--- Log opened Wed Apr 11 00:00:00 2012 | ||
--- Day changed Wed Apr 11 2012 | ||
n4nd0 | so tomorrow I will make sure that this cover tree is actually giving the right answers | 00:00 |
---|---|---|
n4nd0 | and if it does, I will start to port it :) | 00:00 |
blackburn | n4nd0: okay | 00:00 |
n4nd0 | good night :) | 00:00 |
blackburn | good night | 00:00 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 00:01 | |
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking] | 00:03 | |
-!- wiking [~wiking@78-23-191-201.access.telenet.be] has joined #shogun | 00:08 | |
-!- wiking [~wiking@78-23-191-201.access.telenet.be] has quit [Changing host] | 00:08 | |
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun | 00:08 | |
harshit_ | good night everyone :) | 00:58 |
-!- harshit_ [~harshit@182.68.210.237] has quit [Quit: Leaving] | 00:59 | |
PhilTillet | blackburn, not sleeping? | 01:07 |
PhilTillet | :p | 01:07 |
-!- av3ngr [av3ngr@nat/redhat/x-fslqwfzaqdfhtuml] has joined #shogun | 01:12 | |
-!- PhilTillet [~Philippe@npasserelle10.minet.net] has quit [Remote host closed the connection] | 01:15 | |
-!- blackburn [5bde8018@gateway/web/freenode/ip.91.222.128.24] has quit [Quit: Page closed] | 01:23 | |
@sonney2k | wiking, btw the qnorm 'bug' should not have been critical because for mkl we only considered positive kernel weights | 02:30 |
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking] | 04:16 | |
-!- pluskid [~chatzilla@111.120.64.123] has joined #shogun | 04:55 | |
-!- pluskid [~chatzilla@111.120.64.123] has quit [Client Quit] | 04:56 | |
-!- pluskid [~chatzilla@111.120.81.178] has joined #shogun | 05:41 | |
-!- gsomix [~gsomix@188.168.128.186] has quit [Ping timeout: 245 seconds] | 05:51 | |
-!- pluskid [~chatzilla@111.120.81.178] has quit [Quit: ChatZilla 0.9.88.2 [Firefox 11.0/20120314124128]] | 06:31 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 07:29 | |
-!- wiking [~wiking@78-23-191-201.access.telenet.be] has joined #shogun | 08:33 | |
-!- wiking [~wiking@78-23-191-201.access.telenet.be] has quit [Changing host] | 08:33 | |
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun | 08:33 | |
n4nd0 | sonne|work: I finished cheking all the results, this cover tree gives the right output :) | 08:33 |
n4nd0 | sonne|work: at least the same as our current KNN with qsort | 08:34 |
n4nd0 | and the interesting part, the query time is quite shorter! | 08:35 |
-!- wiking [~wiking@huwico/staff/wiking] has quit [Remote host closed the connection] | 08:35 | |
-!- wiking [~wiking@vpna213.ugent.be] has joined #shogun | 08:36 | |
-!- wiking [~wiking@vpna213.ugent.be] has quit [Changing host] | 08:36 | |
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun | 08:36 | |
-!- blackburn [5bdfb203@gateway/web/freenode/ip.91.223.178.3] has joined #shogun | 08:51 | |
-!- wiking [~wiking@huwico/staff/wiking] has quit [Quit: wiking] | 08:59 | |
n4nd0 | blackburn: hey, good morning! | 09:01 |
n4nd0 | blackburn: I just sent a small pull request for KNN | 09:02 |
blackburn | n4nd0: morgen :) | 09:02 |
blackburn | damn I thought I pushed my changes | 09:02 |
n4nd0 | I didn't see them in the repo | 09:03 |
n4nd0 | no problem :) | 09:03 |
n4nd0 | blackburn: a detail for the cover tree | 09:03 |
n4nd0 | blackburn: should I substitute directly the current one we have or use another name (CoverTree2 for example) | 09:04 |
@sonney2k | n4nd0, so JL's covertree works right? | 09:04 |
@sonney2k | morning blackburn | 09:04 |
