--- Log opened Tue Jun 26 00:00:07 2012 | ||
-!- heiko [~heiko@host86-180-159-222.range86-180.btcentralplus.com] has joined #shogun | 00:34 | |
-!- ckwidmer [8ca3fe9d@gateway/web/freenode/ip.140.163.254.157] has quit [Ping timeout: 245 seconds] | 01:27 | |
-!- heiko [~heiko@host86-180-159-222.range86-180.btcentralplus.com] has quit [Ping timeout: 272 seconds] | 01:28 | |
-!- shogun-toolbox [~shogun@7nn.de] has quit [Ping timeout: 240 seconds] | 01:32 | |
--- Log closed Tue Jun 26 01:32:38 2012 | ||
--- Log opened Tue Jun 26 01:32:48 2012 | ||
-!- shogun-toolbox [~shogun@7nn.de] has joined #shogun | 01:32 | |
-!- Irssi: #shogun: Total of 7 nicks [1 ops, 0 halfops, 0 voices, 6 normal] | 01:32 | |
-!- Irssi: Join to #shogun was synced in 6 secs | 01:32 | |
-!- romi_ [~mizobe@187.66.121.115] has joined #shogun | 02:40 | |
-!- av3ngr [av3ngr@nat/redhat/x-fpfmjewzpoauxtlb] has joined #shogun | 03:03 | |
-!- av3ngr [av3ngr@nat/redhat/x-fpfmjewzpoauxtlb] has quit [Client Quit] | 03:03 | |
-!- cheng [80fa500f@gateway/web/freenode/ip.128.250.80.15] has joined #shogun | 03:07 | |
CIA-18 | shogun: Sergey Lisitsyn master * ra937587 / (5 files in 2 dirs): Updated multitask logistic regression - http://git.io/oEeUnA | 03:08 |
---|---|---|
CIA-18 | shogun: Sergey Lisitsyn master * r3ee917a / (2 files in 2 dirs): Fixed a bunch of bugs in multitask logistic regression - http://git.io/02T5NQ | 03:08 |
-!- blackburn [~blackburn@31.28.43.76] has quit [Ping timeout: 240 seconds] | 03:13 | |
-!- Cheng2 [~cheng@211.29.182.146] has joined #shogun | 03:27 | |
-!- Cheng2 [~cheng@211.29.182.146] has quit [Client Quit] | 03:27 | |
-!- pluskid [~pluskid@202.130.113.141] has joined #shogun | 03:53 | |
pluskid | hi cheng | 03:54 |
pluskid | just get up from bed :D | 03:54 |
cheng | Hi. Are we in the same timezone? | 03:56 |
pluskid | not sure, I'm in GMT+8, what about you? | 03:56 |
cheng | I think I'm GMT+10 | 03:56 |
cheng | It's almost lunchtime for me... :-) | 03:56 |
pluskid | hmm, you are a little ahead of me :) | 03:57 |
cheng | Just means that I can wake up later. | 03:57 |
pluskid | haha, but you also have to sleep later | 03:57 |
pluskid | shall we talk about our mid-term? or are you planing to have lunch now? | 03:57 |
cheng | Looking at the IRC logs, it looks like you're having productive discussions with blackburn and sonney2k | 03:58 |
pluskid | hmm, I just finished the ShareBoost the last week | 03:58 |
pluskid | failed to send the weekly report last night | 03:58 |
pluskid | just sent it a minute ago | 03:58 |
cheng | So, you did figure out the convex optimization problem? | 03:59 |
pluskid | yeah | 03:59 |
pluskid | Finally i used lbfgs to solve it | 03:59 |
pluskid | a by-product is that now I integrated lib-lbfgs in shogun | 03:59 |
pluskid | anyone else can also use that if needed | 03:59 |
cheng | lbfgs is not bad. | 03:59 |
cheng | liblibfgs from Nocedal? | 03:59 |
pluskid | I'm not sure, I'll check the page | 04:00 |
pluskid | it's a C library | 04:00 |
cheng | Doesn't that use fortran or something? | 04:00 |
pluskid | no | 04:00 |
pluskid | it is ported from fortran, but not machine-translation | 04:00 |
cheng | Ah, from chokkan. | 04:00 |
pluskid | http://www.chokkan.org/software/liblbfgs/ | 04:00 |
pluskid | this one | 04:00 |
pluskid | ah, yeah | 04:00 |
cheng | Good library. | 04:01 |
cheng | Nice. | 04:01 |
pluskid | sonney2k said it is not maintained for 1 year (so that we don't have to worry about future upgrade), so I modified it heavily to make it more integrated with shogun | 04:01 |
cheng | Did you contact the author? | 04:02 |
pluskid | no | 04:02 |
cheng | He may be interested in your updates. | 04:02 |
pluskid | maybe I should inform him | 04:02 |
cheng | It is nice for authors to know that someone modifies their code. | 04:02 |
pluskid | my updates are concerned with shogun | 04:02 |
pluskid | e.g. use shogun's SGVector for vector computing | 04:02 |
pluskid | but I didn't modify the algorithm | 04:02 |
pluskid | however, I think you are right, I still should inform him | 04:03 |
pluskid | that his great lib is being integrated in shogun :D | 04:03 |
cheng | Let him know that you've added it to shogun. | 04:03 |
pluskid | yeah, OK | 04:03 |
cheng | So, shareboost finished? | 04:03 |
pluskid | not completely | 04:04 |
pluskid | I'm still working on to provide the python-modular interface | 04:04 |
pluskid | after that, it will be finished | 04:04 |
pluskid | or I might try to test it on some none-toy dataset | 04:04 |
cheng | Depends what you mean by non-toy | 04:04 |
pluskid | I mean not those in shogun's data/toy directory :D | 04:05 |
pluskid | e.g. MNIST or USPS? | 04:05 |
cheng | Sounds good. | 04:05 |
pluskid | however, I'm currently having some problem with the python modular interface | 04:05 |
cheng | SWIG? | 04:06 |
pluskid | the SWIG compiler segfault when I tried to compile | 04:06 |
pluskid | I tried to trace the git history | 04:06 |
pluskid | it seems to be related to a commit from SO Learning | 04:06 |
pluskid | but that commit looks very ordinary, nothing dangerous | 04:06 |
pluskid | I'll have to dig deeper to debug this out | 04:07 |
pluskid | hope I can figure this out today or tomorrow | 04:07 |
pluskid | so back to our mid-term | 04:07 |
cheng | I just looked at the GSoC calendar. http://www.google-melange.com/gsoc/events/google/gsoc2012 | 04:08 |
cheng | The mid term is due on 9 July. | 04:08 |
pluskid | hmm | 04:08 |
pluskid | I looked at my original proposal | 04:08 |
cheng | After mid term, there are 6 more weeks of work. | 04:09 |
pluskid | around the time by mid-term, I planed to implement ECOC, ShareBoost, Fast Label Tree and a preliminary draft of survey | 04:09 |
pluskid | it looks our progress are not bad | 04:09 |
pluskid | I haven't implemented Fast Label Tree yet | 04:09 |
pluskid | but I implemented conditional prob tree instead | 04:10 |
cheng | I thought that the original proposal was quite ambitious. | 04:10 |
pluskid | yeah, my original proposal was also very vague about the work during Jun. 21 and Aug. 10 :D | 04:10 |
pluskid | so we have (a bit less than) 2 weeks before the mid-term | 04:11 |
cheng | So, what's left from the orignal list? 1. survey draft 2. fast label tree | 04:12 |
pluskid | hmm, survey draft is already there (really a draft only :p), but there's large space for improvement, coz many entries are only titles without content | 04:13 |
pluskid | and yeah, fast label tree | 04:13 |
cheng | where's the draft? | 04:13 |
pluskid | I sent it in one of the weekly report earlier | 04:13 |
pluskid | shall I send a copy to you again? | 04:14 |
pluskid | did you get that? | 04:14 |
cheng | Had a look again at my email, and didn't see any attachments. | 04:16 |
pluskid | OK, I'll send you a copy now | 04:16 |
cheng | I'm just thinking that it would be convenient to have it in the repository too. | 04:17 |
pluskid | hmm, not sure where to put it | 04:17 |
pluskid | I can discuss that with sonney2k and blackburn later | 04:18 |
pluskid | just sent | 04:18 |
cheng | One of the "usability improvements" could be such a survey of all algorithms implemented in shogun. | 04:19 |
cheng | So, what do you want to do until 9 July? | 04:19 |
cheng | BTW, just in case it wasn't clear, I think you're doing fantastically well. | 04:20 |
pluskid | I think basically two options: fast label tree or make the survey more complete | 04:20 |
pluskid | cheng: thanks! | 04:20 |
pluskid | cheng: how about we decide what's for the work due 9 July after I (completely) finish ShareBoost? | 04:21 |
pluskid | If the time is enough | 04:21 |
pluskid | I'll go with fast label tree | 04:22 |
pluskid | if not, I'd prefer make the survey complete | 04:22 |
pluskid | (and I'm usually helping sonney2k or blackburn to improve some basic things (like refactoring) in shogun during writing surveys) :D | 04:22 |
cheng | I'd say we say work on survey document. Even if you have lots of time. Because it would be important for you to figure out what else you want to do for the last 6 weeks. | 04:23 |
pluskid | yeah! that's what wondering me, too | 04:23 |
cheng | So, move fast label tree to 2 weeks away, move survey to the front. | 04:24 |
pluskid | yeah | 04:24 |
pluskid | so we can't decided what to do during the second half of GSoC, yet? | 04:24 |
cheng | 1. Fast label tree :-) | 04:25 |
pluskid | according to my previous progress, it seems that I can implement an algorithm within 2~3 weeks if things don't go wrong | 04:25 |
cheng | How about this old paper about learning the ECOC? | 04:25 |
pluskid | so if as you said, there will be 6 weeks, we might be able to implement 2~3 important algorithms | 04:25 |
pluskid | which old paper? | 04:26 |
cheng | http://books.nips.cc/papers/files/nips15/AA04.pdf | 04:27 |
pluskid | ha, it's from "Gun" | 04:28 |
cheng | There are some more recent papers ICML2011 I think related to this, but I cannot remember them at the moment. Perhaps a task for you. ;-) | 04:28 |
pluskid | OK, I'll write that down | 04:28 |
pluskid | but maybe we could also discuss with sonney2k | 04:28 |
cheng | Yes, this paper is ahead of its time, and has a terrible title. | 04:28 |
pluskid | he seems to be not very interested in ECOC related stuff | 04:29 |
pluskid | ahead of its time? what's that mean? | 04:29 |
cheng | I think it would be nice for shogun (soeren probably too) to have really many classes multiclass. | 04:29 |
cheng | I mean that nobody in 2002 thought about learning the relationship between classes. Today, it is more popular. | 04:29 |
cheng | I'm open to suggestions. | 04:30 |
pluskid | hmm | 04:31 |
cheng | Here's another simple algorithm. http://research.microsoft.com/en-us/um/people/ohadsh/2010_AISTATS_DekelSham.pdf | 04:31 |
pluskid | more class than examples, interesting | 04:32 |
pluskid | so some of the classes won't even has training examples | 04:33 |
-!- romi_ [~mizobe@187.66.121.115] has quit [Ping timeout: 245 seconds] | 04:34 | |
cheng | It is two stage. There's some assumption about how the classes are related. Something like what blackburn is doing with multitask. | 04:34 |
pluskid | hmm, that would make sense | 04:35 |
pluskid | or else no information can be obtained for those classes (without training samples) | 04:35 |
cheng | More john langford stuff: http://hunch.net/~beygel/tournament.pdf | 04:36 |
cheng | He calls it a "Filter Tree". | 04:37 |
pluskid | ah, this one was in my reading list, but to be frank I haven't read it yet | 04:37 |
cheng | It may be superceeded by the conditional probability tree. I don't know. | 04:38 |
pluskid | they also got a paper named "multiclass classification with filter-trees" | 04:39 |
pluskid | not sure that is the same as this one | 04:39 |
cheng | I think it is the same paper, but please check. I'm often wrong. | 04:40 |
pluskid | maybe the same, I found that paper from the references of another one, and it was said to be a "preprint", maybe they changed title when publishing | 04:40 |
pluskid | anyway, I'll check it | 04:40 |
cheng | I think with 4 potential ideas, you should be fine. sonney2k probably has some other thoughts. | 04:41 |
pluskid | ah, yeah! | 04:41 |
cheng | While writing your survey, you can choose 2-3 of the 4. Basically rank them | 04:41 |
pluskid | OK, I'll try | 04:41 |
pluskid | I'll ask your suggestions when I'm more familiar with those algorithms | 04:42 |
-!- romi_ [~mizobe@187.66.121.115] has joined #shogun | 04:42 | |
cheng | Different topic: did anything happen with the datasets idea? | 04:42 |
pluskid | unfortunately, no | 04:42 |
pluskid | everyone is super busy | 04:42 |
pluskid | even the doc-shogun project proposed long before by soeren and blackburn is postponed | 04:43 |
cheng | For your large number of classes methods, it would be nice to have some non-toy data. | 04:43 |
cheng | Oh well, cannot be helped. | 04:44 |
pluskid | haha | 04:44 |
pluskid | I think I can try during my surveying | 04:44 |
cheng | Hope you are learning lots while working on GSoC. | 04:45 |
pluskid | yeah, definitely :) | 04:45 |
pluskid | but I hope I can dig more deep into the theory of those algorithms I implemented | 04:45 |
cheng | Well, you have a PhD in front of you. | 04:45 |
