Open in new window / Try shogun cloud
--- Log opened Tue Jul 12 00:00:09 2011
-!- in3xes_ [~in3xes@] has joined #shogun06:42
-!- in3xes [~in3xes@] has quit [Ping timeout: 250 seconds]06:45
-!- [1]warpy [] has joined #shogun06:49
-!- todun [] has joined #shogun07:18
-!- in3xes [~in3xes@] has joined #shogun08:04
-!- in3xes_ [~in3xes@] has quit [Ping timeout: 240 seconds]08:06
@sonney2k[1]warpy, still there?08:39
@sonney2kI think you don't need those loadlibrary things08:39
@sonney2kIMHO csharp will do that automagically08:39
@sonney2k[1]warpy, note however that currently none of the *.cs examples compiles - looks like there is something wrong in daniels typemaps08:40
-!- in3xes_ [~in3xes@] has joined #shogun08:42
-!- todun [] has quit [Quit: :q]09:26
-!- in3xes_ [~in3xes@] has quit [Remote host closed the connection]09:53
-!- [1]warpy [] has quit [Ping timeout: 255 seconds]10:14
-!- VojtechFranc [~quassel@2001:718:2:1612:221:6aff:fe58:151a] has joined #shogun10:14
--- Log closed Tue Jul 12 10:24:22 2011
--- Log opened Tue Jul 12 10:24:30 2011
-!- shogun-i1clog [] has joined #shogun10:24
-!- Irssi: #shogun: Total of 10 nicks [2 ops, 0 halfops, 0 voices, 8 normal]10:24
-!- Irssi: Join to #shogun was synced in 9 secs10:24
-!- Netsplit *.net <-> *.split quits: shogun-irclog10:29
-!- blackburn [~blackburn@] has joined #shogun10:52
-!- in3xes_ [~in3xes@] has joined #shogun11:22
-!- in3xes [~in3xes@] has quit [Ping timeout: 240 seconds]11:25
@sonney2kalesis-novik, VojtechFranc any status updates? I haven't seen yours this week?11:28
-!- in3xes1 [~in3xes@] has joined #shogun11:29
blackburnmidterm evaluation took me 5 minutes hehe11:30
* sonney2k too11:32
* sonney2k hopes that blackburn didn't mention vodka or anything alike :-)11:33
-!- in3xes_ [~in3xes@] has quit [Ping timeout: 258 seconds]11:33
blackburnyeah it was hard but I didn't11:34
-!- in3xes_ [~in3xes@] has joined #shogun11:40
-!- blackburn1 [~blackburn@] has joined #shogun11:41
-!- in3xes1 [~in3xes@] has quit [Ping timeout: 264 seconds]11:44
-!- in3xes1 [~in3xes@] has joined #shogun11:45
-!- blackburn [~blackburn@] has quit [Ping timeout: 255 seconds]11:45
-!- in3xes_ [~in3xes@] has quit [Ping timeout: 246 seconds]11:48
-!- heiko [] has joined #shogun11:50
-!- in3xes_ [~in3xes@] has joined #shogun12:10
-!- in3xes1 [~in3xes@] has quit [Ping timeout: 252 seconds]12:14
@sonney2kheiko, did you find the bug?12:20
heikosonney2k, hi12:20
heikocurrently testing from sratch again :(12:21
heikoso many possibilities12:21
heikoso many changes12:21
heikohope to get it fixed today12:21
@sonney2kheiko, I really always use valgrind for testing and pipe this to some file12:21
@sonney2kthen search for shogun in there12:21
@sonney2kusually there are unitialized memory reads12:22
@sonney2kor delete[]/delete  / free mismatches or even accidental del/frees12:22
heikoyes, am doing this :)12:26
@sonney2kheiko, that is the best you can do I am afraid...12:29
@sonney2kgdb and SG_PRINTS are the other options...12:29
-!- in3xes_ [~in3xes@] has quit [Ping timeout: 260 seconds]12:35
-!- blackburn1 [~blackburn@] has quit [Quit: Leaving.]12:39
-!- in3xes_ [~in3xes@] has joined #shogun12:47
-!- in3xes1 [~in3xes@] has joined #shogun13:00
-!- in3xes_ [~in3xes@] has quit [Ping timeout: 240 seconds]13:03
-!- in3xes_ [~in3xes@] has joined #shogun14:02
-!- in3xes1 [~in3xes@] has quit [Ping timeout: 240 seconds]14:05
heikoah, just found a bunch of bugs :)14:07
-!- in3xes1 [~in3xes@] has joined #shogun14:08
-!- in3xes_ [~in3xes@] has quit [Ping timeout: 240 seconds]14:11
@sonney2kheiko, and? does it help?14:52
heikosonney2k, yes some lower parts are working again14:53
heikomaking my way up to the top14:53
heikoI implemented a special feature in DynamicObjectArray that deleted elements unwanted :)14:53
@sonney2kas in delete_unwanted_elements :D14:54
heikoyeah ;D14:54
heikowell elements that you just put in but that the array did not want14:55
@sonney2kheiko, still isn't really clear to me...15:02
-!- [1]warpy [] has joined #shogun15:13
