00 |
| alexlovesdata | similar developments exist for other kernels ... see the paper by vedaldi | 23:00 |
| @sonney2k | RF? | 23:00 |
| alexlovesdata | random fourier features | 23:02 |
| alexlovesdata | one needs to be careful about kernel weight | 23:02 |
| alexlovesdata | but one can get a transform such that the transformed features approximate a gaissuan kernel | 23:03 |
| n4nd0 | sonney2k: you won't be around during the workshop? :( | 23:08 |
| @sonney2k | n4nd0, I will be around during workshop but not on saturday/sunday in the hands-on session :/ | 23:16 |
| @sonney2k | alexlovesdata, that you can do such transformation is not sufficient | 23:16 |
| n4nd0 | oh that's a pity :( | 23:16 |
| @sonney2k | alexlovesdata, you cannot precompute these features | 23:16 |
| n4nd0 | sonney2k: but we will see you during the workshop :) | 23:16 |
| @sonney2k | alexlovesdata, you have to do that on-the-fly and very very fast | 23:16 |
| @sonney2k | otherwise you gain nothing | 23:16 |
| alexlovesdata | you just need to save the generating coefficients | 23:17 |
| alexlovesdata | then RF acts as a preprocessors | 23:17 |
| alexlovesdata | which transforms on the fly | 23:17 |
| alexlovesdata | or do you see any problems about that? | 23:17 |
| @sonney2k | alexlovesdata, no - I am just thinking that this would be some easy student task to prepare for gsoc | 23:27 |
| @sonney2k | http://berkeley.intel-research.net/arahimi/papers/rahimi-recht-random-features.pdf | 23:27 |
| @sonney2k | implement algorithm 1 here | 23:27 |
| @sonney2k | just a a shogun preprocessor | 23:28 |
| alexlovesdata | it is implemented alreade :D | 23:28 |
| @sonney2k | for the gaussian kernel | 23:28 |
| @sonney2k | alexlovesdata, what? | 23:28 |
| @sonney2k | name? | 23:28 |
| alexlovesdata | wait ... I find the name of it in shogun | 23:28 |
| @sonney2k | ohh randomfouriergausspreproc :D | 23:28 |
| alexlovesdata | http://www.shogun-toolbox.org/doc/en/2.0.1/classshogun_1_1CRandomFourierGaussPreproc.html | 23:28 |
| @sonney2k | obviously | 23:28 |
| alexlovesdata | hahaha, yes | 23:28 |
| @sonney2k | haha | 23:29 |
| alexlovesdata | :) | 23:29 |
| @sonney2k | stso then randombinning too :D | 23:29 |
| n4nd0 | nobody really knows everything in SHOGUN :D | 23:29 |
| alexlovesdata | stso = ? | 23:29 |
| alexlovesdata | shogun ... the dark dungeons | 23:29 |
| n4nd0 | hehe | 23:30 |
| * sonney2k is impressed | 23:30 | |
| * n4nd0 too | 23:30 | |
| @sonney2k | I guess I just use 10% of shogun | 23:30 |
| alexlovesdata | most of us use 0.5% I would say | 23:30 |
| alexlovesdata | anyway it might be useful to read about fastfood | 23:31 |
| alexlovesdata | I just want to get this tourine here going and then drive home | 23:31 |
| alexlovesdata | tourine =routine | 23:32 |
| @sonney2k | alexlovesdata, I think this preprocessor should better be defined as DotFeatures... then any linear SVM could directly work on this | 23:35 |
| alexlovesdata | you are right ... I could do that on Sunday | 23:35 |
| @sonney2k | would be faster than preprocessing all the time | 23:35 |
| alexlovesdata | or try ;) | 23:35 |
| @sonney2k | alexlovesdata, btw where did you do the FFT? | 23:36 |
| alexlovesdata | + write an example which compares against a real gaussian on small scale data | 23:36 |
| @sonney2k | and where is the sin() ? | 23:36 |
| alexlovesdata | you need only the cos ... if I remember correctly | 23:37 |
| @sonney2k | int the paper above they have sin(w'x) and cos(w'x) | 23:38 |
| alexlovesdata | if I remember correctly there was a way to get around the sin ... | 23:40 |
| alexlovesdata | I will check that on sunday | 23:40 |
| alexlovesdata | below equation (2) in the paper | 23:42 |
| alexlovesdata | other mappings such as ... | 23:42 |
| alexlovesdata | also satisfy the condition ... | 23:42 |
| alexlovesdata | there is a mapping cos(wx +b) | 23:43 |
| alexlovesdata | that I had implemented | 23:43 |
| @sonney2k | sqrt(2) cos(wx+b) ok | 23:44 |
| @sonney2k | alexlovesdata, I think in fast food they have RF and also random hadamart (sp?) features | 23:45 |
| @sonney2k | but this stuff really only kicks ass if we have it as dot features | 23:45 |
| alexlovesdata | hadamard ... could be | 23:45 |
| alexlovesdata | I will take care of it on sunday for dot features | 23:45 |
| alexlovesdata | and of an example | 23:46 |
| alexlovesdata | ... will do the example first | 23:46 |
| @sonney2k | but still consider we need 1000 w's for the approximation - it might actually be very expensive compared to learning piece-wise linear functions | 23:46 |
| @sonney2k | but I never did care enough to compare this | 23:47 |
| alexlovesdata | it blows up dimensions, though 100 was enough when i tested it | 23:47 |
| alexlovesdata | thats why I asked abt strategies when the features do not fit into mem | 23:47 |
| @sonney2k | alexlovesdata, with dot features you don't need factor 100 more memory | 23:48 |
| @sonney2k | but just the SVM-w will be 100 times bigger | 23:48 |
| @sonney2k | that can be very fast - since also very cache-friendly | 23:48 |
| @sonney2k | maybe we can broaden the fastfood project in this respect | 23:52 |
| -!- sumit [75e10b98@gateway/web/freenode/ip.117.225.11.152] has quit [Ping timeout: 245 seconds] | 23:53 | |
| alexlovesdata | I go home | 23:54 |
| alexlovesdata | will continue from home :| | 23:54 |
| @sonney2k | alexlovesdata, see you! | 23:54 |
| -!- alexlovesdata [~binder@2001:638:806:e001:eda1:1d04:97b:c9f] has left #shogun [] | 23:55 | |
| @sonney2k | traceroute 216.81.59.173 | 23:57 |
| @sonney2k | *lol* | 23:57 |
| -!- shogun-notifier- [~irker@7nn.de] has quit [Quit: transmission timeout] | 23:58 | |
| n4nd0 | it got lost after the fifth hop here | 23:59 |
| --- Log closed Sat Feb 16 00:00:14 2013 | ||
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!