pluskid | hmm, you are right | 04:46 |
pluskid | a long way to go | 04:46 |
cheng | Who's going to be your potential supervisors? | 04:46 |
pluskid | they haven't told me yet | 04:46 |
cheng | Normal for US universities. | 04:46 |
pluskid | why? | 04:47 |
pluskid | because they are too lazy to work? :p | 04:47 |
cheng | Just curious. You can probably then choose supervisors that can teach you some theory. | 04:47 |
pluskid | hmm, yeah, they said supervisor can be changed easily when you arrive if you really think the one is not suitable with you | 04:48 |
cheng | There are plenty of good courses at MIT, e.g. game theory, stochastic optimization, etc. | 04:48 |
pluskid | yeah, I'm looking forward for those :D | 04:49 |
pluskid | sad thing would be you only have 24 hours a day | 04:49 |
cheng | Yes, I'm a bit jealous that I don't really have too much time to code. | 04:50 |
pluskid | haha | 04:50 |
cheng | It is getting better. I just moved this year, so I'm kind of doing 2 jobs. | 04:50 |
pluskid | doing 2 jobs would be super busy! | 04:51 |
cheng | The irritating thing is that my previous job was in Europe, so I do a lot of night time meetings/discussions. | 04:51 |
cheng | All very fun and interesting though... | 04:52 |
pluskid | haha | 04:52 |
pluskid | CS people usually do not have regular sleeps | 04:52 |
cheng | Keep up the good work. | 04:52 |
pluskid | that's nog good | 04:52 |
cheng | I'll go for lunch soon... | 04:53 |
pluskid | OK | 04:53 |
pluskid | btw: the link is broken? http://www.ong-home.my/ | 04:53 |
cheng | Yes, Soeren is hosting for me, and just moved the nameserver. | 04:53 |
pluskid | first time of me heard of some .my domain name :D | 04:54 |
cheng | Cool heh? | 04:54 |
pluskid | have a nice meal | 04:54 |
pluskid | cool! | 04:54 |
-!- cheng [80fa500f@gateway/web/freenode/ip.128.250.80.15] has quit [Quit: Page closed] | 04:56 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 06:48 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Client Quit] | 06:52 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 06:55 | |
-!- pluskid [~pluskid@202.130.113.141] has quit [Quit: Leaving] | 07:11 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 08:43 | |
-!- uricamic [~uricamic@2001:718:2:1634:3d67:c999:8cc9:d056] has joined #shogun | 08:52 | |
-!- shogun-buildbot [~shogun-bu@7nn.de] has quit [Remote host closed the connection] | 09:01 | |
--- Log opened Tue Jun 26 09:02:26 2012 | ||
-!- shogun-toolbox [~shogun@7nn.de] has joined #shogun | 09:02 | |
-!- Irssi: #shogun: Total of 6 nicks [0 ops, 0 halfops, 0 voices, 6 normal] | 09:02 | |
-!- Irssi: Join to #shogun was synced in 6 secs | 09:02 | |
-!- shogun-buildbot [~shogun-bu@7nn.de] has joined #shogun | 09:02 | |
-!- sonney2k [~shogun@7nn.de] has joined #shogun | 09:02 | |
-!- mode/#shogun [+o sonney2k] by ChanServ | 09:03 | |
-!- assistance [9320543b@gateway/web/freenode/ip.147.32.84.59] has joined #shogun | 09:10 | |
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun | 09:23 | |
-!- ipondering [~iponderin@58.194.224.104] has joined #shogun | 09:35 | |
-!- gsomix [~gsomix@95.67.169.72] has joined #shogun | 09:57 | |
gsomix | hi | 09:57 |
-!- heiko [~heiko@host86-180-159-222.range86-180.btcentralplus.com] has joined #shogun | 09:59 | |
-!- gsomix_ [~gsomix@95.67.191.81] has joined #shogun | 10:06 | |
-!- gsomix [~gsomix@95.67.169.72] has quit [Ping timeout: 252 seconds] | 10:09 | |
-!- gsomix_ is now known as gsomix | 10:42 | |
-!- nicococo_ [5e8b1a4f@gateway/web/freenode/ip.94.139.26.79] has joined #shogun | 11:01 | |
-!- naywhayare [~ryan@spoon.lugatgt.org] has quit [Ping timeout: 272 seconds] | 11:04 | |
-!- naywhayare [~ryan@spoon.lugatgt.org] has joined #shogun | 11:05 | |
-!- pluskid [~pluskid@111.120.64.13] has joined #shogun | 11:21 | |
pluskid | away | 11:34 |
-!- ipondering [~iponderin@58.194.224.104] has quit [Quit: ipondering] | 11:37 | |
-!- ipondering [~iponderin@2001:da8:203:1823:4919:ecd1:c3d5:4242] has joined #shogun | 12:25 | |
-!- Cheng [~cheng@58.108.31.213] has joined #shogun | 12:32 | |
-!- pluskid [~pluskid@111.120.64.13] has quit [Ping timeout: 248 seconds] | 12:40 | |
-!- pluskid [~pluskid@202.130.113.141] has joined #shogun | 12:41 | |
Cheng | Hi again pluskid | 12:47 |
pluskid | Cheng: hi | 12:47 |
@sonney2k | Cheng, did you manage to get your domain name pointed to the right IP address now? | 12:47 |
-!- cheng_ [73406f11@gateway/web/freenode/ip.115.64.111.17] has joined #shogun | 12:50 | |
-!- Cheng [~cheng@58.108.31.213] has quit [Ping timeout: 248 seconds] | 12:51 | |
cheng_ | Pointed nameserver already, but website still broken. Dunno why. | 12:54 |
@sonney2k | cheng_, when did you do it? | 12:54 |
cheng_ | ssh works. I'm guessing there is something strange about primary and secondary name servers. | 12:55 |
@sonney2k | cheng_, ahh! | 12:55 |
@sonney2k | I am your NS right? | 12:55 |
@sonney2k | in this case I have to do it | 12:55 |
@sonney2k | src | 12:55 |
@sonney2k | sry | 12:55 |
cheng_ | I see. | 12:55 |
@sonney2k | then please don't do anything... | 12:55 |
@sonney2k | or undo whatever I said :D | 12:55 |
cheng_ | I should change the name server back? | 12:55 |
cheng_ | I see. | 12:56 |
@sonney2k | totally forgot that I am doing this | 12:56 |
cheng_ | Do you guys have any tips for detecting OSX veclib? | 12:58 |
-!- cwidmer [~chris@cpe-69-201-177-140.nyc.res.rr.com] has joined #shogun | 12:58 | |
cheng_ | configure fails for shogun. | 12:58 |
@sonney2k | cheng_, updated the dns entry | 12:59 |
-!- puffin444 [62e3926e@gateway/web/freenode/ip.98.227.146.110] has joined #shogun | 12:59 | |
@sonney2k | cheng_, now it needs some time to propagate... | 12:59 |
cheng_ | I've pointed the registry back to the old name server. | 13:00 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 13:01 | |
cheng_ | I guess I don't need blas lapack for the toy problems I have. | 13:01 |
cheng_ | I still wish I understood the whole linking thing under osx. | 13:01 |
uricamic | hi | 13:02 |
@sonney2k | nicococo_, around? | 13:02 |
nicococo_ | hi | 13:03 |
@sonney2k | nicococo_, do you know where alex is? | 13:04 |
nicococo_ | i don't know, sorry | 13:04 |
@sonney2k | uricamic, is vojtech around? | 13:04 |
assistance | yes, hi | 13:04 |
@sonney2k | nicococo_, can you check? | 13:04 |
@sonney2k | assistance == vojtech? | 13:04 |
@sonney2k | ok | 13:04 |
assistance | sorry, I'll log again | 13:04 |
-!- assistance [9320543b@gateway/web/freenode/ip.147.32.84.59] has quit [Quit: Page closed] | 13:04 | |
-!- vojtech [9320543b@gateway/web/freenode/ip.147.32.84.59] has joined #shogun | 13:05 | |
nicococo_ | sonney: can't check, i'm at home | 13:05 |
@sonney2k | ok | 13:05 |
@sonney2k | ipondering, you are not alex right? | 13:06 |
@sonney2k | so who is missing? | 13:06 |
n4nd0 | blackburn? | 13:06 |
@sonney2k | alex and blackburn | 13:06 |
@sonney2k | anyone else? | 13:06 |
@sonney2k | gsomix, can you ping blackburn? | 13:07 |
heiko | sonney2k, unfortunately, I havent heard anything from arthur since yesterday, but I will probably meet him this week to update in case he wont make it | 13:07 |
@sonney2k | heiko, I guess arthur won't join right? | 13:07 |
gsomix | sonney2k, just a moment | 13:07 |
@sonney2k | heiko, ok | 13:07 |
-!- blackburn [d5578b92@gateway/web/freenode/ip.213.87.139.146] has joined #shogun | 13:09 | |
blackburn | hey there sorry | 13:09 |
cwidmer | hi | 13:09 |
@sonney2k | blackburn, hope you did shower already | 13:09 |
@sonney2k | so only alex is missing... then lets start | 13:09 |
blackburn | shower? | 13:09 |
@sonney2k | Welcome everyone to our third meeting. This is our last meeting before the midterm evaluations are due. | 13:09 |
blackburn | I just ran into room :D | 13:09 |
gsomix | blackburn, hey, got sms? | 13:09 |
@sonney2k | So mentors please listen up - starting from July 9 you will be able to submit midterm evaluations. | 13:10 |
blackburn | gsomix: yes but it was late - I was connecting already | 13:10 |
@sonney2k | These are crucial to students - they won't get paid if you don't evaluate them. | 13:10 |
@sonney2k | You can pass or fail a student, though hopefully the latter is not necessary. | 13:10 |
@sonney2k | Students, you know you are on track when you are halfway done and plenty of code has already been merged into shogun - in that case there is nothing to worry about! | 13:10 |
@sonney2k | Some comments: Please students stay in #shogun during normal working ours. | 13:11 |
@sonney2k | Recall GSoC is a full time job/commitment so that really shouldn't be too hard. | 13:11 |
blackburn | I'd like to say it is ok to bias to night a little still (I do that this way ;) | 13:12 |
@sonney2k | Please do not forget about your weekly reports and send them on time. These are really important to track progress not only for your mentors. | 13:13 |
@sonney2k | blackburn, yes sure | 13:13 |
@sonney2k | whatever timezone you want to live is is fine :D | 13:13 |
blackburn | amount of hours you devote matters | 13:13 |
@sonney2k | so samoa time or anything | 13:13 |
@sonney2k | Do many small pull requests if possible instead of one big hard to review PR. | 13:14 |
@sonney2k | even if it is not yet working! | 13:14 |
-!- alexlovesdata [82955843@gateway/web/freenode/ip.130.149.88.67] has joined #shogun | 13:14 | |
@sonney2k | argh | 13:14 |
@sonney2k | dorbell | 13:14 |
cwidmer | pizza time | 13:15 |
heiko | I would like to add: The not-working code shouldnt be a big problem since most of the stuff is new and nothing depends on it. | 13:15 |
-!- os252 [c07c1afb@gateway/web/freenode/ip.192.124.26.251] has joined #shogun | 13:16 | |
@sonney2k | heiko, yes exacly | 13:18 |
@sonney2k | alexlovesdata, os252 you missed one important sentence: So mentors please listen up - starting from July 9 you will be able to submit midterm evaluations. | 13:19 |
wiking | sonney2k: but mine is currently breaks some api if i rebase it to the HEAD. i.e. did some changes in some qp solvers | 13:19 |
os252 | sonney2k: ok, thanks for the headsup (and sorry for being late). | 13:20 |
@sonney2k | so mentors, by end of july 10 I expect everyone to have the evaluations submitted | 13:20 |
alexlovesdata | sonney2k: until when we have to deliver the miodterm evalutions? You said: from 9.7. on | 13:20 |
alexlovesdata | aha | 13:20 |
@sonney2k | it is a 10 minute effort (click through from on gsoc website) | 13:20 |
blackburn | wiking: I would really like to merge it asap | 13:20 |
nicococo_ | ales loves data? from star strek? | 13:20 |
n4nd0 | alexlovesdata: July 13th is the deadline http://google-melange.appspot.com/gsoc/events/google/gsoc2012 | 13:21 |
@sonney2k | wiking, maybe you stay a little longer after the meeting - we then figure out what has to be done | 13:21 |
wiking | sonney2k: ok | 13:21 |
@sonney2k | wiking, but we really really need your code merged ASAP | 13:21 |
@sonney2k | Finally, if you cannot be online in #shogun or are unavailable please notify your mentor *beforehand*. | 13:21 |
@sonney2k | so any further questions to that? | 13:22 |
alexlovesdata | ye | 13:23 |
@sonney2k | Alright then as usual, please every mentor/student team - give us a status update, telling us what you did since the last meeting, whether you are behind or on schedule and what has already been merged into shogun. | 13:23 |
alexlovesdata | anything which I cna read how to do the midterm evaluation ^^? | 13:23 |
@sonney2k | alexlovesdata, I will send and email then | 13:23 |
n4nd0 | sonney2k: do the students have to prepare something for the midterm evaluation? | 13:23 |
@sonney2k | with further instructions | 13:23 |
@sonney2k | blackburn, please say what you had to do last year | 13:23 |
alexlovesdata | sonney2k: Thank! | 13:24 |
blackburn | sonney2k: had to do with midterm? | 13:24 |
blackburn | well there was some form to fill in | 13:24 |
@sonney2k | blackburn, that's it right? | 13:24 |
blackburn | yeah | 13:25 |
blackburn | only one form | 13:25 |
@sonney2k | alexlovesdata, same for mentors | 13:25 |
@sonney2k | really fast job | 13:25 |
heiko | its was just a few click questions and some form fields with comments: number of hours, satisfied?, etc | 13:25 |
