--- Log opened Fri Aug 24 00:00:17 2012 | ||
-!- zxtx [~zv@cpe-75-83-151-252.socal.res.rr.com] has quit [Read error: Operation timed out] | 03:48 | |
shogun-buildbot_ | build #75 of nightly_default is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_default/builds/75 | 03:54 |
---|---|---|
-!- av3ngr [av3ngr@nat/redhat/x-unomdjhlngrsqlva] has joined #shogun | 06:45 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 06:52 | |
-!- zxtx [~zv@guest-wireless-nup-198-147-120-231.usc.edu] has joined #shogun | 07:16 | |
-!- av3ngr [av3ngr@nat/redhat/x-unomdjhlngrsqlva] has quit [Ping timeout: 246 seconds] | 08:12 | |
-!- av3ngr [av3ngr@nat/redhat/session] has joined #shogun | 08:24 | |
-!- av3ngr [av3ngr@nat/redhat/session] has quit [Ping timeout: 248 seconds] | 08:55 | |
-!- uricamic [~uricamic@2001:718:2:1634:e9e5:5ab4:359d:813d] has joined #shogun | 08:56 | |
-!- av3ngr [av3ngr@nat/redhat/x-canvtavaulipugsg] has joined #shogun | 09:14 | |
-!- zxtx [~zv@guest-wireless-nup-198-147-120-231.usc.edu] has quit [Ping timeout: 264 seconds] | 09:16 | |
-!- av3ngr [av3ngr@nat/redhat/x-canvtavaulipugsg] has left #shogun ["Leaving"] | 09:35 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 10:09 | |
-!- zxtx [~zv@cpe-75-83-151-252.socal.res.rr.com] has joined #shogun | 10:33 | |
-!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has joined #shogun | 10:37 | |
-!- emrecelikten [~emre@trir-4d0d9a53.pool.mediaWays.net] has joined #shogun | 10:59 | |
-!- yoo_ [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has joined #shogun | 11:17 | |
-!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has quit [Quit: Page closed] | 11:29 | |
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.83.179.44.135] has joined #shogun | 13:15 | |
-!- yoo_ [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has quit [Quit: Page closed] | 13:19 | |
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.83.179.44.135] has quit [Quit: Page closed] | 13:52 | |
-!- nicococo [~nico@guest0.ml.tu-berlin.de] has joined #shogun | 13:57 | |
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.83.179.44.135] has joined #shogun | 14:00 | |
-!- blackburn [~blackburn@109.226.79.69] has joined #shogun | 14:21 | |
blackburn | HELLO | 14:24 |
-!- yoh_ [~yoh@washoe.dartmouth.edu] has joined #shogun | 14:37 | |
blackburn | n4nd0: around? | 14:40 |
n4nd0 | blackburn: yeah | 14:41 |
n4nd0 | blackburn: at skype with Nico :) | 14:41 |
blackburn | oh any news on SO improvements? | 14:41 |
blackburn | does he have any cool ideas? | 14:41 |
-!- Netsplit *.net <-> *.split quits: yoh | 14:42 | |
n4nd0 | blackburn: I have been asking here about the bundle methods | 14:46 |
n4nd0 | blackburn: I have the ideas more clear now :) | 14:46 |
blackburn | what about subtracting <w,Psi(x,y_truth)> in argmax? | 14:47 |
n4nd0 | yeah we talked about that | 14:47 |
n4nd0 | he said that the score in the ResultSet (the return value of argmax) can be defined as we like | 14:48 |
n4nd0 | so yes, we can make that the score is | 14:48 |
n4nd0 | score = <w,Psi(xi,y)> + Delta(yi, y) - <w,Psi(xi,yi)> | 14:48 |
n4nd0 | that way in the generic risk we'll just do | 14:49 |
n4nd0 | call to argmax | 14:49 |
n4nd0 | and | 14:49 |
n4nd0 | risk += ret->score | 14:49 |
n4nd0 | blackburn: do you like it like that? | 14:49 |
blackburn | n4nd0: yes that's ok | 14:49 |
n4nd0 | cool | 14:50 |
blackburn | n4nd0: but we need to not forget about subtracting loss from multiclass model | 14:50 |
n4nd0 | blackburn: yeah, sure | 14:50 |
blackburn | however this is kind of inefficient it is more general this way | 14:50 |
blackburn | n4nd0: what about HM SVM? | 14:50 |
blackburn | is it correct now with provided risk function? | 14:51 |
n4nd0 | yes | 14:51 |
n4nd0 | the only thing is that now (in the PR I submitted) the loss is taken into account twice | 14:51 |
n4nd0 | but I'll fix that within minutes now | 14:51 |
n4nd0 | nicococo: anything to add? :) | 14:53 |
-!- n4nd0 [53b32c87@gateway/web/freenode/ip.83.179.44.135] has quit [Quit: Page closed] | 14:54 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 14:54 | |
n4nd0 | blackburn: what do you think? | 14:55 |
-!- nicococo [~nico@guest0.ml.tu-berlin.de] has quit [Ping timeout: 265 seconds] | 14:55 | |
n4nd0 | does it sound good? | 14:55 |
blackburn | n4nd0: about what? | 14:56 |
blackburn | :) | 14:56 |