heikosonney2k, the array did eat up elements, this was not intended.15:34
heikoa bug15:34
[1]warpyhi sonney2k saw your msg from earlier15:34
@sonney2kheiko, you mean the array didn't shrink?15:37
heikono, just a bug15:38
@sonney2kI see15:38
heikowrong element deleted15:38
-!- VojtechFranc [~quassel@2001:718:2:1612:221:6aff:fe58:151a] has quit [Quit: No Ping reply in 180 seconds.]15:48
-!- VojtechFranc [] has joined #shogun15:50
-!- heiko [] has quit [Ping timeout: 258 seconds]16:05
[1]warpysonney2k, is there a way to issue a pull via the git website ?16:06
[1]warpyso basically i push it first16:09
[1]warpyand then i pull (ask you to look at it)16:09
[1]warpyokay, so can i push stuff or do i need permission16:19
-!- blackburn [~blackburn@] has joined #shogun16:20
-!- in3xes1 [~in3xes@] has quit [Quit: Leaving]16:24
-!- in3xes [~in3xes@] has joined #shogun16:25
@sonney2kyou can push as much as you want :)16:31
blackburni'd like to push some cats into shogun16:36
[1]warpysonney2k, what is the problem with the type map can you possibly email me the errors ?16:41
[1]warpyand also to daniel ?16:41
@sonney2kblackburn, shogun takes out is sword and cuts the cat into nice sashimi pieces to be served for dinner16:43
blackburnsonney2k: have you vietnamese ancestors? :D strange tastes ;)16:45
-!- blackburn1 [~blackburn@] has joined #shogun16:50
-!- blackburn [~blackburn@] has quit [Ping timeout: 255 seconds]16:51
[1]warpysonney2k, i forked the project, i have no idea how to push/commit16:53
[1]warpyfrom the web obviously16:55
[1]warpynot command line16:55
blackburn1just make changes, pull origin, commit it, push it to your fork and press pull request button on web16:56
-!- f-x [~user@] has joined #shogun16:57
-!- f-x [~user@] has quit [Remote host closed the connection]17:10
-!- VojtechFranc [] has quit [Remote host closed the connection]17:26
-!- heiko [] has joined #shogun18:08
heikosonney2k, are you there?18:11
-!- heiko [] has quit [Quit: Leaving.]18:26
-!- blackburn1 [~blackburn@] has quit [Quit: Leaving.]19:02
-!- blackburn [~blackburn@] has joined #shogun19:02
-!- [1]warpy [] has quit [Quit: HydraIRC -> <- Po-ta-to, boil em, mash em, stick em in a stew.]19:12
-!- in3xes is now known as gsoc19:13
-!- gsoc is now known as group19:14
-!- group is now known as gsoc19:14
-!- gsoc is now known as in3xes19:15
-!- f-x [~user@] has joined #shogun19:38
-!- in3xes_ [~in3xes@] has joined #shogun20:10
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]20:19
-!- blackburn [~blackburn@] has joined #shogun20:32
@sonney2kf-x, so how is it going?21:02
@sonney2kdid you start working on liblinear/sgd-qn?21:03
@sonney2kf-x, if not I have one more proposal :)21:03
@sonney2kf-x, you could download the original sgd / sgd-qn and run it on this subset of data that you have21:03
f-xsonney2k: started, but i'd love to hear your proposal :)21:03
@sonney2kthese can read svmlight file format21:04
@sonney2k(the format the python script of the large scale learnign challenge produces)21:04
@sonney2kso you can directly run them and check if they get very similar results21:04
@sonney2kf-x, but it is great that you started already21:05
f-xi ran liblinear on a small subset of data, but it always ended up reaching max iterations21:05
@sonney2kf-x, I am very excited about this...21:05
f-x(the original liblinear from shogun)21:05
@sonney2kf-x, on alpha even?21:05
@sonney2kf-x, which solver did you choose?21:06
f-xsonney2k: everything default21:06
@sonney2k(for liblinear)21:06
@sonney2kf-x, then of course :)21:06
f-xwhich is the fastest one?21:06
@sonney2kf-x, you need to use L2R  L1LOSS21:06
@sonney2kread L_2 norm regularized SVM-w with L_1-norm loss term21:06
f-xah.. i'll try that.. meanwhile i do have some questions about the online version.. i'll ask you a bit later21:07
@sonney2kthat is the SVM21:07
@sonney2kdefault in liblinear is IIRC some logistic regression21:07
f-xoh.. btw liblinear has come out with a new version which reads from datasets that don't fit in the memory21:08