n4nd0 | sonney2k: it is not JL's implementation | 09:04 |
n4nd0 | sonney2k: it is this one https://github.com/mbrucher/CoverTree | 09:04 |
@sonney2k | hmmhh - admittedly I am a big fan of JL's stuff. | 09:05 |
@sonney2k | what he does is reall fast(tm) | 09:05 |
blackburn | sonney2k: morgen. today is the day? ;) | 09:05 |
n4nd0 | sonney2k: yeah but I don't really understand that implementation | 09:05 |
@sonney2k | that is the downside of his code - but just ask him then | 09:06 |
@sonney2k | blackburn, not my day | 09:06 |
n4nd0 | sonney2k: mailing list? | 09:06 |
@sonney2k | tomorrow I would say :D | 09:06 |
@sonney2k | n4nd0, personal email | 09:06 |
n4nd0 | sonney2k: ok, in any case, do you think it pays off to start porting this implementation? | 09:07 |
@sonney2k | I would rather try to get JL's implementation to work | 09:07 |
n4nd0 | ok ... | 09:07 |
n4nd0 | this one presents a good improvement in time though | 09:07 |
n4nd0 | Shogun's time: 00:00:00.328804 | 09:07 |
n4nd0 | CoverTree time 00:00:00.053650 | 09:07 |
blackburn | sonney2k: btw I managed to beat all my competitors with 'brand new libedrt' :D | 09:08 |
-!- uricamic [9320543b@gateway/web/freenode/ip.147.32.84.59] has joined #shogun | 09:08 | |
n4nd0 | that is just 100 queries using 10000 training vectors | 09:08 |
n4nd0 | sonney2k: another thing I was thinking about our KNN | 09:08 |
uricamic | Hi | 09:09 |
n4nd0 | sonney2k: right now we have N?log(N) for every query, because we sort the distances from the query vector to all the training vectors | 09:09 |
@sonney2k | n4nd0, well then benchmark JLs impl against this one | 09:09 |
@sonney2k | uricamic, hi... | 09:09 |
n4nd0 | sonney2k: yeah, I can do that ... but we won't know if the results are correct | 09:10 |
@sonney2k | are you back from vacations? did you think of some small patch? | 09:10 |
n4nd0 | sonney2k: so it won't fell safe in any case | 09:10 |
@sonney2k | n4nd0, but you can compare the output to the qsort/other covertree | 09:10 |
n4nd0 | sonney2k: yes, to the cover tree I am using now, the comparison is done | 09:10 |
n4nd0 | sonney2k: I am quite certain it is correct | 09:11 |
uricamic | sonney2k: yes, I plan to work on it this week, and I have one question | 09:11 |
@sonney2k | n4nd0, yeah - I meant you can do the same with JL | 09:11 |
n4nd0 | sonney2k: I have tried a very long test (10000 training vector, 10000 queries with 5 neighbors) and it works | 09:11 |
n4nd0 | sonney2k: no, I cannot do that with JL | 09:11 |
@sonney2k | and the winner is the one we integrate | 09:11 |
uricamic | sonney2k: is it possible to write very simple interface for the libbmrm as a patch, or hsould I choose something different? | 09:11 |
@sonney2k | n4nd0, why not? | 09:12 |
@sonney2k | uricamic, I would prefer something completely different | 09:12 |
@sonney2k | blackburn, do you have some *small* idea for uricamic? | 09:12 |
uricamic | sonney2k: ok | 09:12 |
n4nd0 | sonney2k: when I query 10 points with k = 10 with JL, I get 10 neighbors for the first query | 09:12 |
n4nd0 | sonney2k: 1 for each of the others ... | 09:12 |
blackburn | sonney2k: I proposed parzen stuff before | 09:12 |
@sonney2k | uricamic, ^ | 09:12 |
@sonney2k | so feel free | 09:13 |
blackburn | sonney2k: do you think it worths? | 09:13 |
@sonney2k | uricamic, other relatively simple algos are also ok | 09:13 |
uricamic | ok, I will start with the Parzen windows | 09:13 |
uricamic | ok | 09:13 |
uricamic | thanks | 09:13 |