n4nd0 | about having this working well finally | 14:57 |
n4nd0 | supported by Nico, it gives more confidence that it is ok | 14:57 |
blackburn | well yeah I don't mind to make it work :) | 14:58 |
blackburn | n4nd0: can you please also document in argmax description | 14:59 |
blackburn | that in training mode it computes blablah | 14:59 |
blackburn | with blahblah assumption | 14:59 |
blackburn | :D | 14:59 |
n4nd0 | ok | 14:59 |
blackburn | I will add director SO model | 15:00 |
blackburn | once we have everything working | 15:00 |
blackburn | this would be really kind of killer feature I hope | 15:01 |
n4nd0 | blackburn: yeah the SO director is very interesting | 15:01 |
n4nd0 | I told gsomix about it | 15:01 |
n4nd0 | blackburn: SO model and StateModel is very interesting too! | 15:02 |
blackburn | well it is pretty easy to set up, not really needed any gsomix's expertise :) | 15:02 |
blackburn | director StateModel? | 15:02 |
n4nd0 | I would welcome StateModel very very much | 15:02 |
n4nd0 | yeah | 15:02 |
blackburn | it is HM right? | 15:02 |
n4nd0 | you know when you use the HM-SVM you need another piece | 15:02 |
blackburn | so i.e. if you have 3 states | 15:03 |
blackburn | you would need that | 15:03 |
n4nd0 | yes | 15:03 |
blackburn | right? | 15:03 |
n4nd0 | apart from the number of states, the type of model too | 15:03 |
blackburn | okay lets implement that too | 15:03 |
blackburn | ah llike semi markovian? | 15:03 |
n4nd0 | e.g. in speech recognition people used right Markov models | 15:03 |
n4nd0 | the transitions to the left (think of a Markov chain of states) are not allowed | 15:04 |
blackburn | I think about starting PGM course *seriously* this time | 15:04 |
n4nd0 | yeah me too | 15:04 |
n4nd0 | end of September | 15:04 |
n4nd0 | :D | 15:04 |
blackburn | n4nd0: well I don't care about that competitive stuff | 15:04 |
blackburn | lections are here anyway | 15:04 |
blackburn | err lectures | 15:04 |
n4nd0 | true | 15:04 |
blackburn | and you always can submit code | 15:04 |
n4nd0 | we should try to take to it and motivate each other not to drop it | 15:05 |
blackburn | motivation should come from research interest | 15:05 |
n4nd0 | that one seems so far away... | 15:06 |
blackburn | what do you mean? | 15:06 |
n4nd0 | I see research something sooo far | 15:08 |
blackburn | no it is already here in front of you | 15:09 |
blackburn | staring into your eyes :) | 15:09 |
n4nd0 | haha we shall see | 15:09 |
n4nd0 | the compilation time in my computer kills me | 15:11 |
n4nd0 | it makes me lose time for real | 15:11 |
blackburn | --disable-optimization, multiple threads? | 15:11 |
blackburn | ccache? | 15:11 |
blackburn | are you using that | 15:12 |
blackburn | or are you talking about interfaces? | 15:12 |
n4nd0 | I use ccache make -j2 normally | 15:12 |
n4nd0 | disable-optimization on | 15:12 |
n4nd0 | normally just with python-modular support | 15:12 |
blackburn | well what can I suggest is to use libshogun only | 15:13 |
blackburn | or lets try to find a way to split interface file | 15:13 |
blackburn | n4nd0: hah were you here already then we disabled optimizations in interfaces? | 15:13 |
blackburn | before it took 2-3 times more | 15:14 |
n4nd0 | I don't think so | 15:14 |
n4nd0 | really? | 15:14 |
blackburn | we are now using -O0 in interfaces | 15:14 |
blackburn | but before it was global | 15:14 |
blackburn | -O3 or -O0 | 15:14 |
blackburn | for everything | 15:14 |
n4nd0 | can imagine | 15:15 |
blackburn | n4nd0: but in general I agree we really need to improve dev experience | 15:16 |
blackburn | n4nd0: what mostly slows down compilation is template classes | 15:20 |
n4nd0 | I see | 15:21 |
blackburn | n4nd0: one of reasons I felt a little pain in my bottom when you introduced matrix features :D | 15:21 |
n4nd0 | it could be so cool if it is just re-compiled what is changed | 15:21 |
n4nd0 | when one changes a header file, all over again | 15:21 |
n4nd0 | haha | 15:22 |
n4nd0 | they were required I think, didn't find anything that could do the job | 15:22 |
n4nd0 | though I looked for it! | 15:22 |
blackburn | yes, dense wouldn't work | 15:22 |
blackburn | okay dense subset is not really needed we could merge it to dense | 15:23 |
blackburn | argh no pluskid around | 15:24 |
blackburn | okay I have to add a unit test then | 15:25 |
n4nd0 | blackburn: PR updated | 15:45 |
blackburn | okay let me check | 15:47 |
