--- Log opened Sun Jun 02 00:00:27 2013 | ||
-!- hushell [~hushell@c-67-189-100-116.hsd1.or.comcast.net] has left #shogun ["WeeChat 0.3.7"] | 04:51 | |
-!- zxtx [~zv@cpe-75-83-151-252.socal.res.rr.com] has joined #shogun | 07:41 | |
-!- Netsplit *.net <-> *.split quits: flxb_ | 08:10 | |
-!- Netsplit over, joins: flxb_ | 08:21 | |
-!- nube [~rho@49.244.74.63] has quit [Ping timeout: 276 seconds] | 08:33 | |
-!- Netsplit *.net <-> *.split quits: flxb_ | 08:36 | |
-!- nube [~rho@49.244.80.228] has joined #shogun | 08:47 | |
-!- Netsplit over, joins: flxb_ | 08:49 | |
-!- lambday [67157d37@gateway/web/cgi-irc/kiwiirc.com/ip.103.21.125.55] has joined #shogun | 09:04 | |
lambday | good morning sonney2k | 09:16 |
---|---|---|
lambday | शुभ प्रभात lisitsyn :P | 09:16 |
-!- gsomix [~gsomix@83.234.54.61] has joined #shogun | 09:19 | |
gsomix | good mornning | 09:19 |
-!- sanyam [uid10602@gateway/web/irccloud.com/x-cgbsdbswjbrxmjzw] has quit [Quit: Connection closed for inactivity] | 09:22 | |
lambday | gsomix: morning :) | 09:27 |
-!- foulwall [~foulwall@2001:da8:215:503:c5a2:4030:17a3:c8ce] has joined #shogun | 09:39 | |
@lisitsyn | lambday: yes lambday exactly! | 09:43 |
gsomix | lisitsyn, hi, listing. | 09:46 |
lambday | lisitsyn: :P | 09:49 |
@lisitsyn | gsomix: hi mogsix | 09:49 |
gsomix | lisitsyn, lol http://www.freelancer.com/u/mogsix.html | 09:51 |
@lisitsyn | shalom | 09:52 |
* gsomix sings "hava nagila, hava nagila" | 09:59 | |
-!- iglesiasg [d58f3264@gateway/web/freenode/ip.213.143.50.100] has joined #shogun | 10:20 | |
-!- mode/#shogun [+o iglesiasg] by ChanServ | 10:21 | |
-!- foulwall [~foulwall@2001:da8:215:503:c5a2:4030:17a3:c8ce] has quit [Remote host closed the connection] | 10:31 | |
gsomix | sonney2k, around? | 10:32 |
-!- vgorbati [~vgorbati@212.2.159.34] has joined #shogun | 10:48 | |
-!- ipondering [~iponderin@202.112.128.135] has joined #shogun | 10:57 | |
-!- ipondering [~iponderin@202.112.128.135] has quit [Client Quit] | 11:01 | |
-!- ipondering [~iponderin@202.112.128.135] has joined #shogun | 11:02 | |
vgorbati | lisitsyn: hello, present? | 11:17 |
@lisitsyn | vgorbati: yes | 11:17 |
vgorbati | lisitsyn: well, at least on the datasets I had, the algorithm performed better compared to simple NN and GMDH:) I guess I will use it as a baseline - if some very complex algorithm, that involves a lot of cool math, has worst performance than this one - then I probably won't use it:) | 11:21 |
@lisitsyn | vgorbati: the one with derivatives? | 11:21 |
vgorbati | lisitsyn: yes | 11:21 |
@lisitsyn | vgorbati: good | 11:22 |
vgorbati | lisitsyn: it usually does not grow more then the 1st derivative, but this is still enough to beat the cool NN's and GMDH:) (for the data I has) | 11:23 |
@lisitsyn | vgorbati: what's NN you compare to? | 11:24 |
vgorbati | lisitsyn: multi-layer perceptron, built-in matlab | 11:24 |
@lisitsyn | I see | 11:24 |
vgorbati | lisitsyn: it is pretty good there, they use levenberg-marquardt training method, logistic hidden layer, random division of data set and all that stuff | 11:26 |
vgorbati | lisitsyn: by default | 11:26 |
@lisitsyn | vgorbati: I see | 11:26 |
@lisitsyn | vgorbati: that's good for you then :) | 11:26 |