@sonney2k | so back to the updates | 13:25 |
@sonney2k | From my side - we now have some dedicated server hosting shogun-toolbox.org that does the webservice, irc bot/logging, buildbot. I | 13:25 |
@sonney2k | It is some core 2 quad 2.4GHz, with 4 G of memory and 1TB harddrive. So we can host some example data sets and lots of other stuff there. So everyone here succeeding GSoC can get a myname@shogun-toolbox.org forward to his address :D If someone has other ideas - notify me | 13:25 |
@sonney2k | comments/questions to that? | 13:26 |
n4nd0 | nice! | 13:26 |
@sonney2k | if not gsomix - can you begin? | 13:26 |
heiko | nice :) | 13:26 |
@sonney2k | yeah thanks to google giving us $500 / student we can rent one now :D | 13:26 |
gsomix | sonney2k, ok. | 13:26 |
@sonney2k | (maybe you all take the time to prepare just now) | 13:28 |
gsomix | Will be brief. I have passed all exams and now I can work on my gsoc project. I and sonney2k plan to implement this director classes: "sparse features, string features, multiclass machines, preprocessors, structured output machines, converters, loss, labels". Before midterm and with examples, of course. | 13:28 |
@sonney2k | (next then vojtech/uricamic) | 13:28 |
vojtech | ok | 13:28 |
@sonney2k | the good news is | 13:28 |
vojtech | Michal implemented the BMRM algorithm of Teo et al. ( = Joachims' StructSVM) to the SO framework of Shogun. | 13:28 |
@sonney2k | that directors are really working | 13:29 |
@sonney2k | vojtech, wait wait | 13:29 |
@sonney2k | after gsomix is done | 13:29 |
@sonney2k | so one can currently implement ones own kernel in python | 13:29 |
@sonney2k | and use it from within any SVM of shogun :D | 13:29 |
gsomix | Since the last meeting we (I and sonney2k) did director machines and dot features. And polymorphic return for features. | 13:29 |
gsomix | sonney2k, it seems, that's all. | 13:30 |
@sonney2k | ...all of it is merged | 13:30 |
@sonney2k | so any comments? | 13:30 |
gsomix | yep | 13:31 |
@sonney2k | from my side only - future will be python buffer protocol | 13:31 |
@sonney2k | that stuff will enable things like X+=1 | 13:31 |
@sonney2k | where X is some shogun DenseFeatures | 13:31 |
@sonney2k | (inplace) | 13:31 |
@sonney2k | and X*Y with Y being a numpy matrix | 13:32 |
blackburn | I'd like to have [] -> CList, CSet and CDynArray typemaps | 13:32 |
@sonney2k | blackburn, file a bug report and assign gsomix :D | 13:32 |
blackburn | I actually tried that and failed :( | 13:32 |
@sonney2k | (though it doesn't make sense for list/set - but lets discuss that later) | 13:32 |
@sonney2k | vojtech, your turn | 13:32 |
gsomix | I also remember that n4nd0 wanted to use some director stuff. | 13:33 |
n4nd0 | yeah, for the StructuredModel | 13:33 |
@sonney2k | yeah notify us when you want to be able to overload some specific functions... | 13:33 |
vojtech | ok, can I start? | 13:33 |
@sonney2k | (after that n4nd0/nico) | 13:33 |
@sonney2k | vojtech, yes please | 13:34 |
vojtech | Michal implemented the BMRM algorithm of Teo et al. ( = Joachims' StructSVM) to the SO framework of Shogun. | 13:34 |
vojtech | As the core QP solver the BMRM solver uses my LIBQP which was also imported to Shogun. | 13:34 |
vojtech | Michal wrote C example which uses the BMRM for learning linear multi-class SVM. | 13:34 |
vojtech | He crosschecked the Shogun implementation with out many-times-verified Matlab code. | 13:34 |
vojtech | The code is submitted to Shogun. | 13:34 |
vojtech | The next plan is to implement the i) Bundle method and ii) use it for one real-life problem. | 13:34 |
vojtech | My suggestion is to select one of the standard SO benchmarks which appear in 90% of SO papers. | 13:35 |
@sonney2k | vojtech, any particular one? | 13:35 |
vojtech | Yes, there are approximately 4 standard benchmarks | 13:35 |
vojtech | texline OCR and problems from NLP | 13:36 |
vojtech | lile name entity recognition | 13:36 |
@sonney2k | ok | 13:36 |
vojtech | Having such benchmark in shogun would increase its value of the scientific community a ,ot | 13:36 |
vojtech | a lot | 13:36 |
blackburn | sounds like advertisement :) | 13:36 |
vojtech | i.e. not just the base-line algorithms but also ready to use benchmark sets | 13:37 |
vojtech | this is what everyone who publishes on SO needs to do anyway | 13:37 |
vojtech | why not to do it once and carrefuly | 13:37 |
alexlovesdata | I can offer kernels from Pascal VOC2009 and ImageCLEF2010, ImageCLEF2011 for MKL testing | 13:37 |
alexlovesdata | Plosone paper is published | 13:37 |
@sonney2k | vojtech, will there be time to do the libqp extensions once the speeded up BMRM is done? | 13:37 |
vojtech | not in one month | 13:38 |
@sonney2k | vojtech, we intend to release shogun 2.0 September 1st | 13:38 |
@sonney2k | so > 2 months :D | 13:39 |
vojtech | if I find someone who will program it then I'm not against :) | 13:39 |
@sonney2k | anyway vojtech, uricamic - thanks for your work | 13:39 |
@sonney2k | vojtech, can uricamic do it? | 13:39 |
vojtech | theoretically yes | 13:40 |
@sonney2k | but? | 13:40 |
vojtech | If he has time | 13:40 |
@sonney2k | ok | 13:40 |
vojtech | I cannot tell it now | 13:40 |
@sonney2k | well then thanks | 13:40 |
nicococo_ | our turn now? | 13:40 |
@sonney2k | n4nd0, nicococo_ yes | 13:40 |
nicococo_ | a primal structured output svm with a mosek solver is implemented and tested on a simple multi-class example. now, we move on to an hidden markov so-svm. hence, we are half way now, i would guess. in addition, we started to write down everything concerning hmsvm implementation. | 13:40 |
@sonney2k | (next is wiking / alexlovesdata) | 13:41 |
alexlovesdata | wiking? | 13:41 |
@sonney2k | actually cool that you document these things | 13:41 |
wiking | alexlovesdata: yes? | 13:41 |
nicococo_ | with neat pictures ;) | 13:41 |
@sonney2k | so we can put them in the to-be-created-doc | 13:41 |
alexlovesdata | you like to tell? | 13:41 |
wiking | alexlovesdata: they are still talking i believe... | 13:42 |
alexlovesdata | noo | 13:42 |
@sonney2k | wiking, yes stay tuned | 13:42 |
alexlovesdata | our turn now ;) | 13:42 |
@sonney2k | alexlovesdata, wait | 13:42 |
wiking | alexlovesdata: ^ sonney2k | 13:42 |
@sonney2k | n4nd0, nicococo_ so you finally get the exact same solution as (what did you compare with?) | 13:42 |
n4nd0 | I have carried out some tests comparing with the results of SVM-struct | 13:43 |
n4nd0 | but just based on accuracy so far | 13:43 |
@sonney2k | n4nd0, nicococo_ what else is on the roadmap - 'only' HMSVM with mosek or something else? | 13:43 |
n4nd0 | I know that this is not good ... | 13:43 |
cheng_ | nicococo_: It seems that for multiclass, we could compare with stuff that pluskid has implemented. | 13:43 |
@sonney2k | cheng_, and uricamic's stuff too | 13:44 |
n4nd0 | but the objectives of the other problems are not the same maybe | 13:44 |
@sonney2k | nicococo_, why not just compute the unconstrained objective? | 13:44 |
nicococo_ | i think, n4and0 tried that already and without bias it was pretty much the same, right? | 13:44 |
vojtech | n4nd0, it might be useful to compare the primal SO solver with the BMRM both should d othe same | 13:44 |
cheng_ | @sonney2k: perhaps it is worth having some multiclass benchmarks (in mldata.org?) | 13:45 |
@sonney2k | I mean some function that computes objective based on the set of w's | 13:45 |
n4nd0 | vojtech: ok thank you, but I think sonney2k told me that the BMRM problem is unconstrained | 13:45 |
nicococo_ | i think, it is okay to have a bundle solver AND a mosek implementation | 13:45 |
@sonney2k | vojtech, do you have a function like that already? | 13:45 |
n4nd0 | or that I shouldn't take into account the constraints, I am not sure right now | 13:45 |
@sonney2k | n4nd0, it is easier to compare unconstrained objective | 13:45 |
n4nd0 | sonney2k: but I am not sure if I understand that properly | 13:46 |
n4nd0 | if we don't take into account the constraints, then the problem is completely different | 13:46 |
vojtech | n4nd0, I think you can also solve unconstrained problems right? | 13:46 |
nicococo_ | sonney: we can approach another solver after hmsvm is implemented, is that okay? | 13:47 |
vojtech | n4nd0, can I look at the formulation? | 13:47 |
n4nd0 | vojtech: sure, here http://iglesiashogun.wordpress.com/2012/06/04/second-weekly-report-gsoc-2012/ | 13:47 |
@sonney2k | vojtech, nicococo_, n4nd0, uricamic maybe you quickly discuss after the meeting | 13:47 |
nicococo_ | sure | 13:47 |
n4nd0 | vojtech: found it? | 13:47 |
blackburn | yes please (I have to leave very very soon :D) | 13:47 |
@sonney2k | so please wiking continue | 13:48 |
wiking | I couldn't participate the last meeting, thus i reckon the best is to start from GSoC beginning: before the last meeting i've started | 13:48 |
wiking | implementing the framework of latent svm (both normal and structural). The P. Felzenszwalb based latent svm has been implemented to work | 13:48 |
wiking | with the given mammal dataset (finding mammals within a given picture based on HoG features). In the meanwhile (from where i've | 13:48 |
wiking | got that latent dataset) i've started to check the implementation of the structural latent svm (both by Chun-Nam and M. Pawan Kumar) | 13:48 |
wiking | and the entropy based latent svm (by M. Pawan Kumar). As Fernando had to go with Mosek implementation of the structural SVM (that's | 13:48 |
wiking | used by Joachim's struct SVM), i've went first with the implementation of CCCP based latent structural SVM (see Chun-Nam). While i had | 13:48 |
wiking | some troubles of testing (not having a data set that i could use to compare my results) i went on with the implementation of M3E by | 13:48 |
wiking | M. Pawan Kumar that is a Renyi entropy based latent svm solver. and since i've already in the beginning proposed that my intention was | 13:48 |
vojtech | n4nd0, yes I have it. But the formulation can be solved as an unconstrained problem as well | 13:48 |
wiking | basically to try to extend the way that entropy is measured, i went on and tried out Tsallis-entorpy based one and experimented | 13:48 |
wiking | with using Jensen-Shannon Divergence based measure (see paper of M. Pawan Kumar for details). Last week Chun-nam got back to me finally | 13:48 |
wiking | with a data set that he used in his paper and he could share it with me in order that i can compare with the results of mine. | 13:48 |
wiking | Currently it's nowhere near to the 'original' implementation, which i cannot understand yet | 13:48 |
wiking | because it's CCCP is based on Mosek solver as well just like Fernando's. | 13:48 |
n4nd0 | vojtech: aham, let's talk about it later then | 13:48 |
wiking | (hope it's not flooded to anybody) | 13:48 |
alexlovesdata | wiking: we will talk after that meeting about that | 13:49 |
vojtech | n4nd0, let's talk after the meeting | 13:49 |
wiking | ok so it hasn't flooded.. | 13:49 |
@sonney2k | wiking, so from what I see you make good progress - but you really really need to get this merged into shogun | 13:49 |
wiking | sonney2k: yes talk now or after meeting about PR | 13:50 |
wiking | ? | 13:50 |
@sonney2k | wiking, there is no way a student can pass midterm without PRs being merged | 13:50 |
@sonney2k | yes | 13:50 |
wiking | yes now or after? | 13:50 |
@sonney2k | after meeting yes | 13:50 |
wiking | ok | 13:50 |
@sonney2k | wiking, so what is next on the plan? | 13:50 |
@sonney2k | (except for the PR) | 13:51 |
wiking | well alexlovesdata pointed me out to go and try SO solver of vojtech/uricamic see how that works concerning accuracy | 13:51 |
@sonney2k | alexlovesdata, and others - further comments? | 13:51 |
alexlovesdata | what is PR ??? I just have presse release on my mind. Do you mean code merge? | 13:52 |
@sonney2k | pull request | 13:52 |
blackburn | hah pull request | 13:52 |
alexlovesdata | ok | 13:52 |
blackburn | press release should be nice too | 13:52 |
wiking | alexlovesdata: pull request (putting code into shogun HEAD) | 13:52 |
alexlovesdata | ok, got that | 13:52 |
heiko | press realease :D | 13:52 |
alexlovesdata | soory, I am not used to chat acronyms, guys | 13:52 |