CIA-52 | shogun: Sergey Lisitsyn master * r093b502 / (7 files): Merge pull request #756 from iglesias/master (+6 more commits...) - http://git.io/_rSCmA | 16:02 |
CIA-52 | shogun: Chiyuan Zhang master * rba02284 / src/shogun/multiclass/MulticlassOneVsOneStrategy.cpp : tie-breaking by decision values in One-vs-One. - http://git.io/uEjFmg | 16:10 |
CIA-52 | shogun: Sergey Lisitsyn master * r2f3688a / src/shogun/multiclass/MulticlassOneVsOneStrategy.cpp : Merge pull request #753 from pluskid/multiclass - http://git.io/kxI4lg | 16:10 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 16:20 | |
-!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has joined #shogun | 16:28 | |
yoo | hi all | 16:32 |
blackburn | hi | 16:33 |
-!- alexlovesdata [~binder@goldenezahl.ml.tu-berlin.de] has joined #shogun | 17:14 | |
-!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has quit [Quit: Page closed] | 17:41 | |
-!- uricamic [~uricamic@2001:718:2:1634:e9e5:5ab4:359d:813d] has quit [Quit: Leaving.] | 17:44 | |
CIA-52 | shogun: Sergey Lisitsyn master * rd50e297 / src/interfaces/modular/modshogun_ignores.i : Ignored SGVector operator+ and operator+= - http://git.io/t-TRzA | 17:45 |
@sonney2k | yoh_, I think you can safely assume that debian wheezy will have shogun from squeeze | 18:03 |
-!- gsomix [~gsomix@178.45.65.87] has joined #shogun | 18:05 | |
yoh_ | sonney2k: 0.9.3? how ? atm there is no single version in wheezy and only viable one to enter is in unstable/sid which is 1.1.0... or do you mean that wheezy will not have any version? ;) | 18:11 |
yoh_ | btw -- probably for my own sake I will just build backports of 1.1.0 and distribute from neuro.debian.net | 18:12 |
blackburn | yoh_: why not 2.0? | 18:12 |
yoh_ | blackburn: is it released? well tested? is there a debian package? ;) | 18:12 |
blackburn | yoh_: is that done for 1.1.0? :D | 18:12 |
blackburn | 2.0 will be released on 1st of sept | 18:13 |
yoh_ | blackburn: 1.1.0 was out there for a while so hopefully it is a bit more tested ;) | 18:13 |
blackburn | yoh_: no it is not | 18:13 |
blackburn | I am sure 2.0 have less bugs than 1.1.0 | 18:14 |
yoh_ | blackburn: it would be great! ;) so, ok -- is 1.1.0 worth hassle over 0.9.3? | 18:15 |
blackburn | yes, sure | 18:15 |
blackburn | but 2.0 is worth even more :) | 18:15 |
blackburn | not feature-wise but bug-wise | 18:16 |
blackburn | there were a few bugs that has to be merged 1.1.0 if it was supported | 18:16 |
blackburn | to 1.1.0* | 18:16 |
blackburn | yoh_: 0.9.3 has no java, no ruby.. etc | 18:19 |
blackburn | yoh_: two major (I do mean major) new features are director swig stuff and structured output framework | 18:20 |
yoh_ | good ;-) FWIW: myself I care about python ;) | 18:23 |
blackburn | yoh_: then directors are for you probably | 18:24 |
blackburn | and zero-copy features | 18:24 |
blackburn | yoh_: directors allow you to take care of e.g. kernels computing using python code | 18:25 |
blackburn | or to define structured model via python code | 18:25 |
blackburn | (in case you didn't know what it is) :) | 18:26 |
yoh_ | so, 1st of sept there also would be debian packages uploaded? ;) | 18:31 |
blackburn | yoh_: I am just a stupid developer, let you rather ask soeren about that :) | 18:32 |
yoh_ | sonney2k: btw -- from the debian build log it is not clear -- are you running unit/regresion tests at pkg build time? it might be cool if you did ;-) | 18:34 |
blackburn | to be serious I am unsure about that - sonney2k takes care about that | 18:34 |
blackburn | yoh_: can I see these logs? | 18:34 |
yoh_ | blackburn: no way -- they are a property of the communist part of C.! | 18:35 |
yoh_ | nah... of D.! | 18:35 |
yoh_ | ;) https://buildd.debian.org/status/fetch.php?pkg=shogun&arch=armhf&ver=1.1.0-6&stamp=1345539853 | 18:35 |
yoh_ | more from here: https://buildd.debian.org/status/package.php?p=shogun | 18:35 |
blackburn | oh my it is pretty big | 18:36 |
blackburn | super lu is absent | 18:37 |
blackburn | strange, okay not a problem | 18:37 |
CIA-52 | shogun: Sergey Lisitsyn master * rcbc0150 / src/shogun/structure/StructuredModel.cpp : Changed type of loop variable in StructuredModel - http://git.io/JZa_1g | 20:07 |
@sonney2k | yoh_, of course we don't run any tests... otherwise the build will take >1d on arm etc | 20:22 |
blackburn | sonney2k: are you willing to have modsel typemaps in 2.0? | 20:22 |