blackburn | ah or some nadaraya-watson or etc | 09:13 |
n4nd0 | sonney2k: but I can try to check again if I am doing correctly this test, but I think so | 09:13 |
blackburn | :D | 09:13 |
@sonney2k | for example some iterative method for ridge regression (not computing pinv) | 09:13 |
@sonney2k | n4nd0, well ask JL directly in CC me in the email why this is the case and give him your code snippet | 09:14 |
n4nd0 | sonney2k: good idea :) thank you | 09:14 |
@sonney2k | uricamic, that was for you: you could for example also do some iterative method for ridge regression (not computing pinv) | 09:15 |
@sonney2k | if that is more in your field of expertise | 09:15 |
@sonney2k | as optimization guru :) | 09:15 |
uricamic | sonney2k: ok, I will look at it and then decide | 09:16 |
@sonney2k | uricamic, just look in the shogun/regression/ directory... these all are pinv based things | 09:16 |
uricamic | sonney2k: ok, thanks | 09:17 |
n4nd0 | sonney2k: about the other strategy for KNN, the qsort one, right now we are in N?log(N) but that can be done in N easy | 09:18 |
n4nd0 | sonney2k: instead of sorting all the distances and taking the K shortest | 09:18 |
n4nd0 | sonney2k: loop through all the distances and keep a record of the K shortest one found | 09:18 |
blackburn | n4nd0: really, how? ;) | 09:20 |
@sonney2k | n4nd0, yes that would be faster for small k | 09:20 |
blackburn | it won't be N still I mean | 09:21 |
@sonney2k | n4nd0, I guess no one is using k>11 | 09:21 |
@sonney2k | blackburn, it is N*k | 09:21 |
n4nd0 | sonney2k: I think it is something like N?k?log(k) | 09:21 |
blackburn | depends on structure you will use to store neighbors | 09:21 |
blackburn | but with lame way it is even Nk^2 | 09:22 |
n4nd0 | using just a vector of k elements | 09:22 |
n4nd0 | that you order in every iteration | 09:22 |
n4nd0 | is just N?k?log(k) | 09:22 |
n4nd0 | and I know nothing lamer than that ... | 09:22 |
blackburn | but you would need to compare with elements | 09:22 |
blackburn | in worst case with each element | 09:22 |
@sonney2k | n4nd0, I thought you meant sth like going through the N elements and keeping a history of the k neighreast | 09:22 |
blackburn | right? | 09:22 |
n4nd0 | sonney2k: yeah, that is the idea | 09:23 |
blackburn | n4nd0: I used fibonacci heap for that before :D | 09:23 |
n4nd0 | blackburn: what complexity gives with fibonacci heap? | 09:23 |
n4nd0 | I normally solve this problem using std::set | 09:23 |
blackburn | minimum is log N amortized | 09:24 |
n4nd0 | you insert in log(k) | 09:24 |
blackburn | so k*log N | 09:24 |
@sonney2k | gtg | 09:24 |
@sonney2k | cu | 09:24 |
blackburn | insert is O(1) in the fib heap | 09:24 |
n4nd0 | sonney2k: bye | 09:24 |
n4nd0 | blackburn: so you insert in O(1) and later you sort it? | 09:24 |
blackburn | yes something like that | 09:25 |
n4nd0 | ok | 09:25 |
blackburn | actually you could measure it too | 09:25 |
n4nd0 | but when you said | 09:25 |
n4nd0 | 09:22 < blackburn> but you would need to compare with elements | 09:25 |
n4nd0 | 09:22 < blackburn> in worst case with each element | 09:25 |
n4nd0 | you are playing there with worst-time complexity of qsort?? | 09:25 |
blackburn | no | 09:25 |
n4nd0 | if not, I don't see what you mean | 09:25 |
blackburn | okay let us have x1,x2,x3,x4,... | 09:25 |
blackburn | when you collect neighbors | 09:26 |
blackburn | let we have some collection already | 09:26 |
blackburn | xi1,xi2,xi3 | 09:26 |
n4nd0 | ok | 09:26 |