vgorbati | lisitsyn: :) btw, if, say, I have such parameters of SVM as: support vectors, weight vector, bias, kernel function, scaling parameters and classes for support vectors, then I can calc output as: | 11:29 |
vgorbati | lisitsyn: sign(sum_over_i(w_i*output_i*kernel_func(supp_vec_i, scaling(input_vector))) - bias) | 11:29 |
vgorbati | right? | 11:29 |
@lisitsyn | what is output_i? | 11:30 |
vgorbati | -1 or 1 for support vector at index i | 11:30 |
@lisitsyn | to avoid confusion rather use \alpha_i to denote support vector coefficient | 11:30 |
vgorbati | but is the overall formula right? | 11:31 |
@lisitsyn | \sign (\sum_i \alpha_i k(x_i, x_input) - b) | 11:31 |
@lisitsyn | there is no need to put a sign of support vector there actually | 11:31 |
@lisitsyn | they can go to alphas | 11:31 |
@lisitsyn | vgorbati: http://latex.codecogs.com/gif.latex?f(x_{input})%20=%20\mathrm{sign}%20\left(\sum_i%20\alpha_i%20k(x_i,%20x_{input})%20-%20b\right) | 11:32 |
@lisitsyn | vgorbati: I mean it is conventional that alphas are to be signed | 11:34 |
@lisitsyn | vgorbati: I guess you might have seen such notation in vorontsov's lectures | 11:35 |
@lisitsyn | it is more strict actually | 11:35 |
@lisitsyn | as alphas are lagrange multipliers and they are positive by definition | 11:35 |
@lisitsyn | non-negative i mean | 11:36 |
vgorbati | lisitsyn: yeah, I got it, thanks very much:) for better performance (forecasting performance I mean), I tried to get not the class label (-1 or 1) from svmclassify, but the actual value under the sign function - and I failed at it, because matlab probably makes it the way you have described - it stores alpha's signed, and when I also multiplied by class_i, I got totally crap results) | 11:37 |
@lisitsyn | vgorbati: hmm why do you use {-1,1}? | 11:38 |
vgorbati | lisitsyn: what is wrong with -1 or 1? | 11:38 |
@lisitsyn | vgorbati: SVM is easily transformed to SVR | 11:38 |
@lisitsyn | where inputs are real-valued | 11:38 |
@lisitsyn | vgorbati: you predict real values right? | 11:39 |
vgorbati | lisitsyn: SVM's predict only the sign of the derivative | 11:39 |
@lisitsyn | ah | 11:39 |
vgorbati | lisitsyn: so, -1 or 1 is actually quite suited for this purpose | 11:39 |
vgorbati | lisitsyn: I then just multiply by it | 11:39 |
@lisitsyn | I see | 11:40 |
@lisitsyn | yes then it is ok | 11:40 |
gsomix | sonney2k, so, CircularBuffer and LineReader are testing now. line counter takes ~3.3 sec for data/asp/Human.dat | 11:40 |
gsomix | sonney2k, docs and clean up are needed for code. I'll send PR at evening. | 11:41 |
gsomix | sonney2k, now I want try to prepare for pass-fail exam - digital signal processing. cu later. | 11:44 |
-!- gsomix [~gsomix@83.234.54.61] has quit [Quit: Leaving] | 11:44 | |
-!- lambday [67157d37@gateway/web/cgi-irc/kiwiirc.com/ip.103.21.125.55] has quit [Quit: http://www.kiwiirc.com/ - A hand crafted IRC client] | 11:44 | |
-!- ipondering [~iponderin@202.112.128.135] has quit [Quit: ipondering] | 11:46 | |
-!- gsomix [~Miranda@83.234.54.61] has joined #shogun | 11:47 | |
-!- vgorbati_ [~vgorbati@212.2.159.34] has joined #shogun | 11:53 | |
-!- vgorbati [~vgorbati@212.2.159.34] has quit [Ping timeout: 246 seconds] | 11:55 | |
-!- vgorbati_ is now known as vgorbati | 11:55 | |
-!- vgorbati [~vgorbati@212.2.159.34] has quit [Read error: Connection reset by peer] | 12:43 | |