@sonney2k | so cwidmer/blackburn - want to continue? | 13:52 |
cwidmer | sure | 13:52 |
blackburn | yeah please | 13:52 |
cwidmer | backburn made quite some progress in implementing Multitask algorithms from the SLEP toolbox, the framework is now there, so it should be easier to bring in additional algos. Also, he helped in merging some of the code for Marius, Nico and mine ECML paper into shogun, which implements a LibLinear Style solver for graph-regularized MTL. Next steps are finishing up (some) stuff from SLEP, but also try what's already been ported on real data. | 13:53 |
cwidmer | 13:53 | |
cwidmer | blackburn, feel free to add things | 13:53 |
blackburn | okay after some struggles I get first algo proved to be correct (assuming SLEP is correct) | 13:53 |
blackburn | what I need to stress is now I have no foreign affairs (like bachelor thesis) | 13:54 |
blackburn | so I am going to continue with good pace | 13:54 |
@sonney2k | which reminds me - is everyone `free' now? | 13:54 |
blackburn | next after SLEP should be Ando/Zhang framework (for logit regression I think) | 13:54 |
@sonney2k | as in no more exams? | 13:54 |
uricamic | I am having an exam in 5 minutes, and then 2 more to come later, but still don't know when, unfortunately | 13:55 |
blackburn | sonney2k: I am sorry I have to leave now - professor is waiting for me unexpectedly - I'll check logs | 13:55 |
@sonney2k | uricamic, good luck then! | 13:55 |
uricamic | thanks | 13:55 |
@sonney2k | uricamic, and thanks for attending nevertheless... | 13:55 |
uricamic | no problem | 13:55 |
cwidmer | blackburn, alright, see you! | 13:55 |
@sonney2k | so I assume no comments to blackburn/cwidmer | 13:56 |
blackburn | so see you guys I will be online once I get to another city again | 13:56 |
@sonney2k | then heiko - your turn - tell us about your vacations in poland ;-) | 13:56 |
nicococo_ | nice ecml paper :) | 13:56 |
cwidmer | hehehe | 13:56 |
uricamic | ok, I also have to go now, see you | 13:56 |
heiko | well yes, I was sitting in a little house in the woods in the middle of nowhere and did quite some work :) | 13:56 |
heiko | I finished all MMD based two-sample tests (Quadratic time and linear time) - these are already committed. | 13:56 |
heiko | All tests assert the results against these of the MATLAB implementation. | 13:57 |
-!- blackburn [d5578b92@gateway/web/freenode/ip.213.87.139.146] has quit [Quit: Page closed] | 13:57 | |
heiko | Most work since last meething went into the approximations of null-distrubutions (Eigenspectrum, Gamma moment matching, Gaussian approximation, bootstrapping) | 13:57 |
heiko | For that I added some new pdf/cdf functions for Gamma, Normal, errorfunction etc. from ALGLIB | 13:57 |
heiko | The framework works well after extending the first draft a bit here and there -- and its pretty easy to use all tests now. | 13:57 |
heiko | This was about the first half of work from my proposal so I am in time. | 13:57 |
heiko | Next step are independence tests (HSIC) and kernel choice for large scale two-sample tests | 13:57 |
heiko | (based on linear MMD, I will need the qp solver for that) | 13:57 |
heiko | Also, there was this idea of creating a LaTeX document on the new methods which I want to start with soon. | 13:57 |
heiko | A word on SHOGUN 2.0. I would like to mention that there are many things which are not part of GSoC projects that should be finished before | 13:57 |
heiko | (serialisation, testing, documentation, etc) So I think it might be a bit early. | 13:57 |
@sonney2k | heiko, might be - we will have to see then. it is still > 2 months aways | 13:58 |
heiko | sonney2k, true, just that there are so many things. but youre right, lets see | 13:59 |
@sonney2k | and last release was ages ago so we have to release at least something (alpha or so) | 13:59 |
heiko | I mean | 13:59 |
heiko | when you release a 2.0 , people expect everythign to work | 13:59 |
heiko | thats just my point | 13:59 |
@sonney2k | well yes - it should work otherwise we cannot release | 14:00 |
@sonney2k | any questions to heiko? | 14:00 |
heiko | This year, we got gsomix to do all the usability stuff :) | 14:00 |
@sonney2k | if not os252/ puffin444 want to continue? | 14:00 |
puffin444 | sure | 14:00 |
os252 | sure - | 14:00 |
@sonney2k | heiko, yeah.. :D | 14:01 |
puffin444 | The whole GP framework is implemented in shogun. | 14:01 |
cheng_ | puffin444: "whole" == gpml in matlab? | 14:01 |
puffin444 | The model selection framework has been expanded to include | 14:01 |
puffin444 | gradient search | 14:01 |
puffin444 | as in the general infrastructure supports arbitrary inference methods, likelihood models, etc. | 14:02 |
@sonney2k | heiko, puffin444 btw please file bugs using the bugtracker on github whenever you notice some! | 14:02 |
heiko | sonney2k, yes good idea | 14:02 |
-!- ipondering [~iponderin@2001:da8:203:1823:4919:ecd1:c3d5:4242] has left #shogun [] | 14:02 | |
puffin444 | New inference methods, approx. methods, etc. are simply a matter of adding an extra class | 14:02 |
puffin444 | will do. | 14:03 |
puffin444 | The model selection part is just one memory leak bug away from a pull request. | 14:03 |
-!- ipondering [~iponderin@2001:da8:203:1823:4919:ecd1:c3d5:4242] has joined #shogun | 14:04 | |
heiko | I really like puffin444's extensions to the model selection framework. Its very flexible and modular :) | 14:04 |
-!- ipondering [~iponderin@2001:da8:203:1823:4919:ecd1:c3d5:4242] has left #shogun [] | 14:04 | |
puffin444 | This week I will add on/cleanup some of the details of the framework to make sure its ready for general use. | 14:04 |
-!- ipondering [~iponderin@58.194.224.114] has joined #shogun | 14:05 | |
-!- ipondering [~iponderin@58.194.224.114] has left #shogun [] | 14:05 | |
puffin444 | Add caching for GPs, add some more kernel derivatives, add some demos, etc. | 14:06 |
os252 | Perhaps relevant for cheng's question: after this is done, shogun will pretty much mirror the state where GPLML is for the standard regression stuff (not logistic regression and non-standard likelihoods). | 14:06 |
heiko | puffin444, Id like to push the ARD thing again :) | 14:06 |
@sonney2k | puffin444, os252 so you are on schedule? | 14:06 |
@sonney2k | it sounds like | 14:07 |
os252 | I'd say pretty much. Depending on how long this takes things may lack behind by a few days but Puffin444 really pushed this through up to this point! | 14:07 |
@sonney2k | very nice! | 14:08 |
os252 | heiko: ARd is merely a different kernle. | 14:08 |
puffin444 | I think technically the model selection should have been merged yesterday, but roughly yes. | 14:08 |
os252 | Agree the we should add that kernel though. Let's do this once we have the standard shogun kernels (lin,pol, gauss) ported. | 14:08 |
@sonney2k | further comments? | 14:08 |
@sonney2k | if not ... nice work! | 14:08 |
@sonney2k | and last one is cheng_/pluskid | 14:09 |
@sonney2k | please continue! | 14:09 |
pluskid | cheng_: you will do it or I do? | 14:09 |
cheng_ | pluskid: you go. | 14:09 |
pluskid | ok | 14:09 |
pluskid | Basically, I wrote a draft survey on multiclass learning, implemented two multiclass learning algorithms: Conditional Probability Tree and ShareBoost, ported liblbfgs (not remember correctly the date of the last meeting, so not sure any of the stuff here were already reported in the last meeting :p). | 14:09 |
pluskid | brief and no flooding :D | 14:09 |
@sonney2k | pluskid, cool - the survey is something we can use for doc again | 14:10 |
@sonney2k | pluskid, 25th of may so no | 14:10 |
pluskid | sonney2k: yeah, I discussed with cheng_ this morning about that | 14:10 |
@sonney2k | so at least MC / SO will have some doc | 14:10 |
cheng_ | @sonney2k: On that note, shogun badly needs a doco section about the methods. | 14:10 |
@sonney2k | I fully agree | 14:11 |
pluskid | but not sure yet where in shogun do we have a good place to put PDF( or LaTeX) docs | 14:11 |
@sonney2k | pluskid, well we have an extra repository shogun-doc | 14:11 |
@sonney2k | and blackburn wanted to put his stuff there | 14:11 |
pluskid | ah, cool | 14:11 |
pluskid | I'll check that | 14:11 |
@sonney2k | we should ask him to start taht now | 14:11 |
@sonney2k | that | 14:11 |
@sonney2k | otherwise we will never have a doc | 14:11 |
pluskid | and btw: a word for n4nd0, I think I just fixed the SWIG segfault right now | 14:12 |
n4nd0 | pluskid: good, what was it? | 14:12 |
@sonney2k | pluskid, what was it? | 14:12 |
cheng_ | Aim high... http://www.cs.waikato.ac.nz/ml/weka/book.html | 14:12 |
@sonney2k | I was betting on bad memory | 14:12 |
pluskid | some header file not being included | 14:12 |
pluskid | sonney2k, n4nd0: we can discuss about that after the meeting, I do have some questions about SWIG | 14:13 |
@sonney2k | ok | 14:13 |
pluskid | I wrote some in the pull request I sent just now | 14:13 |
@sonney2k | any further comments? | 14:13 |
@sonney2k | pluskid, what will you do next 6 weeks :D | 14:14 |
pluskid | sonney2k: I had a discussion about that this morning with cheng_ | 14:14 |
pluskid | I suggest that I go back to complete the survey before the midterm | 14:14 |
pluskid | and choose 2~3 algorithms out of those he suggested me | 14:14 |
pluskid | for the upcoming 6 weeks after mid-term | 14:15 |
@sonney2k | I vote for other massively multiclass stuff | 14:15 |
pluskid | btw, sonney2k do you hate ECOC related stuff? :p | 14:15 |
@sonney2k | all the other one-vs-rest is still very competitive | 14:15 |
@sonney2k | no | 14:15 |
@sonney2k | why should I? | 14:15 |
@sonney2k | only MKL :D | 14:15 |
@sonney2k | alright | 14:15 |
pluskid | I mean you said they do not perform very well (IIRC)? | 14:16 |
@sonney2k | one if there is nothing else | 14:16 |
@sonney2k | Mentors/Students please think about some nice, fancy graphical applications of your methods - some kind of eye candy we can use to advertise the great work you all were doing once GSoC is over that can also be used as examples. | 14:16 |
cheng_ | For multiclass, there is image-net. | 14:16 |
@sonney2k | cheng_, the book idea is nice - but it would require someone with lots of time :D | 14:18 |
@sonney2k | but in principle we got asked a couple of times... | 14:18 |
@sonney2k | and we have many mentors / students that could contribute chapters | 14:18 |
@sonney2k | anyway | 14:18 |
n4nd0 | sonney2k: I think we can do it together if we organize it good | 14:18 |
@sonney2k | lets end this | 14:18 |
@sonney2k | Finally when do we meet again? Is 3rd of August 11 UTC OK for everyone? | 14:18 |
cwidmer | fine with me | 14:19 |
@sonney2k | n4nd0, it is a 1-2 years project though | 14:19 |
os252 | 3rd is good, yes. | 14:19 |
alexlovesdata | fine with me | 14:19 |
vojtech | ok for me | 14:19 |
n4nd0 | for me too | 14:19 |
pluskid | I might be on trip that day | 14:19 |
nicococo_ | i'm on vacation until 11th of august | 14:19 |
cheng_ | Ok for me | 14:20 |
puffin444 | yep | 14:20 |
@sonney2k | heiko, ? | 14:20 |
heiko | yes ok | 14:20 |
@sonney2k | alright then we have one for each student /mentor combo o | 14:21 |
@sonney2k | wiking, will you be there? | 14:21 |
alexlovesdata | he is eating now | 14:21 |
alexlovesdata | will be back in 20 | 14:21 |
alexlovesdata | I can join anyway | 14:21 |
alexlovesdata | on 3rd august | 14:22 |
@sonney2k | alright | 14:22 |
@sonney2k | then thanks all for attending the meeting - let the submeetings start now | 14:22 |
@sonney2k | (and sorry that it was our longest one) | 14:22 |
cheng_ | Night guys! | 14:22 |
pluskid | I'm wondering, is it 3rd Aug. or 11th Aug. ? | 14:22 |
@sonney2k | 3rd | 14:22 |
pluskid | 11 UTC | 14:22 |
os252 | night cheng | 14:22 |
pluskid | I'm fool | 14:22 |
pluskid | then I'll be there, too | 14:22 |
@sonney2k | very good | 14:22 |
-!- cheng_ [73406f11@gateway/web/freenode/ip.115.64.111.17] has quit [Quit: Page closed] | 14:23 | |
@sonney2k | maybe we should use 11th at 11 o'clock always :D | 14:23 |
nicococo_ | i'm there :) | 14:23 |
cwidmer | ok, thanks sonney2k | 14:23 |
cwidmer | bye guys | 14:23 |
nicococo_ | bye chris | 14:23 |
n4nd0 | see you cwidmer | 14:23 |
-!- cwidmer [~chris@cpe-69-201-177-140.nyc.res.rr.com] has left #shogun ["Konversation terminated!"] | 14:24 | |