f-xsonney2k: they split the problem into multiple blocks, and process one block at a time21:08
@sonney2kf-x, I know21:08
@sonney2kf-x, but it is not in liblinear directly AFAIK it is some addon21:09
f-xthey call it the 'block' version iirc21:09
@sonney2kf-x, yes but it is some extra c++ code that just uses liblinear or?21:09
f-xsonney2k: they modify the train and predict functions of the original liblinear..21:10
f-xso i'd consider it to be a standalone package21:10
@sonney2kI just checked
@sonney2kso yes indeed21:10
@sonney2kit is not in core liblinear21:10
@sonney2kf-x, if you have time later - this is something you could do21:11
f-xsonney2k: and one more thing - i'm finding it easier to be using DynArray instead of int* etc for liblinear21:11
f-xis that okay?21:11
@sonney2kfor what?21:11
@sonney2kblackburn, my replacement harddrive just arrived today21:11
@sonney2kblackburn, isn't that great news :D21:11
blackburnsonney2k: of course it is ;)21:12
@sonney2kblackburn, the only sad thing is that the drive's label says Certified Repaired HDD21:12
@sonney2kthat scares me quite a bit21:13
@sonney2kf-x, int* for what?21:13
blackburnbastards, I thought it is possible only in russia :D21:13
f-xsonney2k: umm.. the y values21:13
f-xdoesn't liblinear depend on multiple iterations?21:13
f-xis convergence guaranteed without this?21:14
@sonney2kf-x, not more than SGD21:14
@sonney2kf-x, no convergence is not guaranteed21:14
f-xbecause i see it retains some examples and discards others, and runs another iteration on the retained examples21:15
f-xi thought that was essential21:15
blackburnsonney2k: I started glancing PCACut now. is it PCA *Cut*??21:16
@sonney2kf-x, it is if you want to converge to some epsilon precision21:16
@sonney2kf-x, the discarding stuff is 'only' for speeding up21:17
@sonney2knot essential in the online variant21:17
f-xaah.. that solves most of my worries.. i thought somehow convergence could be guaranteed even through this online method21:17
@sonney2kf-x, well with infinite data SGD and liblinear will converge to the same solution :)21:18
f-xgreat.. so the basic mechanism is nearly identical21:18
f-xsonney2k: but i'll have to modify some fundamental structures also -- like 'problem'21:20
f-xagain CDotFeatures -> CStreamingDotFeatures21:20
f-xi don't see a way without creating something like online_problem21:21
@sonney2kf-x, yeah or do without problem21:23
f-xyes - another possibility21:23
f-xi guess i'll do it without the problem stuff, that's good21:24
@sonney2kf-x, just to be clear: you only need to convert solve_l2r_l1l2_svc21:26
f-xsonney2k: right. i've only looked at that for now21:26
@sonney2kthat is liblinear's svm like training algorithm21:26
@sonney2kno more than that21:26
@sonney2kthe rest is not necessary / not so useful21:27
f-xok. so the generic online features loading stuff, then train using solve_l2r_l1l2_svc, and again the generic predict stuff21:27
f-xonly solve_l2r_l1l2_svc is specific to liblinear21:28
@sonney2kf-x, yeah so just copy that function to OnlineLibLinear's train() function and modify it such that it can use streamingfeatures21:35
@sonney2kthat's it21:35
@sonney2kStreamingDotFeatures that is21:35
f-xyeah.. that way is much more compact.. thanks21:36
@sonney2kblackburn, I named it this way because it is doing PCA and could cut off features with low EV's21:36
@sonney2kblackburn, but you are the preproc master now21:36
blackburnsonney2k: I guess it is time to rename it to PCA - it transforms data, not cut off it21:37
@sonney2kf-x, yeah it is not a lot of work when you look at it directly from that angle21:37
blackburnor I got wrong?21:37
@sonney2kblackburn, as you wish :)21:39
blackburnsonney2k: well it transforms data, right? :)21:40
@sonney2kblackburn, as I said - you are the master of preprocs now - so just do it - I am fine with either name :)21:41
* sonney2k bites the dust and works on c# typemaps22:08
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]22:16
-!- in3xes_ [~in3xes@] has quit [Ping timeout: 260 seconds]22:40
--- Log closed Wed Jul 13 00:00:14 2011