blackburn | we have to freeze at some point | 20:22 |
blackburn | sonney2k: has TU bb gone? | 20:27 |
blackburn | cygwin one | 20:27 |
@sonney2k | blackburn, yeah | 20:28 |
@sonney2k | blackburn, extra typemaps for these kind would be nice but if they dont' make it we put them in 2.1 right? | 20:29 |
@sonney2k | or 2.0.1 | 20:29 |
blackburn | exactly | 20:29 |
blackburn | but 2.0 should be more or less stable | 20:29 |
@sonney2k | blackburn, however these are totally non-intrusive and won't hurt if not yet ready | 20:29 |
@sonney2k | sure | 20:29 |
@sonney2k | as stable as possible | 20:29 |
blackburn | I have time now but a little lost what to do | 20:29 |
blackburn | now I am finishing multiclass output requested by yoo a few times | 20:30 |
@sonney2k | blackburn, update NEWS? try to get tests to work perfectly (even old ones?) | 20:30 |
blackburn | oh NEWS | 20:30 |
blackburn | I love to update news :D | 20:30 |
blackburn | sonney2k: however warnings are at soft zero | 20:31 |
blackburn | that's actually nice | 20:31 |
blackburn | ahh I need to implement director structured model | 20:32 |
blackburn | very important thing to have | 20:32 |
@sonney2k | yeah but could as well be in 2.0.1 | 20:32 |
@sonney2k | tests / doc is probably most important now | 20:32 |
blackburn | no, disagree here | 20:32 |
blackburn | sounds very important | 20:32 |
blackburn | however | 20:32 |
@sonney2k | blackburn, very few people will use directors! | 20:32 |
blackburn | may be I am wrong | 20:32 |
blackburn | sonney2k: BMRM + python code sounds vvverrry interesting | 20:33 |
blackburn | pure prototyping | 20:33 |
@sonney2k | ahh and we have to write some blog entry with nice figures ... | 20:33 |
@sonney2k | though I didn't receive any further ones... | 20:34 |
blackburn | lol | 20:34 |
blackburn | well I really thought about that | 20:34 |
blackburn | but still so hard to came with any idea for me | 20:34 |
blackburn | :( | 20:34 |
@sonney2k | blackburn, well multiclass is easy ... just illustrate e.g. 3-5 gaussians and decision surfaces for all the MC methods | 20:34 |
@sonney2k | multitask is more difficult | 20:34 |
blackburn | hmm multitask could be | 20:35 |
@sonney2k | but one could create one toy example showing the solution for the 'vanilla' svm and one for a dual task one | 20:35 |
blackburn | like | 20:35 |
@sonney2k | so show how solutions changes... | 20:35 |
blackburn | pictures coming from 3 sources | 20:35 |
blackburn | I'll try something | 20:35 |
blackburn | dimension reduction is the easiest here | 20:36 |
@sonney2k | for SO / latent variable SO we really need some fancy example like license plate recognition, speech recognition, background detection (in images)... | 20:36 |
@sonney2k | but that is lots of work | 20:36 |
blackburn | sonney2k: good example, sound like Phd | 20:36 |
blackburn | :D | 20:36 |
@sonney2k | blackburn, no | 20:36 |
@sonney2k | not when you have all the methods available already | 20:37 |
@sonney2k | blackburn, it is of same effort as your traffic sign detector | 20:37 |
blackburn | sonney2k: I still didn't get how could that even work (background detection with SO) | 20:37 |
blackburn | sonney2k: I totally failed with L1/Lq btw :( | 20:37 |
blackburn | I thought it is worth to regularize cells with L2 | 20:37 |
@sonney2k | blackburn, it is a segmentation problem you have to decide for each pixel if it is background or foreground | 20:38 |
blackburn | sonney2k: yes, I understand | 20:38 |
blackburn | but how technically not yet | 20:38 |
@sonney2k | however doing so independently (naiive way) misses correlations... so SO can help here | 20:38 |
@sonney2k | blackburn, I am not an image guy so better ask wiking - he did things like this IIRC | 20:39 |
blackburn | sonney2k: can that be done with HM having #pixels states? | 20:39 |
@sonney2k | blackburn, ahh for GPs we have nice figures right? | 20:39 |
@sonney2k | blackburn, between 2 ... #pixels states :D | 20:40 |
@sonney2k | #pixels sounds a bit stupid | 20:40 |
blackburn | why? | 20:40 |
blackburn | argh | 20:40 |
blackburn | so hidden states are | 20:40 |
@sonney2k | would then be like SVM | 20:40 |
blackburn | back and fore? | 20:40 |
@sonney2k | yes | 20:40 |
blackburn | what are observations? | 20:41 |
blackburn | pixels? | 20:41 |
@sonney2k | but maybe some more | 20:41 |
blackburn | intensities? | 20:41 |