blackburn | and if we need to know if new vector x is a neighbor | 09:26 |
blackburn | we would need to compare all these distances right? | 09:27 |
n4nd0 | no | 09:27 |
blackburn | why? | 09:27 |
n4nd0 | you just compare with the one that gives the largest distance | 09:27 |
n4nd0 | among these neighbors | 09:27 |
n4nd0 | that's why you need to have it sorted | 09:27 |
blackburn | sorted? | 09:27 |
n4nd0 | yeah | 09:27 |
n4nd0 | so you have this collection of k tentative neighbors | 09:27 |
n4nd0 | sorted | 09:28 |
blackburn | okay and if you have distance 1.5 | 09:28 |
blackburn | while largest is 5 | 09:28 |
blackburn | and smallest is 1.25 | 09:28 |
n4nd0 | then 5 pops out | 09:28 |
blackburn | you would need to compare to place it? | 09:28 |
n4nd0 | 1.5 enters | 09:28 |
n4nd0 | no | 09:28 |
n4nd0 | the largest goes out | 09:28 |
blackburn | ah | 09:28 |
n4nd0 | the new one goes in, and you sort again | 09:28 |
blackburn | okay got it | 09:29 |
n4nd0 | N?k?log(k) :D | 09:29 |
n4nd0 | "assuming" that qsort is n?log(n) for n elements | 09:29 |
n4nd0 | so we will normally have complexity N because N >> k | 09:30 |
n4nd0 | and that would make our current KNN to work in O(N) instead of O(N?log(N)) | 09:30 |
n4nd0 | blackburn: did I convince you? | 09:30 |
blackburn | n4nd0: just measure :) | 09:30 |
n4nd0 | sure | 09:30 |
n4nd0 | blackburn: I think that is my homework for today | 09:31 |
n4nd0 | I am going to change it, measure and mail JL | 09:31 |
blackburn | yeah a lot of things to do | 09:31 |
-!- nickon [~noneedtok@dD5774105.access.telenet.be] has joined #shogun | 09:31 | |
n4nd0 | blackburn: I was thinking of that because I remember you said that scikits people wrote that our KNN is slow | 09:34 |
n4nd0 | blackburn: and AFAIK they don't have cover tree either | 09:34 |
blackburn | n4nd0: kdtree | 09:37 |
blackburn | or balltree | 09:37 |
n4nd0 | I think kd-trees work fine in low dimensionality problems | 09:38 |
n4nd0 | blackburn: I have just realized that KNN comes with SG_PROGRESS, how is that used? | 09:43 |
blackburn | I would avoid kdtree | 09:43 |
blackburn | n4nd0: knn.io.enable_progress() | 09:43 |
n4nd0 | blackburn: ok, thank you | 09:44 |
n4nd0 | blackburn: why would you avoid kd-trees? | 09:44 |
blackburn | n4nd0: that thing should be everywhere actually | 09:44 |
blackburn | n4nd0: because it would work only for simplefeatures | 09:44 |
sonne|work | n4nd0: mail JL first then start the play stuff :) | 09:46 |
sonne|work | n4nd0: btw JL is located in NY so you cannot quickly get an answer | 09:46 |
n4nd0 | sonne|work: :( haha ok, I was starting to play first, such a kid :P | 09:46 |
n4nd0 | sonne|work: the sooner I mail him, the better then | 09:47 |
blackburn | sonne|work: is there anything except covertree and vw he has we can integrate? | 09:47 |
sonne|work | blackburn: his vw stuff changed completely | 09:47 |
sonne|work | they have some massiviely parallel thing in there now | 09:48 |
blackburn | sonne|work: I have seen that - bfgs-l is here now | 09:48 |
n4nd0 | particle filters! :P | 09:48 |
-!- av3ngr [av3ngr@nat/redhat/x-fslqwfzaqdfhtuml] has quit [Ping timeout: 246 seconds] | 10:06 | |
n4nd0 | sonne|work: there is sth weird, JL's non-templated implementation gives the correct number of neighbors while the templated (the one I was using for my test) not | 10:07 |
sonne|work | then sb introduced a bug :) | 10:07 |
n4nd0 | sonne|work: I am going to try to find out why is this, the mail may wait | 10:08 |