nicococo_ | thx to soeren | 14:24 |
@sonney2k | so vojtech / n4nd0 - isn't the unconstrained / constrained problem equivalent - I mean if you just compute unconstrained objectives f(W) for both - aren't you good? | 14:24 |
vojtech | n4nd0, I looked at the formulation in http://iglesiashogun.wordpress.com/2012/06/04/second-weekly-report-gsoc-2012/ and this can be easily expressed as an unconstrained problem hence BMRM can solve it | 14:24 |
nicococo_ | it is equivalent, but cannot solved with qp solver , right? | 14:24 |
vojtech | nicococo_, I don't understand it. You can solve both formulations using cutting planes | 14:25 |
n4nd0 | vojtech: is it expressing it as unsconstrained the same as using the dual formulation? | 14:25 |
vojtech | n4nd0, no | 14:25 |
n4nd0 | I am still a beginner with optimization :S | 14:26 |
vojtech | n4nd0, no problem. I can write a short draft explaining the realation | 14:26 |
nicococo_ | vojtech: i mean with a qp formulatation it is not solvable but bundle methods, subgradient descent do the job, right | 14:26 |
nicococo_ | ? | 14:27 |
vojtech | nicococo_, I'm not sure what you mean | 14:27 |
n4nd0 | vojtech: that would be great, if it doesn't take much time for you of course | 14:27 |
nicococo_ | vojtech: i cannot use quadprog to solve the unconstraint version | 14:27 |
vojtech | n4nd0, no problem I'll do it | 14:27 |
alexlovesdata | I think what you talk guys is also valuable for us | 14:28 |
vojtech | nicococo_, you cannot up to toy examples | 14:28 |
alexlovesdata | since we have in mind to try the BMRM solver | 14:28 |
alexlovesdata | vojtech: you means replace \xi_i = max(0, \Delta - \delta\Psi ) right? | 14:29 |
vojtech | yes | 14:29 |
nicococo_ | so we have a cutting plane solver and a nice bmrm, i think it is a good idea to have both | 14:30 |
nicococo_ | and they are interchangeable (more or less) | 14:30 |
nicococo_ | nice for comparisons :) | 14:30 |
vojtech | nicococo_, BMRM is a cutting plane solver in fact | 14:30 |
nicococo_ | oh yes, sorry :) | 14:31 |
vojtech | n4nd0, In my view both algorithms should solve the same problems and thus they can be easily compared | 14:32 |
vojtech | n4nd0, the only difference is the computational time | 14:33 |
n4nd0 | ok | 14:33 |
n4nd0 | it is a very good new that they should give the same results | 14:34 |
vojtech | n4nd0, BTW: additional linear constraints on W can be also easily added to BMRM | 14:34 |
nicococo_ | vojtech: i have first results concering slack-rescaled hmsvm we were talking about the last meeting | 14:34 |
vojtech | n4nd0, but then we will need to use Mosek QP | 14:34 |
n4nd0 | vojtech: aham I think those were interesting for an application nicococo_ said | 14:34 |
vojtech | nicococo_, I'm eager to see them if possible | 14:35 |
vojtech | n4nd0, we also use constrained BMRM when learning submodular MRF for image segmentation | 14:36 |
-!- cheng [~Adium@115-64-111-17.tpgi.com.au] has joined #shogun | 14:36 | |
nicococo_ | vojtech: just a number on toy data.. but let me tell you that it is really disappointing :( performance drops heavily and the OP takes much longer | 14:36 |
n4nd0 | nicococo_: you are talking about our method, or? | 14:36 |
n4nd0 | well, haha "our" :P | 14:37 |
n4nd0 | the method in PrimalMosekSOSVM | 14:37 |
nicococo_ | n4nd0: hehe.. no WE ;) do the margin rescaling.. but we recently tested slack rescaling | 14:37 |
-!- cheng [~Adium@115-64-111-17.tpgi.com.au] has left #shogun [] | 14:38 | |
n4nd0 | WE? | 14:38 |
nicococo_ | you and me :) | 14:38 |
n4nd0 | ah, ok | 14:38 |
vojtech | nicococo_, what exactly did you test? I think slack-rescaling and margin rescaling are are equivalent for multi-class classification | 14:38 |
vojtech | or did you consider some SO problem ? | 14:39 |
nicococo_ | i tested on sequences | 14:39 |
vojtech | I see | 14:39 |
nicococo_ | any idea? | 14:39 |
vojtech | how do you evaluate slack-rescaling loss on sequences ? | 14:39 |
nicococo_ | you mean the argmax? | 14:39 |
vojtech | the slack-rescaling loss is not tractable, namely, loss augmented classification problem | 14:40 |
nicococo_ | well it is in quadratic time.. it is the work of one of our master students | 14:41 |
nicococo_ | but it is not an easy problem and its a pitty that there is no gain... | 14:42 |
vojtech | slack-recaling requires solving argmax_y L(y,y')(1-w,psi(x,y)+w,psi(x,y')) , how can you solve it is y is a sequence? | 14:43 |
vojtech | do you solve it exactly or use a kind of approximation? | 14:44 |
nicococo_ | its still a dynamic program | 14:44 |
nicococo_ | we have both | 14:44 |
vojtech | what kind of loss do you use ? | 14:44 |
nicococo_ | i was talking about exact inference | 14:44 |
nicococo_ | we compared to brute force method | 14:45 |
nicococo_ | 0-1 hamming loss | 14:45 |
vojtech | does brute force mean trying all Y^n sequences ? | 14:46 |
vojtech | I don't see how this can be solved by DP | 14:46 |
nicococo_ | yes :) well, at some point you have to compare if the implementation works | 14:46 |
nicococo_ | as soon as the student finishs the master thesis i can send it to you | 14:47 |
vojtech | ok, I'm looking forward | 14:47 |
nicococo_ | (may take some more month :) ) | 14:47 |
vojtech | last question, you really mean the slack-rescaling loss: max_y L(y,y')(1+<w,psi(x,y')-psi(x,y)>) where y is from Y^N ? | 14:48 |
vojtech | and L(y,y') Hamming distance | 14:49 |
nicococo_ | yes, i really mean that.. the loss is no longer decompasable as in margin rescaling | 14:49 |
vojtech | ok, very good | 14:49 |
nicococo_ | no it isn't, thats the prob :) | 14:50 |
nicococo_ | i thought slack rescaling would be superior, but the results show the opposite | 14:51 |
vojtech | this is also interesting negative results | 14:52 |
vojtech | ok, I need to go. | 14:52 |
vojtech | n4nd0, I'll send you the draft | 14:52 |
nicococo_ | ok, ciao | 14:52 |
vojtech | CU | 14:52 |
n4nd0 | vojtech: thank you, bye! | 14:52 |
-!- vojtech [9320543b@gateway/web/freenode/ip.147.32.84.59] has quit [Quit: Page closed] | 14:53 | |
nicococo_ | n4nd0: in the first pdf i sent around, all so formulation are unconstraint | 14:53 |
nicococo_ | n4nd0: did you had a look at the new doc i sent you yesterday? | 14:54 |
n4nd0 | do you mena in this one https://dl.dropbox.com/u/11020840/shogun/so.pdf? | 14:54 |
n4nd0 | nicococo_: yeah sure, I have been thinking of the design for the hmsvm | 14:55 |
n4nd0 | nicococo_: I have a couple of questions about it, as usual :) | 14:55 |
nicococo_ | n4nd0: cool :) .. (and yes so.pdf e.g. Eqn.1 is unconstraint.. if you disregard the general ax<=b aso) | 14:58 |
n4nd0 | nicococo_: ok I see | 14:58 |
n4nd0 | nicococo_: what I don't have so clear is what kind of solver one needs to use there since the argument of minimize (w) is affected by the loss function in the sum term | 14:58 |
nicococo_ | yes, you can't use anymore mosek if you want to solve it like that | 14:59 |
n4nd0 | I guess that the solver uricamic has used solves that | 15:00 |
nicococo_ | correct ;) it is an advanced solver that is likely to need less iterations than ours.. | 15:01 |
n4nd0 | but the results, the solution for w, doesn't need to be the same right? | 15:01 |
n4nd0 | or is there some kind of equivalence in the problem representation? | 15:01 |
nicococo_ | theoretically yes, in practice one would not optimize ad infinitum | 15:02 |
n4nd0 | does theoretically yes mean yes into our problem? | 15:02 |
n4nd0 | I mean I solve what we currently have | 15:02 |
n4nd0 | minimize_w,xi 0.5 w'*C*w + sum xi_i | 15:03 |
n4nd0 | s.t. dPsi_i(y) >= Delta(y_i, y) - xi_i for all i and y in Y \ y_i | 15:03 |
nicococo_ | n4nd0: practically they are not exactly same but within an epsilon region (user can set this eps) | 15:03 |
n4nd0 | where is the epsilon in these formulations? | 15:04 |
nicococo_ | so if your objective is 2845 there might be 2844.99 | 15:04 |
n4nd0 | ok, I could try to do that | 15:05 |
n4nd0 | interesting | 15:05 |
n4nd0 | in any case, let's talk about hmsvm that is what we are concerned about now | 15:05 |
n4nd0 | ok? | 15:05 |
nicococo_ | yes, but is good to have both algorithm | 15:05 |
nicococo_ | okay | 15:05 |
n4nd0 | so one thing I found a bit new from what I expected is in the data representation | 15:06 |
n4nd0 | I have noted that each *feature* is a sequence itself | 15:06 |
n4nd0 | not each training example | 15:06 |
n4nd0 | I mean a training example, an x_i | 15:06 |
nicococo_ | a single training example is a matrix now | 15:07 |
n4nd0 | can be formed of several features, e.g. 10 features is the number they use in the hmsvm toolbox | 15:07 |
n4nd0 | ok, I understand | 15:07 |
n4nd0 | I am not quite sure how this is going to affect the use of a CFeatures shogun type | 15:08 |
n4nd0 | I think that we shouldn't use a CDotFeatures for this then | 15:08 |
n4nd0 | do you know what I mean? | 15:08 |
nicococo_ | you have free hands on that | 15:09 |
n4nd0 | I think I should do something new that inherits from CFeatures for that | 15:10 |
nicococo_ | of course one could reshape matrices into vector form | 15:10 |
nicococo_ | but that would be a hack (to reshpae) | 15:10 |
n4nd0 | yeah ... but still, it won't be a good idea to use something like DotFeatures or DenseFeatures for that, after reshaping | 15:10 |
n4nd0 | just because of the semantics of the operations | 15:11 |
n4nd0 | we would have something like get_feature_vector or the like | 15:11 |
n4nd0 | that would return something that it is not a feature vector actually | 15:11 |
n4nd0 | do you understand what I mean? | 15:11 |
nicococo_ | i understand its not a good idea, so any solution yet? | 15:12 |
n4nd0 | yes, to create a new type :) | 15:13 |
n4nd0 | but before I do that, I'd like to know what operations do we want to do with this type | 15:13 |
n4nd0 | as I understand right now, we would need something like creating a combined feature vector right? | 15:14 |
nicococo_ | perfect, do we talk about Cstructureddata again :) | 15:14 |
-!- puffin444 [62e3926e@gateway/web/freenode/ip.98.227.146.110] has quit [Quit: Page closed] | 15:14 | |
n4nd0 | mmm not exactly | 15:14 |
n4nd0 | this won't be a CStructuredData but a sub-type of CFeatures | 15:15 |
nicococo_ | something like a block-feature | 15:15 |
n4nd0 | what do you mean with that? | 15:15 |
nicococo_ | i guess, it is the same as your combined feature vector | 15:16 |
n4nd0 | I don't think so | 15:16 |
n4nd0 | because the combined feature vector is a "well-defined" vector | 15:16 |
n4nd0 | but we have just said that each training example is a matrix right? | 15:17 |
nicococo_ | go on | 15:18 |
n4nd0 | ok... | 15:18 |
n4nd0 | so what do we need to do with the features (I am talking about the x)? | 15:18 |
n4nd0 | just to compute joint features from them? | 15:19 |
nicococo_ | i don't understand: x is a matrix and the joint feature vectors are computed in CStructuredModel given x | 15:20 |
n4nd0 | ok, let's see | 15:20 |
n4nd0 | x (each of the training examples) is a matrix | 15:20 |
n4nd0 | ok? | 15:20 |
nicococo_ | check | 15:20 |
n4nd0 | I need to have all the training examples in a data type | 15:21 |
n4nd0 | something that inherits from CFeatures | 15:21 |
n4nd0 | ok? | 15:21 |
nicococo_ | ok | 15:21 |
n4nd0 | let's call it, mmm .... CHMSVMFeatures for example | 15:21 |
n4nd0 | let's define the operations we need to do with CHMSVMFeatures | 15:22 |
nicococo_ | ok | 15:22 |
n4nd0 | I just see one operation | 15:22 |
n4nd0 | take one feature in CHMSVMFeatures and on label (also a sequence) an create a Psi | 15:22 |
nicococo_ | stop | 15:23 |
nicococo_ | do you mean something like a psi=get_jfv_from(example id) in CHMSVMFeatures | 15:24 |
nicococo_ | ? | 15:24 |
n4nd0 | I am not thinking yet where it should be | 15:25 |
n4nd0 | I am just thinking what do we want to do with them | 15:25 |
nicococo_ | but what we want to do is decoupled from the CHMSVMFeature, so the above function should be in CStructModel | 15:26 |
nicococo_ | i only see a matrix=get_example(int i) for CHMSVMfeatures | 15:26 |
n4nd0 | yes | 15:26 |
n4nd0 | my point was that at the current level of asking what do we want to do with CHMSVMFeatures it is not relevant where the code for the operation is, that should come later I think | 15:27 |