blackburn | or some fancy stuff like hogs would work too? | 20:41 |
@sonney2k | some features computed based on pixels / neighborhood | 20:41 |
@sonney2k | I would guess so | 20:41 |
blackburn | lbp etc? | 20:41 |
blackburn | I see | 20:41 |
@sonney2k | yeah | 20:41 |
blackburn | well | 20:41 |
blackburn | I could develop an example then | 20:41 |
blackburn | if I understand correctly | 20:42 |
blackburn | we have everything | 20:42 |
blackburn | to do that | 20:42 |
@sonney2k | well one thing maybe not | 20:42 |
blackburn | which one? | 20:42 |
@sonney2k | I only understand how things work for 1d images | 20:42 |
blackburn | :D | 20:42 |
@sonney2k | so you would go through an image from left to right | 20:42 |
blackburn | 1d images? | 20:42 |
blackburn | lol | 20:42 |
@sonney2k | (one line only :) | 20:42 |
@sonney2k | and get a state sequence F-B-B-B-B-F-F ... etc | 20:43 |
blackburn | why not to flatten it row by row then? | 20:43 |
@sonney2k | blackburn, the dependencies matter | 20:43 |
blackburn | ahh wait wait | 20:44 |
blackburn | so we observe only neighborhood | 20:44 |
blackburn | of each pixel | 20:44 |
blackburn | not the whole image | 20:44 |
blackburn | right? | 20:44 |
@sonney2k | yes | 20:44 |
blackburn | I start to understand these pictures I have seen with CRFs segmentation | 20:44 |
blackburn | there they had 9 observations and some network of it | 20:44 |
@sonney2k | but in SO you have the state model (simplest 2 states F & B) | 20:44 |
@sonney2k | and you learn some weights going from F -> B and back | 20:44 |
@sonney2k | so this somehow needs to be taken into account | 20:45 |
@sonney2k | yeah it is more like a network | 20:45 |
@sonney2k | (e.g. markov network) | 20:45 |
blackburn | okay I see | 20:46 |
blackburn | you made me interested I should try | 20:46 |
@sonney2k | blackburn, in this sense speech recognition is much simpler | 20:46 |
blackburn | this is the application I do not understand even more | 20:47 |
blackburn | in means of features | 20:47 |
blackburn | how do they construct features? | 20:47 |
@sonney2k | you just have a number of states (phonems) and combine these phonems into most likely words | 20:47 |
blackburn | how do they detect phonems? | 20:47 |
@sonney2k | blackburn, well a sample of a phonem | 20:47 |
blackburn | is it a solved problem? | 20:47 |
@sonney2k | blackburn, yes | 20:48 |
blackburn | okay I see | 20:48 |
@sonney2k | well basically you have some known phonems in training | 20:48 |
@sonney2k | and then you can fit gaussians or so on them | 20:48 |
@sonney2k | (HMMs with guassians are the standard approach for that task) | 20:48 |
blackburn | sonney2k: can I fit segmentation HM with one image? | 20:49 |
blackburn | would that work at all? | 20:49 |
blackburn | not in generalization means but at least would it segment the image I used? | 20:49 |
@sonney2k | well if you used say 30% of the image for training then maybe | 20:50 |
@sonney2k | blackburn, the methods here are all supervised | 20:50 |
blackburn | in other words | 20:51 |
@sonney2k | so not like a HMM that could (in theory) also segment the image in one go | 20:51 |
blackburn | would it reproduce the same pattern? | 20:51 |
blackburn | at least approximately | 20:51 |
@sonney2k | blackburn, yes | 20:51 |
blackburn | sounds like pretty interesting thing | 20:52 |
@sonney2k | depending on #states it will reproduce the exact image :D | 20:52 |
blackburn | I have to do that because I hardly believe :D | 20:52 |
@sonney2k | let me better work on the website... | 20:54 |
blackburn | instead of? | 20:54 |
blackburn | sonney2k: do you have any (just any) idea what could make our HMM class fail? | 20:56 |
blackburn | I assume it was working before? | 20:57 |
@sonney2k | blackburn, does it fail now? | 20:58 |
blackburn | test - yes | 20:58 |
@sonney2k | ahh test... ohh well it is difficult. randomly initialized parameters NP hard problem... | 20:59 |
@sonney2k | I would need to check... | 20:59 |
@sonney2k | if the 3 dice problem still works then it is good though | 20:59 |
blackburn | I see | 21:00 |
@sonney2k | (3 dice with different distributions are drawn lots of times - a 3 state HMM can learn the distribution of the dice figures and when they were drawn) | 21:00 |