-!- vgorbati [~vgorbati@212.2.159.34] has joined #shogun | 12:43 | |
-!- lambday [67157d37@gateway/web/cgi-irc/kiwiirc.com/ip.103.21.125.55] has joined #shogun | 13:02 | |
lambday | there are some files (SGSparseVector.cpp, SGMatrix.cpp) which misses the GPL header | 13:16 |
lambday | I'm adding complex64_t support for those.. should I add the header too? | 13:17 |
-!- vgorbati [~vgorbati@212.2.159.34] has quit [Read error: Connection reset by peer] | 13:21 | |
-!- vgorbati [~vgorbati@212.2.159.34] has joined #shogun | 13:22 | |
-!- vgorbati_ [~vgorbati@212.2.159.34] has joined #shogun | 13:35 | |
-!- vgorbati [~vgorbati@212.2.159.34] has quit [Ping timeout: 264 seconds] | 13:36 | |
-!- vgorbati_ is now known as vgorbati | 13:36 | |
-!- iglesiasg [d58f3264@gateway/web/freenode/ip.213.143.50.100] has quit [Ping timeout: 250 seconds] | 14:04 | |
@lisitsyn | lambday: yeah please add one | 14:54 |
-!- lambday [67157d37@gateway/web/cgi-irc/kiwiirc.com/ip.103.21.125.55] has quit [Quit: http://www.kiwiirc.com/ - A hand crafted IRC client] | 15:01 | |
-!- vgorbati_ [~vgorbati@212.2.159.34] has joined #shogun | 15:58 | |
-!- vgorbati [~vgorbati@212.2.159.34] has quit [Ping timeout: 252 seconds] | 15:59 | |
-!- vgorbati_ is now known as vgorbati | 15:59 | |
-!- vgorbati_ [~vgorbati@212.2.159.34] has joined #shogun | 16:32 | |
-!- vgorbati [~vgorbati@212.2.159.34] has quit [Ping timeout: 245 seconds] | 16:32 | |
-!- vgorbati_ is now known as vgorbati | 16:32 | |
-!- votjakovr [~votjakovr@host-46-241-3-209.bbcustomer.zsttk.net] has joined #shogun | 16:36 | |
-!- iglesiasg [~iglesiasg@213.143.50.76] has joined #shogun | 16:45 | |
-!- mode/#shogun [+o iglesiasg] by ChanServ | 16:45 | |
-!- gsomix [~Miranda@83.234.54.61] has quit [Quit: Miranda IM! Smaller, Faster, Easier. http://miranda-im.org] | 17:21 | |
-!- lisitsyn [~blackburn@109-226-97-52.clients.tlt.100megabit.ru] has left #shogun [] | 17:22 | |
-!- iglesiasg [~iglesiasg@213.143.50.76] has quit [Ping timeout: 252 seconds] | 17:38 | |
-!- iglesiasg [~iglesiasg@213.143.50.111] has joined #shogun | 18:05 | |
-!- mode/#shogun [+o iglesiasg] by ChanServ | 18:05 | |
-!- vgorbati [~vgorbati@212.2.159.34] has quit [Quit: vgorbati] | 18:20 | |
-!- n4nd0 [~iglesiasg@213.143.50.81] has joined #shogun | 18:50 | |
-!- iglesiasg [~iglesiasg@213.143.50.111] has quit [Ping timeout: 245 seconds] | 18:53 | |
-!- n4nd0 is now known as iglesiasg | 18:54 | |
-!- mode/#shogun [+o iglesiasg] by ChanServ | 18:55 | |
-!- iglesiasg [~iglesiasg@213.143.50.81] has quit [Ping timeout: 252 seconds] | 19:32 | |
-!- van51 [~van51@athedsl-320452.home.otenet.gr] has joined #shogun | 20:45 | |
-!- gsomix [~gsomix@46.20.65.107] has joined #shogun | 20:50 | |
gsomix | good evening | 20:50 |
van51 | hello | 20:52 |
-!- votjakovr [~votjakovr@host-46-241-3-209.bbcustomer.zsttk.net] has quit [Quit: Leaving] | 21:20 | |
@sonney2k | gsomix, hey there | 21:51 |
gsomix | sonney2k, hey | 21:51 |
@sonney2k | van51, did you meet already? | 21:51 |
van51 | sonney2k: hey, no | 21:51 |
van51 | they said tomorrow | 21:51 |
van51 | if I want to perform some file I/O | 21:55 |
van51 | should I use classes like SimpleFile? | 21:55 |
@sonney2k | van51, what file i/o? | 21:57 |
@sonney2k | van51, ascii? then gsomix' thing | 21:58 |