n4nd0 | anyway | 15:27 |
n4nd0 | I think we are both thinking of the same so it is ok | 15:27 |
n4nd0 | just note that probably, in order to be able to support coffin, this is something that should probably change | 15:27 |
n4nd0 | but I don't think we are concerned with that yet | 15:28 |
n4nd0 | are we ok so far? | 15:29 |
nicococo_ | okay, we could still use coffin afterwards.. yes, we are ok | 15:29 |
n4nd0 | there is something that I don't know what role plays in all of this | 15:30 |
n4nd0 | that is the state | 15:31 |
n4nd0 | as I undestand it | 15:31 |
n4nd0 | a state is an element of the label sequence | 15:31 |
nicococo_ | yes.. if the label sequence is 0 1 2 3 1 2 3 the numbers correspond to the state | 15:32 |
n4nd0 | BUT 0 is an state, 1 another, 2 another | 15:32 |
n4nd0 | and so on | 15:32 |
n4nd0 | right? | 15:32 |
nicococo_ | yes | 15:32 |
n4nd0 | ok | 15:32 |
n4nd0 | then there is no need to have an object itself for each state | 15:33 |
n4nd0 | ? | 15:33 |
nicococo_ | no, i don't think thats necessary | 15:33 |
n4nd0 | all right | 15:34 |
nicococo_ | but... mmmhh | 15:34 |
n4nd0 | tell me | 15:34 |
nicococo_ | i mean for our example it would be overkill.. | 15:34 |
nicococo_ | no, its okay.. we don't use classes for each state | 15:35 |
n4nd0 | ok | 15:35 |
n4nd0 | so for the moment I don't know exactly how the argmax should work | 15:36 |
n4nd0 | I have read about the Viterbi algorithm but not especifically for this application | 15:37 |
nicococo_ | i will write that in the dcoument.. i also have a c-code viterbi | 15:37 |
n4nd0 | ok | 15:37 |
n4nd0 | I was planning to look into the hmsvm toolbox | 15:37 |
nicococo_ | hopefully by tomorrow afternoon i will sent it to you | 15:37 |
n4nd0 | but I think that would be even greater | 15:37 |
n4nd0 | sure, no problem | 15:37 |
n4nd0 | I can start sketching the classes | 15:38 |
n4nd0 | and after that working out the psi comptuation | 15:38 |
nicococo_ | you can also start to think/implement a toy example generator | 15:38 |
n4nd0 | I think I have not more doubts ... | 15:38 |
n4nd0 | would the simulate_data in hmsvm toolbox be enough for that? | 15:38 |
n4nd0 | I have read it this morning, it is quite straightforward in my opinion | 15:39 |
nicococo_ | yep, totally :) | 15:39 |
nicococo_ | it could be even simplier. | 15:39 |
n4nd0 | I think it is simple enough ;) | 15:39 |
n4nd0 | there is a thing that kind of bugs me though ... | 15:40 |
nicococo_ | ..and i will have a look for real data. | 15:40 |
nicococo_ | tell me | 15:40 |
n4nd0 | right now I have a better idea of the hmsvm than I had a couple of weeks ago | 15:40 |
n4nd0 | when I told you that I thought it was just to implement a psi function, an argmax and a loss | 15:40 |
n4nd0 | you answered that it was much more than that | 15:40 |
n4nd0 | but still, currently I see no more than that :) | 15:41 |
nicococo_ | well, the hmsvm toolbox has more than 3 files, right ;) | 15:41 |
nicococo_ | but we agreed to have another setting | 15:41 |
n4nd0 | yes... but that's kind of ambiguous | 15:42 |
n4nd0 | I meant the the thing of the 3 files | 15:42 |
nicococo_ | which is easier to implement and thx to coffin (later) we have something powerfull | 15:42 |
n4nd0 | there are also other things that are already done, like the optimization algo | 15:42 |
nicococo_ | still, if you have a look a the state model directory there are 7 files or so.. and this is only the state model | 15:43 |
n4nd0 | mmm | 15:43 |
n4nd0 | so we still don't use that or? | 15:43 |
nicococo_ | we start simple.. but i already have some nice additions in mind :) | 15:44 |
n4nd0 | cool, I like the idea | 15:45 |
n4nd0 | then let me do some work today and let's talk tomorrow again? | 15:45 |
nicococo_ | okay, so, i guess i will have to write a really nice midterm evaluation ;) | 15:46 |
n4nd0 | :) | 15:46 |
n4nd0 | when is it good for you to meet tomorrow? | 15:46 |
nicococo_ | see you tomorrwo and bye .. | 15:46 |
n4nd0 | is 16h ok? | 15:46 |
nicococo_ | ok | 15:46 |
n4nd0 | cool | 15:47 |
-!- nicococo_ [5e8b1a4f@gateway/web/freenode/ip.94.139.26.79] has quit [Quit: Page closed] | 15:47 | |
n4nd0 | tty tomorrow then | 15:47 |
n4nd0 | thanks for the help! | 15:47 |
n4nd0 | sonney2k: around? | 15:49 |
@sonney2k | n4nd0, back | 16:14 |
@sonney2k | wiking, alexlovesdata are you back? | 16:14 |
wiking | sonney2k: me yes | 16:14 |
wiking | alexlovesdata: is going to eat | 16:15 |
@sonney2k | wiking, so what hinders the merge? | 16:15 |
wiking | so i had 2 branches | 16:15 |
wiking | branch a) was started before any gsoc project | 16:15 |
wiking | that never got rebased, and then i hadn't rebased for a long time | 16:16 |
wiking | and e.g. labels, machines had been changed | 16:16 |
wiking | so now when i've tried rebasing it breaks apart | 16:16 |
@sonney2k | sure | 16:16 |
@sonney2k | ok and now? | 16:16 |
wiking | and with branch b) i have a cherrypicked SO branch of fernando | 16:17 |
wiking | plus that 'broken' a) | 16:17 |
@sonney2k | wiking, which files did you touch? | 16:17 |
wiking | i mean it's forked from a) applied some of fernando's so development | 16:17 |
@sonney2k | I understand | 16:17 |
@sonney2k | so which files did you add / modify? | 16:18 |
wiking | so i've touched some minors in example.. then the whole label system and svmocas and machine as well | 16:18 |
wiking | so now i need to rebase this stuff | 16:18 |
wiking | somehow | 16:18 |
wiking | i can merge by hand | 16:18 |
wiking | i've started to do that | 16:19 |
@sonney2k | if it really is that far away now - you probably have no other option than merge everything by hand (either checkout HEAD and cherry pick patches and merge or copy stuff over manually) | 16:19 |
@sonney2k | for example using meld | 16:19 |
wiking | meld? | 16:19 |
@sonney2k | meld | 16:19 |
wiking | just a sec | 16:19 |
wiking | it's like a visual merger? | 16:19 |
@sonney2k | nice program to compare two dirs | 16:19 |
@sonney2k | yes | 16:19 |
wiking | ah ok | 16:19 |
@sonney2k | visual diff | 16:19 |
@sonney2k | very convenient for big changes | 16:20 |
wiking | ok | 16:20 |
wiking | lemme install | 16:20 |
n4nd0 | sonney2k: I don't know if you read something from my conversation with Nico; anyway, the case is that for the hmsvm I need to use a special kind of features since each training example is a matrix itself (each of the features are sequences of integers), that's all right for me, I thought of doing a new CHMSVMFeatures (or so) that inherits from CFeatures. However, I wonder if there's something already in shogun for this purpose since I believe tha | 16:20 |
@sonney2k | wiking, in any case - this really is the top priority for you now | 16:20 |
wiking | OMFG (starts with installing perl...) | 16:20 |
wiking | aparently i didn't have perl | 16:20 |
@sonney2k | wiking, get the merge done this week | 16:20 |
n4nd0 | sonney2k: sorry for the big message, I didn't want to interfere with your conver with wiking | 16:20 |
wiking | sonney2k: ok! | 16:20 |
@sonney2k | wiking, we really cannot let you pass w/o this being in shogun main/head/master! | 16:21 |
wiking | sonney2k: thursday midnight,ok | 16:21 |
wiking | ? | 16:21 |
wiking | just to set something in stone | 16:21 |
@sonney2k | sure, the earlier the better | 16:21 |
wiking | i'll try my best | 16:21 |
wiking | but i want to say that i can arrive to that moment for sure | 16:21 |
wiking | on thursday midnight | 16:21 |
@sonney2k | wiking, then do it | 16:21 |
wiking | ok | 16:21 |
wiking | sonney2k: if there's something that u want to talk about with alexlovesdata and me then we'll have a meeting around 19:30 | 16:22 |
@sonney2k | wiking, that really is my main concern | 16:24 |
@sonney2k | wiking, so basically only you | 16:24 |
wiking | sonney2k: ok, i'm just saying | 16:24 |
@sonney2k | wiking, one suggestion | 16:24 |
@sonney2k | merge small pieces | 16:24 |
@sonney2k | if you can do PR's for each of them | 16:24 |
wiking | sonney2k: ok... i'll try by concept | 16:24 |
wiking | but maybe it'll be by file... | 16:24 |
@sonney2k | so if you did extensions to stuff that is already there, merge these first | 16:25 |
@sonney2k | and PR | 16:25 |
@sonney2k | and btw it seems like thursday is the euro 2012 game germany vs italy | 16:26 |
@sonney2k | so I will be awake :D | 16:26 |
@sonney2k | wiking, n4nd0 - regarding n4nd0 Q above - that seems a bit like image features ... | 16:29 |
n4nd0 | interesting match ... I won't say who I am supporting :D | 16:29 |
@sonney2k | n4nd0, one thing that would work is use StringFeatures | 16:29 |
@sonney2k | n4nd0, spain I guess :D | 16:30 |
n4nd0 | yeah, but I meant in Germany vs Italy ;) | 16:30 |
@sonney2k | n4nd0, you don't want italy to be in the final with spain don't you? | 16:30 |
n4nd0 | sonney2k: but I think that the sequences are of the same length so maybe StringFeatures are not really the case | 16:31 |
n4nd0 | sonney2k: I have not yet decided if I want Italy or Germany | 16:31 |
n4nd0 | sonney2k: Germany is pretty strong this eurocup, although it was in the last one and in the worlcup ... | 16:31 |
n4nd0 | but first, let's see what happens against Portugal :O | 16:32 |
@sonney2k | n4nd0, it was always before meeting spain :D | 16:32 |
@sonney2k | and italy for that matter | 16:32 |
@sonney2k | n4nd0, well the germans did beat portugal | 16:32 |
@sonney2k | so it shouldn't be too difficult | 16:32 |
CIA-18 | shogun: Chiyuan Zhang master * rd3bb757 / (2 files): fix SO modular compile segfault. - http://git.io/nQoqIQ | 16:33 |
CIA-18 | shogun: Soeren Sonnenburg master * r7c20b8c / (2 files): Merge pull request #602 from pluskid/fix-swig-segfault - http://git.io/JTH3kw | 16:33 |
n4nd0 | sonney2k: so is it ok for you if I do this CHMSVMFeatures or do you think it is unnecessary? | 16:35 |
n4nd0 | I'd rather do them since I think that to think of them as StringFeatures might lead to confussion | 16:35 |
@sonney2k | n4nd0, then rather do CMatrixFeatures | 16:36 |
n4nd0 | good idea :) | 16:36 |
@sonney2k | n4nd0, you should check if you can derive from DenseFeatures for that | 16:37 |
@sonney2k | hmmhh maybe not | 16:37 |
@sonney2k | well yes | 16:37 |
@sonney2k | you can | 16:37 |
@sonney2k | if matrices are all of same shape | 16:37 |
@sonney2k | then you derive from DenseFeatures and add num_rows / num_cols | 16:38 |
n4nd0 | mmm I was thinking of having an SGNDArray instead of an SGMatrix | 16:38 |
n4nd0 | for the example I have seen in the hmsvm toolbox, all the matrices were of the same shape | 16:39 |
n4nd0 | I don't know if that is always the case though, I should ask Nico about that | 16:39 |
shogun-buildbot | build #1115 of libshogun is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/libshogun/builds/1115 | 16:39 |
@sonney2k | n4nd0, well or you copy everything from DenseFeatures and just replace SGMatrix with SGNDArray | 16:39 |
@sonney2k | shogun-buildbot, finally! | 16:39 |
shogun-buildbot | What you say! | 16:39 |
@sonney2k | and SGVector with SGMatirx | 16:39 |
n4nd0 | all right, thank you | 16:45 |
n4nd0 | I have to go now, bye | 16:46 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 16:46 | |
-!- pluskid [~pluskid@202.130.113.141] has quit [Quit: Leaving] | 17:09 | |
-!- os252 [c07c1afb@gateway/web/freenode/ip.192.124.26.251] has quit [Quit: Page closed] | 17:25 | |
-!- uricamic [~uricamic@2001:718:2:1634:3d67:c999:8cc9:d056] has quit [Quit: Leaving.] | 17:32 | |
-!- blackburn [~blackburn@31.28.43.76] has joined #shogun | 18:05 | |
@sonney2k | the black burn blackburn is back | 18:50 |
blackburn | sonney2k: yeah back | 19:00 |
blackburn | sonney2k: enjoying my new ssd powered book | 19:00 |
heiko | ah, the joys of ssd :) | 19:01 |
blackburn | full reboot - 20s | 19:01 |
-!- alexlovesdata [82955843@gateway/web/freenode/ip.130.149.88.67] has quit [Ping timeout: 245 seconds] | 19:01 | |
heiko | yeh same here :) | 19:01 |
blackburn | however time to remove shindows | 19:01 |
heiko | I didnt even install it on my new computer :;) | 19:02 |
blackburn | was installed already | 19:02 |
heiko | oh, cant you buy without? | 19:02 |
blackburn | I can return money after I buy it | 19:03 |