yoh_ | sonney2k: well -- it would be neat if tests were run... at least selectively for some architectures... and yeah -- it does take a while to build some of my packages on ARM because of that... ARM people should make faster chips/ports ;) | 21:06 |
yoh_ | imho it is better to make machines sweat than later on troubleshoot bug reports (although it is indeed unlikely to see shogun users on eg s390 ;) ) ;) | 21:08 |
blackburn | We have processed the evaluation for your project named Implement multitask and domain adaptation algorithms with Shogun Machine Learning Toolbox. | 21:09 |
@sonney2k | blackburn, and failed as usual ;-) | 21:11 |
@sonney2k | I have 8 emails of this kind | 21:12 |
blackburn | sonney2k: how do I manage to fail it year by year? | 21:20 |
blackburn | I am tired of it already :( | 21:20 |
@sonney2k | at least now the Rubel rolls right :D | 21:20 |
@sonney2k | blackburn, http://94.23.237.10:5000/ | 21:47 |
* sonney2k Zzzz... | 22:00 | |
blackburn | sonney2k: doesn't work for me | 22:06 |
-!- AndChat49209 [~AndChat49@host86-178-156-19.range86-178.btcentralplus.com] has joined #shogun | 22:33 | |
AndChat49209 | . | 22:33 |
AndChat49209 | . | 22:34 |
-!- AndChat49209 [~AndChat49@host86-178-156-19.range86-178.btcentralplus.com] has quit [Client Quit] | 22:34 | |
-!- AndChat49209 [~AndChat49@host86-178-156-19.range86-178.btcentralplus.com] has joined #shogun | 22:35 | |
blackburn | . :) | 22:35 |
AndChat49209 | Its heiko with mobile but dont know how to change name | 22:36 |
blackburn | hello mobile heiko | 22:36 |
blackburn | how are you and your thesis? | 22:36 |
-!- heiko1 [~heiko@host86-178-156-19.range86-178.btcentralplus.com] has joined #shogun | 22:37 | |
heiko1 | hey | 22:37 |
blackburn | hey | 22:37 |
heiko1 | dude this is horros on the mobile keyboard :) | 22:37 |
heiko1 | thesis is coming along, has been a prett yintense week | 22:37 |
blackburn | I imagine | 22:38 |
heiko1 | as always: its more work than I thought :) | 22:38 |
blackburn | haha | 22:38 |
blackburn | yeah | 22:38 |
heiko1 | but still 2 weeks left | 22:38 |
heiko1 | so this will be ok | 22:38 |
blackburn | I thought I will do my bachelor thesis in a week LOL | 22:38 |
heiko1 | thats optimistic :) | 22:39 |
blackburn | very | 22:39 |
heiko1 | but I have talked to people who havent start writing things down yet | 22:39 |
-!- AndChat49209 [~AndChat49@host86-178-156-19.range86-178.btcentralplus.com] has quit [Client Quit] | 22:39 | |
blackburn | I knew quite a few guys who acted this way | 22:39 |
blackburn | they are in army now | 22:39 |
blackburn | :D | 22:39 |
heiko1 | lol :) | 22:39 |
blackburn | won't work for you though | 22:39 |
heiko1 | no too old already | 22:40 |
heiko1 | (and did civil service) | 22:40 |
blackburn | I have 6 more years of hiding | 22:40 |
blackburn | :D | 22:40 |
heiko1 | oh man | 22:40 |
heiko1 | come to Germany :) | 22:40 |
blackburn | heiko1: well it won't work :) | 22:40 |
blackburn | everybody would come to germany if it allows to avoid army :D | 22:41 |
heiko1 | ok | 22:41 |
heiko1 | is your thesis processed already? | 22:41 |
blackburn | there is a traditional russian way | 22:41 |
blackburn | processed in which mean? | 22:41 |
heiko1 | like feedback | 22:41 |
blackburn | I am unsure what do you mean :) | 22:42 |
blackburn | feedback from? | 22:42 |
heiko1 | supervisors | 22:42 |
heiko1 | like marks | 22:42 |
blackburn | ah | 22:42 |
blackburn | yes I defended it already | 22:42 |
heiko1 | oh really, nice | 22:42 |
heiko1 | congrats then :) | 22:42 |
blackburn | hah thanks | 22:42 |
blackburn | but I am a bachelor for two months already | 22:43 |
heiko1 | cool | 22:43 |
heiko1 | whats next? | 22:43 |
blackburn | masters here | 22:43 |
heiko1 | computer science? | 22:43 |
heiko1 | what courses and stuff? | 22:44 |
blackburn | oh well | 22:44 |
blackburn | you know we are crazy here | 22:44 |
blackburn | I will study just what they give me | 22:44 |
blackburn | :D | 22:44 |
heiko1 | and that is? | 22:44 |
blackburn | I have nothing to choose but it should be related to supercomputing | 22:44 |
blackburn | I don't care much about that - I am coming along with shogun and other stuff | 22:45 |
blackburn | I already had to learn everything by myself | 22:45 |
heiko1 | ok | 22:45 |
blackburn | so no reason to change that | 22:45 |
blackburn | :) | 22:45 |
heiko1 | well then :) | 22:45 |
blackburn | heiko1: what about you then? :) | 22:46 |
blackburn | are you going to do phd there? | 22:46 |