n4nd0 | yeah, either I did it in the test or it was in this templated version | 10:08 |
sonne|work | n4nd0: maybe you can then even compare the 2 cover trees first | 10:08 |
n4nd0 | I think the first alternative is more likely | 10:08 |
n4nd0 | sonne|work: which two ones? | 10:08 |
sonne|work | the one you dig up and JLs | 10:09 |
blackburn | n4nd0: could you please remind me how does non-templated stuff works? | 10:15 |
n4nd0 | blackburn: the point is fixed to: typedef float* point | 10:16 |
blackburn | n4nd0: how distances are computed then? | 10:17 |
sonne|work | floats are really crazy ... http://www.altdevblogaday.com/2012/04/05/floating-point-complexities/ | 10:17 |
n4nd0 | blackburn: normal Euclidean distance | 10:18 |
n4nd0 | blackburn: I think I didn't get the point of the question sorry | 10:18 |
blackburn | n4nd0: hardcoded? can you change it? | 10:18 |
n4nd0 | blackburn: you mean like how many elements do they use? | 10:18 |
blackburn | n4nd0: no how hard it could be to adapt non-templated? | 10:18 |
blackburn | or in other words is it any possible ? | 10:19 |
n4nd0 | blackburn: so take a look to the distance function for example | 10:19 |
-!- wiking [~wiking@vpnb121.ugent.be] has joined #shogun | 10:19 | |
-!- wiking [~wiking@vpnb121.ugent.be] has quit [Changing host] | 10:19 | |
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun | 10:19 | |
blackburn | sorry overwhelmed with stuff at job but was curious :) | 10:19 |
blackburn | n4nd0: I mean can you provide you own distance there? | 10:19 |
wiking | mmm | 10:19 |
wiking | anybody knows how can i force an ext3 fs to be clean? | 10:19 |
wiking | :) | 10:19 |
blackburn | no idea | 10:20 |
n4nd0 | blackburn: http://snipt.org/uhGh8 | 10:20 |
n4nd0 | blackburn: I think that by changing that function that can be done | 10:20 |
blackburn | yes | 10:20 |
blackburn | ok | 10:20 |
n4nd0 | wiking: no idea either | 10:20 |
wiking | hehehe trying to hack with tune2fs and dumpfs :> | 10:21 |
sonne|work | wiking: hi ... what errors where you getting with clang? because here it compiles fine w/ clang... | 10:21 |
n4nd0 | blackburn: there are things I don't get anyway | 10:21 |
wiking | sonne|work: will give you the list | 10:21 |
n4nd0 | blackburn: if I set k = 9, I get 10 neighbors | 10:21 |
wiking | sonne|work: just now i'm trying to save my work while the FS is collapsing :) | 10:21 |
n4nd0 | blackburn: the k used is the k given +1 | 10:22 |
blackburn | n4nd0: Iguess you get object itself? | 10:22 |
n4nd0 | blackburn: aham, let me check | 10:23 |
sonne|work | wiking: why do you want the fs to become clean | 10:23 |
sonne|work | wiking: just mount it | 10:23 |
wiking | sonne|work: i cannot boot | 10:23 |
wiking | just in single user mode | 10:23 |
n4nd0 | blackburn: I don't really know if it is because of that | 10:24 |
sonne|work | wiking: then boot in single user mode copy stuff over and run fsck | 10:24 |
wiking | hehehe i'm actually on it | 10:24 |
wiking | the problem here is that it's an nfs | 10:24 |
wiking | a | 10:24 |
wiking | and i don't have space to save that much data only when i have network | 10:25 |
wiking | but i think | 10:25 |
wiking | iiii think i've made it now :) | 10:25 |
wiking | damn raid controllers... | 10:39 |
-!- sonne|work [~sonnenbu@194.78.35.195] has left #shogun [] | 10:46 | |
-!- blackburn [5bdfb203@gateway/web/freenode/ip.91.223.178.3] has quit [Quit: Page closed] | 10:50 | |
-!- sonne|work [~sonnenbu@194.78.35.195] has joined #shogun | 11:10 | |