blackburn | but it is just 18$ and takes 6 or 7 days | 19:03 |
heiko | oh, well then :) | 19:03 |
blackburn | I'd rather just remove it | 19:03 |
blackburn | huh i5 is a warm guy | 19:04 |
heiko | what? | 19:04 |
blackburn | I have intel core i5 here | 19:04 |
heiko | same here | 19:04 |
blackburn | in 13.3" pretty small notebook | 19:04 |
blackburn | makes my legs warm :D | 19:04 |
heiko | oh | 19:04 |
heiko | mmh well depends on electronics in there | 19:05 |
heiko | my thinkpad rather gets warm in programming use | 19:05 |
heiko | but does get *really* warm when I start 4 matlab jobs :D | 19:05 |
blackburn | ok time to choose linux distribution! | 19:05 |
heiko | oh, thats a pain, there are so many :) | 19:05 |
heiko | I currently use Ubuntu because its so easy to install | 19:06 |
heiko | but regulary have problems with it since they change things all the time | 19:06 |
blackburn | I tend to ubuntu too heh | 19:06 |
@sonney2k | blackburn, heiko I need some feedback - what should the new buildbot do? | 19:06 |
heiko | well its nice and robust, some minor problems but these are not too bad | 19:06 |
@sonney2k | slave that is | 19:06 |
blackburn | sonney2k: popcorn, cocacolar | 19:06 |
blackburn | :D | 19:06 |
blackburn | cocacola* | 19:06 |
@sonney2k | options are: we have 2 buildslaves | 19:06 |
heiko | what? :) | 19:06 |
@sonney2k | both doing the same thing | 19:06 |
@sonney2k | or alternatively one is building on debian - unstable all the time | 19:07 |
@sonney2k | the other on stable | 19:07 |
blackburn | latter | 19:07 |
@sonney2k | or one is building every few hours | 19:07 |
@sonney2k | and potentially updating the documentation | 19:07 |
@sonney2k | such that we have the latest one under /doc/latest | 19:07 |
blackburn | sonney2k: what is version of gnome-shell in your debian (sid?) | 19:07 |
heiko | sonney2k, I dont get the problem | 19:08 |
@sonney2k | blackburn, 3.4.1-8 | 19:08 |
blackburn | thanks | 19:08 |
@sonney2k | heiko, we have a buildmaster to trigger stuff | 19:08 |
@sonney2k | heiko, then we have a buildslave hosted at TU | 19:08 |
@sonney2k | (the current one) | 19:09 |
@sonney2k | now we have new resources and could either just use 2 build slaves | 19:09 |
@sonney2k | doing the same thing | 19:09 |
@sonney2k | (so if someone pushes a lot all the commits are quickly rebuild one by one) | 19:09 |
@sonney2k | or it can do other shogun stuff | 19:09 |
heiko | mmh | 19:10 |
heiko | is it that urgent that everything is rebuilt so quickly? | 19:11 |
@sonney2k | I dont' think so | 19:11 |
heiko | and what would be the other stuff you mentioned? | 19:11 |
@sonney2k | I am also no longer sure we really need a build for each interface | 19:11 |
@sonney2k | why not just build it once | 19:11 |
@sonney2k | for everything | 19:11 |
@sonney2k | if it fails | 19:11 |
@sonney2k | it fails | 19:12 |
heiko | lol :) | 19:12 |
@sonney2k | http://shogun-toolbox.org/buildbot/waterfall | 19:12 |
heiko | yeah seems a bit weird, there are not bug differences in the interfaces, right? if one fails and the other one doesnt, its likely to be a minor mistake | 19:12 |
@sonney2k | I mean this picture could be reduced to 3 columns | 19:12 |
heiko | oh yes that would be nice | 19:12 |
heiko | I tend to not really look because its so messy | 19:13 |
heiko | ok, then I am for that | 19:13 |
heiko | reduce it to one | 19:13 |
heiko | use the other computer for different stuff | 19:13 |
@sonney2k | heiko, I think I will do it the following way: | 19:14 |
@sonney2k | 1) build libshogun (this is quick) | 19:14 |
@sonney2k | 2) build static interfaces (if libshogun was OK) | 19:14 |
@sonney2k | 3) build modular interfaces (if libshogun was OK) | 19:15 |
@sonney2k | so 3 columns | 19:15 |
heiko | sounds good ! | 19:15 |
@sonney2k | heiko, yeah but this is currently only built for debian stable | 19:16 |
@sonney2k | maybe we should build it for some other architecture too | 19:16 |
heiko | and you want to add mac/win? | 19:16 |
heiko | yes that would be a good idea | 19:16 |
@sonney2k | mac is tough to get it seems :/ | 19:16 |
heiko | so many people use macs nowadays, would be good if the max port would always run | 19:16 |
@sonney2k | windows -> cygwin yes | 19:16 |
blackburn | f ck win | 19:16 |
heiko | and windows also, my old professor in germany only had windows machines | 19:17 |
blackburn | I have no idea how to use cygwin | 19:17 |
heiko | I had to include a live debian with shogun installed for my bachelor thesis | 19:17 |
blackburn | most inconvenient shit ever | 19:17 |
heiko | still widely used, blackburn | 19:17 |
heiko | (Most scientists here in London have macs) | 19:18 |
@sonney2k | yeah many have | 19:18 |
blackburn | macs are cool but supporting win is pain | 19:18 |
heiko | actually ALL ML lecturers had macs :D | 19:18 |
@sonney2k | heiko, problem really is that one can not easily get a mac as build machine | 19:18 |
heiko | not that I think of it | 19:18 |
heiko | mmh | 19:18 |
@sonney2k | or if not a mac emulate one to have it | 19:19 |
heiko | yeah what about that? | 19:19 |
@sonney2k | and I don't like the idea of buying some mac from gsoc money just for that | 19:19 |
heiko | with a VM perhaps? | 19:19 |
heiko | I understand | 19:19 |
@sonney2k | heiko, is not so easy | 19:19 |
heiko | no dont do that | 19:19 |
heiko | dont many people have that problem? I mean many software developers? | 19:20 |
@sonney2k | that new server is just 26 EUR/month | 19:20 |
blackburn | better buy pro with retina for yourself | 19:20 |
-!- alexlovesdata_ [82955843@gateway/web/freenode/ip.130.149.88.67] has joined #shogun | 19:20 | |
heiko | what about renting a mac server? | 19:20 |
blackburn | you deserved it :) | 19:20 |
@sonney2k | blackburn, you know *I* don't get the money right? | 19:20 |
heiko | i dont know if thats possible | 19:20 |
blackburn | sonney2k: who will? | 19:20 |
@sonney2k | I put it on some extra bank account - 'shogun toolbox foundation' | 19:21 |
@sonney2k | no idea though how to turn us into some official non profit organization | 19:21 |
@sonney2k | so the money is sitting there and is used only for the benefit of shogun | 19:21 |
blackburn | well I am pretty sure you may do what you want | 19:22 |
blackburn | why? | 19:22 |
@sonney2k | hardware, servers, and maybe for a meeting | 19:22 |
heiko | meeting would be cool :) | 19:22 |
@sonney2k | it is not a lot though right | 19:22 |
@sonney2k | $ 500 / student | 19:22 |
heiko | I dont need money to get to Berlin, or to stay there. | 19:22 |
@sonney2k | and I have no idea what german tax will do with it | 19:22 |
@sonney2k | can very well be that they just halve the money as they consider it income | 19:23 |
@sonney2k | (my income) | 19:23 |
blackburn | I have one month to come with some paper for ICPRAM @ Barcelona | 19:23 |
heiko | Perhaps we should find out how to register shogun as non-profit orga to avoid this? | 19:23 |
@sonney2k | I am actually a bit afraid of that - google paid attendence for GSoC meeting last year | 19:23 |
blackburn | I want to visit barcelona! | 19:23 |
blackburn | :D | 19:23 |
heiko | sonney2k, I dont have any income currently and wouldnt have to pay taxes if that is a problem | 19:24 |
@sonney2k | only because you two don't want to be mentors you cannot get invited | 19:24 |
blackburn | sonney2k: send your money to some cyprus bank account :D all russians do that | 19:24 |
heiko | aah | 19:24 |
@sonney2k | hehe | 19:24 |
@sonney2k | heiko, I will do my tax declaration for last year soon | 19:25 |
heiko | sonney2k, dont worry, I will be mentor next year, this year was a bit hard money wise | 19:25 |
@sonney2k | and if someone complaints we do it your way :D | 19:25 |
heiko | ok :) | 19:25 |
@sonney2k | there is always an if we get in and if we survive this one :D | 19:25 |
@sonney2k | no but it is actually looking good | 19:25 |
@sonney2k | I was pretty pleased with the lots and lots of progress today | 19:26 |
heiko | yeah for sure, if all conditions are met I mean | 19:26 |
heiko | and if I dont get a heart-attack from all the stress here :D | 19:26 |
blackburn | hey do you know log sum exp trick? | 19:26 |
heiko | sonney2k, I think its awesome whats currently happening to SHOGUN | 19:26 |
@sonney2k | blackburn, you mean the one in shogun? | 19:27 |
blackburn | no generally | 19:27 |
heiko | blackburn, yes :) was in one of my exams last month | 19:27 |
blackburn | I didn't know about that | 19:27 |
@sonney2k | blackburn, CMath::log_sum ? | 19:27 |
@sonney2k | maybe I don't | 19:27 |
@sonney2k | what is it? | 19:27 |
@sonney2k | (if not the CMath:: one) | 19:27 |
blackburn | sonney2k: in slep I encountered the following code | 19:27 |
heiko | no its to avoid numerical problems when evaluating a log of a special form | 19:27 |
blackburn | double aa = ...; | 19:27 |
blackburn | double bb = CMath::max(aa,0); | 19:28 |
@sonney2k | heiko, sounds like the one in CMath | 19:28 |
heiko | might be | 19:28 |
blackburn | fun_x += CMath::log(CMath::exp(bb) + CMath::exp(aa-bb)) + bb | 19:28 |
blackburn | I had no idea what is it | 19:28 |
heiko | its used a lot in all these graphical models | 19:28 |
blackburn | encountered that in logistic regression | 19:29 |
@sonney2k | log( exp(a) + exp(b)) = a + log (1 + exp (b-a)) where a = max(p,q) and b min(p,q). | 19:29 |
blackburn | right | 19:29 |
heiko | yeah | 19:29 |
@sonney2k | that is the one I implemented (documented in Cmath) | 19:29 |
blackburn | cool | 19:29 |
@sonney2k | back in the last century | 19:29 |
@sonney2k | for HMMs | 19:29 |
@sonney2k | heiko, yes HMMs... :D | 19:29 |
heiko | indeed :) | 19:30 |
@sonney2k | heiko, blackburn so what other architectures do we want? | 19:30 |
@sonney2k | maybe latest ubuntu? | 19:30 |
blackburn | sonney2k: archs like? mac, win? | 19:31 |
blackburn | ah | 19:31 |
@sonney2k | and redhat? | 19:31 |
heiko | so far only debian stable? | 19:31 |
@sonney2k | I would say latest debian - stable | 19:31 |
@sonney2k | for sure | 19:31 |
blackburn | redhat?? | 19:31 |
@sonney2k | no idea | 19:31 |
blackburn | I'd rather choose suse | 19:31 |
heiko | perhaps the most popular one | 19:31 |
heiko | that would be ubuntu or mint | 19:31 |
gsomix | android, hehe | 19:31 |
blackburn | ubuntu is debian | 19:31 |
blackburn | what is different? | 19:31 |
heiko | true but sometimes they change the packet structure a bit | 19:31 |
heiko | also its debian unstable | 19:32 |
@sonney2k | gsomix, btw heiko complained about some errors in CMap ... | 19:32 |
@sonney2k | heiko, difference might be minor | 19:32 |
heiko | I did not complain, just saw some, already corrected, will submit them soon | 19:32 |
@sonney2k | we could of course create .deb's automagically | 19:32 |
heiko | that would be a great thing to have | 19:32 |
blackburn | and set up a ppa | 19:32 |
heiko | many people dont like compiling | 19:33 |
@sonney2k | blackburn, what is a ppa? | 19:33 |
blackburn | personal package archive | 19:33 |
blackburn | ubuntu stuff | 19:33 |
blackburn | sonney2k: you add a ppa to your sources and it gets updated daily | 19:33 |
gsomix | heiko, thanks. .___. | 19:33 |
@sonney2k | blackburn, so just something one can put in apt/sources.lst | 19:33 |
@sonney2k | ? | 19:33 |
blackburn | yes | 19:33 |
@sonney2k | well we can have daily snapshots | 19:34 |
heiko | that would be a cool thing | 19:34 |
@sonney2k | (only if the buildbot is happy of course) | 19:35 |
@sonney2k | blackburn, which distributions are popular? is there a ranking somewhere? | 19:35 |
blackburn | yes | 19:36 |
blackburn | distrowatch | 19:36 |
@sonney2k | I would say debian, ubuntu, ... what else? | 19:36 |
blackburn | http://distrowatch.com/ | 19:36 |
blackburn | sonney2k: mint | 19:36 |
blackburn | fedora, suse | 19:36 |
blackburn | and probably arch | 19:36 |
heiko | mint is also debian based | 19:36 |
heiko | suse would be rpm based so probably a good idea | 19:36 |
heiko | but I think the most important thing would be mac | 19:37 |
heiko | isnt there some way? | 19:37 |
blackburn | we have no mac developer | 19:37 |
blackburn | :( | 19:37 |
blackburn | wait | 19:37 |
blackburn | wiking is | 19:37 |
blackburn | :D | 19:37 |