heiko1 | well, thesis first | 22:46 |
heiko1 | I got a small RA position here for about half a year afterwards | 22:46 |
heiko1 | and then I need another one :) | 22:46 |
blackburn | research assistant? | 22:46 |
blackburn | I have no clue what it is | 22:46 |
heiko1 | but i will apply for a phd that starts in sept | 22:46 |
heiko1 | yeah assistant | 22:46 |
blackburn | so it is just like a normal job? | 22:47 |
heiko1 | no its uni and research stuff | 22:47 |
heiko1 | but applied | 22:47 |
heiko1 | kernel methods to detect xy | 22:47 |
blackburn | yeah I mean normal research job | 22:47 |
heiko1 | yes | 22:47 |
blackburn | I see | 22:47 |
blackburn | is it well-paid in england btw? | 22:47 |
heiko1 | yes, especially in london | 22:48 |
blackburn | I heard londongrad is not cheap to live | 22:48 |
blackburn | :) | 22:48 |
heiko1 | no, its not cheap at all | 22:48 |
heiko1 | but my girlfriend has a well paid job, and I will join in that fashion, so its fine -- finally last year was so anoyying money wise | 22:49 |
blackburn | what does she do? | 22:49 |
heiko1 | she works in the citi, for a digital media data mining company | 22:49 |
blackburn | wow so a data mining? | 22:49 |
heiko1 | well, old fashioned data mining | 22:49 |
heiko1 | nothing advanced | 22:50 |
heiko1 | but collecting data | 22:50 |
blackburn | still geeky | 22:50 |
blackburn | :) | 22:50 |
heiko1 | shes more in the organisation and planning | 22:50 |
blackburn | I see | 22:50 |
blackburn | is she german too? | 22:50 |
blackburn | I am currently debugging MC output stuff | 22:51 |
blackburn | which collects ROCs | 22:51 |
heiko1 | yeah | 22:51 |
heiko1 | monte carlo ? | 22:51 |
heiko1 | multiclass :) | 22:51 |
blackburn | multiclass :) | 22:51 |
heiko1 | sorry | 22:51 |
blackburn | but I don't mind to have some monte carlo out there | 22:52 |
blackburn | may be some casino royale | 22:52 |
heiko1 | Oh, I want to work on the labels a bit on the weekend | 22:52 |
blackburn | would be interested | 22:52 |
blackburn | :D | 22:52 |
heiko1 | lol :) | 22:52 |
blackburn | have you seen Ernst&Young guy? | 22:52 |
blackburn | in the ML | 22:52 |
heiko1 | yeah | 22:52 |
heiko1 | pretty cool :) | 22:52 |
heiko1 | machine learning arrives in the consulting | 22:53 |
blackburn | we are all perplexed how can it be used | 22:53 |
heiko1 | dont know | 22:53 |
blackburn | yeah | 22:54 |
heiko1 | all the bankers are using (bayesian) linear regression | 22:54 |
blackburn | we was puzzled with the fashion he names himself | 22:54 |
heiko1 | one of our profs here always tells us whats going on in the citi | 22:54 |
heiko1 | linear regresseion :) | 22:54 |
blackburn | heh | 22:54 |
heiko1 | I did not get your last scentence | 22:55 |
blackburn | he names himself | 22:55 |
blackburn | arthur (sean) | 22:55 |
blackburn | at first he wrote sth like | 22:55 |
blackburn | cheers, sean | 22:55 |
blackburn | but email is arthur | 22:55 |
blackburn | that's kind of puzzle :D | 22:55 |
heiko1 | ah, mmh no idea | 22:56 |
heiko1 | blackburn, what about the labels | 22:56 |
heiko1 | with the confidences /scores | 22:56 |
blackburn | ah | 22:56 |
heiko1 | have you seen s?rens email? | 22:56 |
blackburn | oh bad bad | 22:56 |
blackburn | which one? | 22:56 |
heiko1 | I liked the idea of registering multiple vectors with a name | 22:56 |
blackburn | I remember one | 22:56 |
heiko1 | and then you set the "current" one | 22:56 |
heiko1 | and call a method | 22:56 |
heiko1 | this way you can switchj | 22:56 |
blackburn | the most confusing thing for me is that we all wanted to release | 22:57 |
heiko1 | yeah maybe leave that then | 22:57 |
heiko1 | the method is there | 22:58 |
heiko1 | though its only for binary | 22:58 |
blackburn | I have checked multiclass stuff a little | 22:58 |
heiko1 | argh there is still the migration stuff | 22:59 |
blackburn | oops | 22:59 |
heiko1 | that will be horrible | 22:59 |
blackburn | so do we need to set up all the maps now? | 22:59 |
heiko1 | all these changes to the vectors cause loads of memory bugs | 22:59 |
heiko1 | dont know yet | 22:59 |
heiko1 | if we can migrate, maybe | 22:59 |
heiko1 | but I think this wont work | 22:59 |
heiko1 | so we will break | 22:59 |
heiko1 | and start from scratch | 22:59 |
heiko1 | that stuff is a mess, especially in c++ without any libraries | 23:00 |
blackburn | argh mkl storage fails with mkl multiclass | 23:00 |