CIA-64 | shogun: puffin444 master * r6f128a5 / (4 files in 3 dirs): | 11:16 |
CIA-64 | shogun: Mean predictions in GaussianProcessRegression now solved using Cholesky | 11:16 |
CIA-64 | shogun: decomposition. Predicted Covariance Matrix is also accessible. - http://git.io/wJMwHg | 11:16 |
CIA-64 | shogun: Soeren Sonnenburg master * ra27e1ae / (8 files in 5 dirs): | 11:16 |
CIA-64 | shogun: Merge pull request #406 from puffin444/master | 11:16 |
CIA-64 | shogun: Gaussian Process Regression - http://git.io/1hENEw | 11:16 |
CIA-64 | shogun: Soeren Sonnenburg master * r2d37d40 / src/shogun/classifier/KNN.cpp : | 11:17 |
CIA-64 | shogun: Merge pull request #440 from iglesias/knn-covertree | 11:17 |
CIA-64 | shogun: Minor bugfixes in KNN - http://git.io/20opjw | 11:17 |
n4nd0 | it was my bad, there were some bugs in my implementation of the point I was using for JL | 11:22 |
n4nd0 | I am sorry for the confusion | 11:22 |
sonne|work | :) | 11:23 |
sonne|work | n4nd0: so now benchmark please :D | 11:23 |
sonne|work | then we can decide | 11:24 |
n4nd0 | sonne|work: there is one issue though | 11:24 |
n4nd0 | I believe that JL doesn't give the points in order (closest to farthe) | 11:25 |
n4nd0 | the neighbors I mean | 11:25 |
n4nd0 | so a test to ensure that the results are correct would take me a bit more time | 11:26 |
n4nd0 | we can benchmark it though and assume that the results are correct (they must be) | 11:26 |
sonne|work | well try if they are exactly the same... | 11:30 |
sonne|work | then you know for sure that JL's covertree gives (un)ordered results | 11:31 |
sonne|work | however, it is very cheap if you would just compute the distance to these k points and sort accordingly (compared to all the n log(n) stuff before) | 11:32 |
shogun-buildbot | build #662 of cmdline_static is complete: Failure [failed test_1] Build details are at http://www.shogun-toolbox.org/buildbot/builders/cmdline_static/builds/662 blamelist: root@localhost.(none), sonne@debian.org, jacobw@localhost.(none) | 11:50 |
shogun-buildbot | build #640 of r_static is complete: Failure [failed test_1] Build details are at http://www.shogun-toolbox.org/buildbot/builders/r_static/builds/640 blamelist: root@localhost.(none), sonne@debian.org, jacobw@localhost.(none) | 11:54 |
shogun-buildbot | build #641 of octave_static is complete: Failure [failed test_1] Build details are at http://www.shogun-toolbox.org/buildbot/builders/octave_static/builds/641 blamelist: root@localhost.(none), sonne@debian.org, jacobw@localhost.(none) | 11:58 |
-!- gsomix [~gsomix@188.168.2.126] has joined #shogun | 12:03 | |
gsomix | hello all | 12:04 |
shogun-buildbot | build #628 of python_static is complete: Failure [failed test_1] Build details are at http://www.shogun-toolbox.org/buildbot/builders/python_static/builds/628 blamelist: root@localhost.(none), sonne@debian.org, jacobw@localhost.(none) | 12:10 |
-!- Johannes__ [55e5c1c3@gateway/web/freenode/ip.85.229.193.195] has joined #shogun | 12:12 | |
Johannes__ | Hello, where can i find the shogun-java-modular package the installation guide talks about on ubuntu? | 12:13 |
Johannes__ | I can find some of the others through the usual channels but not that one. | 12:14 |
n4nd0 | Johannes__: hey! it doesn't appear directly in my apt-get either, is it said around there to include any special repository for apt-get?? | 12:16 |
n4nd0 | sonne|work: I got results of the benchmark :) | 12:16 |