@sonney2k | wiking, yes | 19:37 |
heiko | oh, poor guy ;) | 19:37 |
@sonney2k | isn't mint just ubuntu? | 19:38 |
blackburn | sonney2k: yes but they rebased to debian IIRC | 19:38 |
blackburn | but still | 19:39 |
blackburn | nothing is changed in mint | 19:39 |
blackburn | it is more about codecs and other media stuff | 19:39 |
heiko | I think mint is only popular because its ubuntu-new gnome shell therefore its popular | 19:40 |
heiko | hoops double popular | 19:40 |
heiko | I read an article some time ago that many ubuntu users switched to mint since its not different but doesnt have this new side-bar thingi | 19:40 |
@sonney2k | yeah so we do ubuntu, debian, fedora, suse | 19:40 |
@sonney2k | mac and cygwin | 19:41 |
heiko | sounds good | 19:41 |
@sonney2k | that is quite a bit though | 19:41 |
@sonney2k | for each we need libshogun, static and modular interfaces | 19:42 |
@sonney2k | and then I still want to see the doc | 19:42 |
@sonney2k | ahh maybe we do it in this nightly thing on debian stable | 19:42 |
@sonney2k | nightly default/ nightly_none / nightly_all | 19:43 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 19:43 | |
@sonney2k | so nightly_default would create a .deb and update the doc | 19:44 |
@sonney2k | heiko, blackburn, n4nd0 btw did you ever get an email from the buildbot about failed builds? | 19:44 |
heiko | sonney2k, no | 19:44 |
n4nd0 | sonney2k: me neither | 19:44 |
blackburn | ?? | 19:45 |
blackburn | No | 19:45 |
@sonney2k | hmmhh only the irc messages then right? | 19:45 |
blackburn | yes | 19:45 |
@sonney2k | then I guess I still have a mistake in the config | 19:45 |
@sonney2k | it is *really complex* | 19:45 |
@sonney2k | I will have to investigate | 19:46 |
@sonney2k | heiko, blackburn - final thing: I do no longer like our website | 19:46 |
@sonney2k | I think we need something with tabs | 19:46 |
blackburn | hah | 19:46 |
heiko | agreed | 19:46 |
@sonney2k | like documentation | 19:46 |
@sonney2k | faq | 19:46 |
heiko | and not so messy :) | 19:47 |
@sonney2k | contact | 19:47 |
@sonney2k | whatever | 19:47 |
heiko | whats your plan about that? | 19:47 |
heiko | should we think of a nice design? | 19:48 |
heiko | make a draft? | 19:48 |
-!- gsomix [~gsomix@95.67.191.81] has quit [Remote host closed the connection] | 19:49 | |
@sonney2k | heiko, I would love to have some nice design from some free template | 19:49 |
@sonney2k | http://www.nuviotemplates.com/ | 19:49 |
@sonney2k | like these | 19:49 |
heiko | yeah these look nice, you took them for your page right? | 19:50 |
@sonney2k | the first one yes | 19:51 |
@sonney2k | another one for mldata.org | 19:51 |
@sonney2k | that one is also OK: http://www.freetemplatesonline.com/templates/Free-Software-Templates-Free-Internet-Templates-2.html | 19:52 |
heiko | yeah, why not, not too much difference | 19:52 |
heiko | I think the main thing is structure | 19:53 |
heiko | and usability | 19:53 |
@sonney2k | this one is cool http://www.nuviotemplates.com/preview.php?template=gr8-01 :) | 19:54 |
heiko | yes thats a good thing with the images :) | 19:54 |
@sonney2k | I really like it a lot | 19:55 |
@sonney2k | blackburn, n4nd0 any comments? | 19:55 |
@sonney2k | I wish pluskid was here - he is certainly deep into graphics | 19:56 |
n4nd0 | I think it is a good idea | 19:56 |
heiko | I always like minimalistic approaches, but I also like that one | 19:56 |
@sonney2k | heiko, the good news is that we have enough resources on the server to dynamically generate the content | 19:57 |
heiko | thats indeed good news :) | 19:57 |
@sonney2k | so I can give you editing capabilities | 19:58 |
@sonney2k | rights I mean | 19:58 |
@sonney2k | I only need to program the website :D | 19:58 |
@sonney2k | django/python that is | 19:58 |
heiko | I read about that in your blog some time ago, sounds nice | 19:58 |
heiko | and its good that more than one person can update stuff | 19:58 |
n4nd0 | we could add tutorials like these! http://docs.opencv.org/doc/tutorials/ml/table_of_content_ml/table_of_content_ml.html | 20:00 |
heiko | n4nd0 hey i didnt know you worked for openCV before :) | 20:02 |
heiko | nice! | 20:02 |
n4nd0 | well "worked" haha | 20:02 |
n4nd0 | I have used opencv a bit and I just wrote those tutorials ;) | 20:03 |
@sonney2k | lots of work... | 20:03 |
@sonney2k | blackburn, have you read logs about documentation? | 20:03 |
@sonney2k | what's the state about your doc project? | 20:03 |
@sonney2k | I think it is about time to create it now such that n4nd0 and pluskid can put their stuff in | 20:04 |
n4nd0 | sonney2k: when would you like to have the new page working? I know someone who can be interested in helping with it or something, if you like the idea of another working on it of course | 20:06 |
@sonney2k | n4nd0, I don't have a schedule or anything | 20:07 |
n4nd0 | sonney2k: and do you like the idea of someone working on it? | 20:08 |
@sonney2k | yes but depends on the person of course :) | 20:10 |
n4nd0 | ok ;) | 20:11 |
@sonney2k | n4nd0, what can he/she do? | 20:12 |
@sonney2k | I think I would just take the template and create some dummy page around it and put it somewhere | 20:12 |
n4nd0 | he has some experience with django | 20:13 |
n4nd0 | and not much to do during July :) | 20:13 |
@sonney2k | sounds good | 20:13 |
n4nd0 | yeah I have just told him that | 20:13 |
@sonney2k | n4nd0, well I have the source code for my homepage somewhere in svn or git or so | 20:13 |
@sonney2k | and it should be easy to modify for that purpose | 20:14 |
n4nd0 | if he's interested, he'll prepare something | 20:14 |
@sonney2k | n4nd0, shouldn't we discuss first? | 20:16 |
@sonney2k | or what do you mean with prepare? | 20:16 |
n4nd0 | sonney2k: I was just thinking about changing a bit a template of the one you pasted the links of | 20:16 |
@sonney2k | n4nd0, you mean to make it look shogun like? | 20:17 |
@sonney2k | *ARGH* | 20:17 |
n4nd0 | yeah | 20:17 |
@sonney2k | we have a u'\ufeff' unicode character somewhere in the source code | 20:17 |
@sonney2k | breaking the build | 20:17 |
@sonney2k | how can one search for it?! | 20:18 |
n4nd0 | where have you seen that? | 20:18 |
n4nd0 | I just see failed test in the buildbot | 20:18 |
n4nd0 | some failed configure too | 20:19 |
@sonney2k | python ../..//.doxy2swig.py --quiet --no-function-definition \ | 20:19 |
@sonney2k | modshogun/doxygen_xml/index.xml modshogun_doxygen.i | 20:19 |
@sonney2k | Traceback (most recent call last): | 20:19 |
@sonney2k | File "../..//.doxy2swig.py", line 450, in <module> | 20:19 |
@sonney2k | main() | 20:19 |
@sonney2k | File "../..//.doxy2swig.py", line 446, in main | 20:19 |
@sonney2k | convert(args[0], args[1], not options.func_def, options.quiet) | 20:19 |
@sonney2k | File "../..//.doxy2swig.py", line 426, in convert | 20:19 |
@sonney2k | p.write(output) | 20:19 |
@sonney2k | File "../..//.doxy2swig.py", line 382, in write | 20:19 |
@sonney2k | o.write("".join(self.pieces)) | 20:19 |
@sonney2k | UnicodeEncodeError: 'ascii' codec can't encode character u'\ufeff' in position 422884: ordinal not in range(128) | 20:19 |
@sonney2k | n4nd0, I am building locally | 20:19 |
blackburn | sonney2k: yeah i have read, time to start | 20:19 |
@sonney2k | does anyone know how to create this unicode character to actually grep the file? | 20:20 |
blackburn | hah no | 20:21 |
@sonney2k | it is a zero space character | 20:22 |
@sonney2k | so not even visible - wtf!? | 20:22 |
@sonney2k | n4nd0, how long will it take your friend to come up with something? | 20:22 |
@sonney2k | or in other words when do we know he is interested :D | 20:23 |
n4nd0 | about a week, he has told me | 20:25 |
@sonney2k | ok so lets say next wednesday - then I will try to get the build bots in shape first | 20:27 |
n4nd0 | cool | 20:28 |
@sonney2k | blackburn, these build errors are on you - http://shogun-toolbox.org/buildbot/builders/cmdline_static/builds/996/steps/test_1/logs/stdio | 21:09 |
@sonney2k | please fix the examples | 21:09 |
blackburn | meaow | 21:30 |
CIA-18 | shogun: Sergey Lisitsyn master * r3eb7bcd / (3 files): Got examples back to work - http://git.io/UeYsyw | 21:38 |
@sonney2k | thx | 21:51 |
@sonney2k | heiko, you did it! | 21:58 |
@sonney2k | you inserted an invisible utf8 code | 21:58 |
@sonney2k | grep shogun -r | 21:58 |
@sonney2k | shogun/statistics/QuadraticTimeMMD.h: * Gretton, A., Borgwardt, K. M., Rasch, M. J., Sch?lkopf, B., & Smola, A. (2012). | 21:58 |
@sonney2k | shogun/statistics/LinearTimeMMD.h: * Gretton, A., Borgwardt, K. M., Rasch, M. J., Sch?lkopf, B., & Smola, A. (2012). | 21:58 |
alexlovesdata_ | @sonney2k: LOL | 21:59 |
@sonney2k | it is a zero width space | 21:59 |
@sonney2k | invisible on the console but causing errors | 21:59 |
@sonney2k | great work! | 21:59 |
CIA-18 | shogun: Soeren Sonnenburg master * rb501667 / (2 files): remove zero width space causing errors - http://git.io/Uf1maQ | 22:00 |
@sonney2k | so the patch looks like this :) | 22:01 |
@sonney2k | diff --git a/src/shogun/statistics/LinearTimeMMD.h b/src/shogun/statistics/LinearTimeMMD.h | 22:01 |
@sonney2k | index fe5cdb8..9714c60 100644 | 22:01 |
@sonney2k | --- a/src/shogun/statistics/LinearTimeMMD.h | 22:01 |
@sonney2k | +++ b/src/shogun/statistics/LinearTimeMMD.h | 22:01 |
@sonney2k | @@ -20,7 +20,7 @@ class CFeatures; | 22:01 |
@sonney2k | 22:01 | |
@sonney2k | /** @brief This class implements the linear time Maximum Mean Statistic as | 22:01 |
@sonney2k | * described in | 22:01 |
@sonney2k | - * Gretton, A., Borgwardt, K. M., Rasch, M. J., Sch?lkopf, B., & Smola, A. (2012). | 22:01 |
@sonney2k | + * Gretton, A., Borgwardt, K. M., Rasch, M. J., Schölkopf, B., & Smola, A. (2012). | 22:01 |
@sonney2k | * A Kernel Two-Sample Test. Journal of Machine Learning Research, 13, 671-721. | 22:01 |
@sonney2k | * | 22:01 |
@sonney2k | * Along with the statistic comes a method to compute a p-value based on a | 22:01 |
@sonney2k | crazy world | 22:01 |
n4nd0 | :D | 22:01 |
heiko | sonney2k, wow | 22:04 |
heiko | How did I do that? | 22:04 |
@sonney2k | heiko, copy paste from some website? | 22:04 |
@sonney2k | totally impossible to detect | 22:04 |
heiko | but I told eclipse to use utf8 | 22:04 |
heiko | man | 22:04 |
heiko | sorry for that :( | 22:04 |
@sonney2k | cause this space is 0 length | 22:04 |
heiko | just saw all the stress you had with that | 22:04 |
@sonney2k | so absolutely invisible | 22:04 |
@sonney2k | well use ascii encoding only | 22:05 |
heiko | yeah Ill change that | 22:05 |
@sonney2k | heiko, blackburn there is one example related to subsets from sparse features still failing: http://shogun-toolbox.org/buildbot/builders/cmdline_static/builds/997/steps/test_1/logs/stdio | 22:06 |
heiko | lol, oh guys, sorry again, but somehow funny these errors ;) | 22:07 |
@sonney2k | then libshogun is back to normal | 22:07 |
alexlovesdata_ | invisible utf8 having zero length ... this should enter a blog! | 22:07 |
@sonney2k | yeah lots of fun | 22:07 |
heiko | blackburn, its a segfault? Any idea? | 22:08 |
alexlovesdata_ | once you solved it is is fun :| | 22:08 |
heiko | happens when the transpose is checked, did you change anything there? | 22:08 |
heiko | I didnt | 22:08 |
blackburn | yeah but let me fix it later | 22:09 |
blackburn | we changed a lot | 22:09 |
blackburn | I changed | 22:09 |
-!- blackburn [~blackburn@31.28.43.76] has quit [Ping timeout: 240 seconds] | 22:36 | |
CIA-18 | shogun: Heiko Strathmann master * r8023b21 / (5 files in 2 dirs): -fixed a problem where combinations of a model selection parameter tree - http://git.io/IdKm7w | 22:59 |
CIA-18 | shogun: Heiko Strathmann master * re53fe46 / src/shogun/lib/Set.h : corrected a wrongly named field - http://git.io/Fdy-Eg | 22:59 |
CIA-18 | shogun: Heiko Strathmann master * r062bc5f / (6 files in 3 dirs): Merge pull request #604 from karlnapf/master - http://git.io/MsyCeA | 22:59 |
-!- heiko [~heiko@host86-180-159-222.range86-180.btcentralplus.com] has quit [Ping timeout: 264 seconds] | 23:11 | |
-!- alexlovesdata_ [82955843@gateway/web/freenode/ip.130.149.88.67] has left #shogun [] | 23:33 | |
--- Log closed Wed Jun 27 00:00:17 2012 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!