heiko1 | and whenever somethings changed, it fails | 23:00 |
heiko1 | oh reall, what happens? | 23:00 |
blackburn | some error message I can't see | 23:01 |
blackburn | hmm | 23:01 |
blackburn | ahhhhh | 23:01 |
blackburn | heiko1: it handles MKLMulticlass as multiclass | 23:02 |
heiko1 | and then? | 23:03 |
blackburn | hmmm | 23:03 |
blackburn | something strange happens then | 23:03 |
heiko1 | mmh, if errors should be easy to track | 23:04 |
blackburn | heiko1: we should mark down all these *outputs as experimental stuff | 23:04 |
blackburn | to aavoid some psycho trying to find us and eat | 23:04 |
heiko1 | mmh | 23:05 |
heiko1 | I mean the idea is that only classes that are supported work with it | 23:05 |
heiko1 | so every class gets another implementation of the output class | 23:05 |
heiko1 | the output class can ignore things | 23:06 |
blackburn | heiko1: are you able to check code? | 23:06 |
heiko1 | yeah, I want to go to bed, but send it :) | 23:07 |
blackburn | ah nevermind I'd rather paste | 23:07 |
blackburn | 84 ????????????SG_PRINT("%smulti-class machine %d:\n", i); | 23:07 |
blackburn | something is wrong out there | 23:07 |
blackburn | what was the intention? | 23:07 |
blackburn | there are two formatters but one var | 23:07 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 23:08 | |
blackburn | n4nd0: ! | 23:08 |
n4nd0 | blackburn: hey! | 23:08 |
n4nd0 | are we all happy with the final evaluations? did everyone pass? | 23:08 |
blackburn | yes I think so | 23:08 |
heiko1 | should :) | 23:08 |
heiko1 | at least I am happy | 23:08 |
blackburn | I wanted to develop a simple application that segments an image | 23:09 |
n4nd0 | blackburn: ok | 23:09 |
blackburn | so it is a two state model I guess | 23:09 |
blackburn | I wanted to ask if we are ready for that | 23:09 |
blackburn | can that be done using current framework? | 23:09 |
n4nd0 | I have not much idea about image segmentation | 23:09 |
n4nd0 | can you tell me more about the application? | 23:09 |
blackburn | n4nd0: okay let me try to describe | 23:10 |
heiko1 | blackburn, talk to you later, I need to sleep | 23:10 |
blackburn | heiko1: okay sure | 23:10 |
n4nd0 | heiko1: good night! | 23:10 |
blackburn | n4nd0: simplest case is image having foreground and background | 23:10 |
heiko1 | night guys! :) | 23:10 |
blackburn | nite heiko1 | 23:10 |
blackburn | n4nd0: so two hidden states | 23:10 |
blackburn | and each pixel has its own state | 23:11 |
blackburn | observations are not so clear to me though | 23:11 |
n4nd0 | blackburn: the actual values of the pixels I guess | 23:12 |
blackburn | yes but in neighborhood of each pixel | 23:12 |
blackburn | I guess | 23:12 |
n4nd0 | and probably something that gives infor about the pixels around | 23:12 |
n4nd0 | exatly | 23:12 |
n4nd0 | exactly* | 23:12 |
blackburn | I still have weak belief how can that work :D | 23:13 |
n4nd0 | I'd say that the framework should work for that | 23:13 |
n4nd0 | but we need to find some information about the features that people use for this | 23:13 |
blackburn | it is not a problem lets try to come along with pixels | 23:13 |
n4nd0 | blackburn: have you read about it in some paper? | 23:13 |
blackburn | no, not really we discussed it with soeren | 23:13 |
blackburn | so it comes from air | 23:14 |
blackburn | and can be wrong :D | 23:14 |
n4nd0 | I think so far sounds reasonable | 23:14 |
blackburn | I have to draw | 23:14 |
-!- heiko1 [~heiko@host86-178-156-19.range86-178.btcentralplus.com] has left #shogun [] | 23:14 | |
n4nd0 | but more research on the topic is probably necessary though | 23:14 |
blackburn | I had a link somewhere | 23:14 |
blackburn | to web service allowing to draw | 23:15 |
blackburn | :D | 23:15 |
blackburn | n4nd0: do you know any service like that? | 23:16 |
n4nd0 | mmm no | 23:16 |
blackburn | http://www.twiddla.com/496234 | 23:16 |
blackburn | n4nd0: here you go | 23:16 |
blackburn | n4nd0: waiting for you ^ :) | 23:18 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Ping timeout: 240 seconds] | 23:38 | |
CIA-52 | shogun: Sergey Lisitsyn master * rcfc8322 / (8 files in 3 dirs): Introduced CV MC storage - http://git.io/a_MgjA | 23:43 |
-!- gsomix [~gsomix@178.45.65.87] has quit [Ping timeout: 256 seconds] | 23:56 | |
--- Log closed Sat Aug 25 00:00:17 2012 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!