Johannes__ | n4nd0: Doesen't mention any special repository to include as far as I can see. | 12:35 |
n4nd0 | Johannes__: aham, then I don't really know | 12:35 |
n4nd0 | Johannes__: have you considered installing from source? | 12:36 |
n4nd0 | sonne|work: JL's is considerably faster | 12:36 |
n4nd0 | sonne|work: it is reasonable to some extent since in his version the algorihtms for creation of the tree and query are batch algorihtms | 12:37 |
Johannes__ | Considered, yes but will probably look around for alternatives first, tried the macport install on my laptop without success aswell ;p | 12:37 |
n4nd0 | Johannes__: but you are on ubuntu or in mac? | 12:37 |
Johannes__ | both, the shogun-java-modular question was in regards to ubuntu | 12:38 |
n4nd0 | Johannes__: ahm | 12:39 |
n4nd0 | Johannes__: anyway, it should not be hard to install in ubuntu from source, just say it if you want some help with that | 12:40 |
Johannes__ | Yea perhaps that's the way to go, we'll see. Thanks for the offer and answers regardless :) | 12:42 |
n4nd0 | no problem :) | 12:44 |
sonne|work | hi gsomix | 13:08 |
sonne|work | n4nd0: with batch you mean - one has to specify *all* train data at once and *all* test data at once? | 13:08 |
n4nd0 | sonne|work: yes, both for the training data and the test data | 13:22 |
n4nd0 | sonne|work: actually the way it works is doing a cover tree for the test data too | 13:22 |
sonne|work | so you end up with k*T indices? | 13:23 |
sonne|work | uhh and what are the memory requirements for that? | 13:23 |
n4nd0 | I have not measured it myself | 13:23 |
n4nd0 | but in the paper they say that a cover tree needs O(n) in memory | 13:24 |
sonne|work | ok | 13:24 |
n4nd0 | http://snipt.org/uhHe8 | 13:24 |
sonne|work | sounds great... | 13:25 |
n4nd0 | the build time of the other's cover tree really sucks | 13:25 |
sonne|work | impressive | 13:26 |
sonne|work | build time? | 13:26 |
sonne|work | why? what? | 13:26 |
sonne|work | you mean compile time | 13:26 |
n4nd0 | no | 13:26 |
sonne|work | or construction of the tree? | 13:26 |
n4nd0 | to create the cover tree | 13:26 |
sonne|work | well I think it is pretty clear that we want JL's tree :) | 13:27 |
n4nd0 | yeah ... bad new for my health status :P | 13:27 |
n4nd0 | sonne|work: so how would you like that we use it? use the same code and provide wrappers? | 13:28 |
n4nd0 | sonne|work: is it better to re-write? | 13:28 |
n4nd0 | btw, I didn't check the correctnes of the results in JL yet! | 13:28 |
n4nd0 | I just assumed that they are ok | 13:28 |
sonne|work | n4nd0: will I said so upfront :D | 13:36 |
sonne|work | ok | 13:36 |
sonne|work | n4nd0: it depends - a shallow wrapper is easier to maintain in case there are fixes to JL's covertree | 13:36 |
-!- nickon [~noneedtok@dD5774105.access.telenet.be] has quit [Quit: ( www.nnscript.com :: NoNameScript 4.22 :: www.esnation.com )] | 13:45 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 14:08 | |
-!- Johannes__ [55e5c1c3@gateway/web/freenode/ip.85.229.193.195] has quit [Quit: Page closed] | 14:15 | |
-!- PhilTillet_ [~Philippe@157.159.42.154] has joined #shogun | 14:20 | |
-!- PhilTillet_ [~Philippe@157.159.42.154] has quit [Ping timeout: 240 seconds] | 14:33 | |
gsomix | sonne|work, moin. I'm working on SGVector, but slow. I think I can turn on "berserk mode" in Friday or Saturday. | 15:10 |
sonne|work | gsomix: before doing that - did you fix the SG_PROGRESS stuff? | 15:11 |
sonne|work | yes ok | 15:11 |
sonne|work | no problem | 15:11 |