@sonney2k | van51, and what for? | 21:58 |
@sonney2k | van51, ohh btw - can you do the combineddotfeature? | 21:58 |
@sonney2k | s | 21:58 |
van51 | sonney2k: a step in the hashing trick is about tokenizing text documents | 22:00 |
van51 | sonney2k: do what about it? :) | 22:00 |
gsomix | sonney2k, btw, I'm preparing PR. | 22:02 |
@sonney2k | van51, errm convert from list -> dynarray | 22:05 |
van51 | sonney2k: yea sure | 22:06 |
@sonney2k | van51, regarding tokenizing | 22:06 |
@sonney2k | gsomix, just wrote a fast line reader which you can use for tokenizing too | 22:06 |
@sonney2k | van51, so basically you can give it stop symbols and then say 'give me next' | 22:06 |
van51 | it's in AsciiFile, right? | 22:06 |
@sonney2k | and then you can do whatever | 22:06 |
@sonney2k | van51, no will be a new file | 22:07 |
@sonney2k | err class | 22:07 |
van51 | is it already integrated in the development branch? | 22:07 |
@sonney2k | van51, gsomix just said he prepares the PR | 22:08 |
@sonney2k | van51, but I would focus on doing it in a way that you have a string in memory | 22:08 |
@sonney2k | because then this will work offline and online | 22:08 |
van51 | sonney2k: ok , was just was wandering if maybe there was a skeleton or an API already | 22:08 |
van51 | yea this would be just a front-end | 22:08 |
van51 | I think in order to compare the results of the soon to be created class | 22:09 |
@sonney2k | van51, well I would do the tokenizer by creating a char[256] array with stop tokens tagged and then just go through the string char by char, computing the hash of words | 22:11 |
@sonney2k | van51, this is something you could already do now | 22:12 |
van51 | sonney2k: yea I thought so too | 22:12 |
van51 | I was just wondering whether to use shogun classes for IO or from the standard libraries | 22:13 |
@sonney2k | van51, no I mean totally w/o IO | 22:15 |
@sonney2k | van51, just consider you get a SGString<char> | 22:15 |
@sonney2k | as input and operate on that | 22:16 |
@sonney2k | then it does not matter if it is a string or sth else | 22:16 |
@sonney2k | I mean input comes from a file | 22:16 |
van51 | sonney2k: ok got it! thanks for the suggestion | 22:18 |
van51 | i'll work a bit on combinedotfeatures and look into it after that | 22:18 |
@sonney2k | van51, the combineddotfeatures should be piece of cake for you now giving your combinedfeatures experience | 22:25 |
van51 | sonney2k: yea I guess :p | 22:43 |
van51 | sonney2k: are the iterators from this class used? | 22:51 |
-!- hushell [~hushell@8-92.ptpg.oregonstate.edu] has joined #shogun | 23:07 | |
van51 | nvm | 23:09 |
gsomix | sonney2k, done. just have updated PR. | 23:31 |
gsomix | sonney2k, I'm little unhappy with my code, because it's ugly and over, I think. hard review is needed. :) | 23:32 |
-!- nube [~rho@49.244.80.228] has quit [Ping timeout: 264 seconds] | 23:33 | |
gsomix | sonney2k, btw PR is still WIP. may be I can find some bugs tomorrow... and then... | 23:33 |
* gsomix z-z-z | 23:34 | |
gsomix | sonney2k, hope I'm not too slow. :( | 23:35 |
gsomix | nite guys | 23:42 |
van51 | gsomix: cu | 23:48 |
-!- nube [~rho@49.244.100.165] has joined #shogun | 23:48 | |
--- Log closed Mon Jun 03 00:00:29 2013 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!