-!- pluskid [~chatzilla@173.254.214.60] has joined #shogun | 15:46 | |
sonne|work | *argh* | 16:02 |
sonne|work | http://www.shogun-toolbox.org/buildbot/builders/octave_static/builds/642/steps/test_1/logs/stdio | 16:02 |
sonne|work | buildbot failing | 16:02 |
gsomix | sonne|work, about SG_PROGRESS. should I do overall progress on all four cycle? | 16:03 |
sonne|work | one overall progress (approximate is sufficient) | 16:04 |
gsomix | sonne|work, ok. | 16:04 |
gsomix | sonne|work, and about references. I figured out, that references is true way. | 16:09 |
sonne|work | gsomix: ahh how come? | 16:19 |
sonne|work | gsomix: what changed your mind | 16:19 |
sonne|work | ? | 16:19 |
sonne|work | pluskid: could you please set a proper username? | 16:20 |
sonne|work | http://help.github.com/set-your-user-name-email-and-github-token/ | 16:20 |
sonne|work | pluskid, uricamic these are the instructions | 16:20 |
pluskid | sonne|work: you mean set username to real name? | 16:21 |
gsomix | sonney2k, wiking wonderful analyzed this situation, hehe. :) | 16:21 |
gsomix | * sonne|work, | 16:22 |
wiking | :> | 16:22 |
-!- harshit_ [~harshit@182.68.45.192] has joined #shogun | 16:22 | |
sonne|work | pluskid: yes | 16:22 |
pluskid | ok, changed | 16:22 |
sonne|work | it helps you to get credit! | 16:22 |
pluskid | but we cannot update already committed code maybe? | 16:23 |
sonne|work | gsomix: I see :)( | 16:23 |
uricamic | sonne|work: I have already set it, sorry that I haven't done it for the first commit | 16:23 |
sonne|work | pluskid: no that is not possible | 16:23 |
sonne|work | but for new pull requests at least | 16:23 |
pluskid | ok, see it | 16:23 |
sonne|work | https://github.com/shogun-toolbox/shogun/contributors | 16:25 |
sonne|work | pluskid: we have a couple of weird contributors in there | 16:25 |
sonne|work | invalid-email-address (You didn't set your git config correctly) | 16:25 |
sonne|work | is the best one (blackburns favourite :) | 16:25 |
pluskid | :D | 16:26 |
pluskid | it consists of a lot of people :D | 16:26 |
sonne|work | uricamic: no problem | 16:29 |
sonne|work | we all are not yet git experts | 16:29 |
-!- vikram360 [~vikram360@117.192.173.208] has joined #shogun | 17:02 | |
pluskid | /quit | 17:11 |
pluskid | er... | 17:12 |
pluskid | failed to quit... | 17:12 |
-!- pluskid [~chatzilla@173.254.214.60] has quit [Quit: ChatZilla 0.9.88.2 [Firefox 11.0/20120314124128]] | 17:12 | |
-!- pythonroar [d2198538@gateway/web/freenode/ip.210.25.133.56] has joined #shogun | 17:16 | |
-!- pythonroar [d2198538@gateway/web/freenode/ip.210.25.133.56] has quit [Quit: Page closed] | 17:26 | |
-!- vikram360 [~vikram360@117.192.173.208] has quit [Ping timeout: 265 seconds] | 17:31 | |
-!- vikram360 [~vikram360@117.192.180.124] has joined #shogun | 17:32 | |
gsomix | I'm going to a dance classes. | 17:39 |
gsomix | cu | 17:39 |
-!- uricamic [9320543b@gateway/web/freenode/ip.147.32.84.59] has quit [Quit: Page closed] | 17:40 | |
-!- puffin444 [230bf329@gateway/web/freenode/ip.35.11.243.41] has joined #shogun | 17:47 | |
puffin444 | Hi | 17:47 |
-!- harshit_ [~harshit@182.68.45.192] has quit [Ping timeout: 244 seconds] | 18:06 | |
-!- harshit_ [~harshit@182.68.45.192] has joined #shogun | 18:24 | |
-!- blackburn [5bde8018@gateway/web/freenode/ip.91.222.128.24] has joined #shogun | 18:32 | |
-!- PhilTillet [~Philippe@npasserelle10.minet.net] has joined #shogun | 18:33 | |
puffin444 | hello | 18:33 |
PhilTillet | hello | 18:34 |
puffin444 | Do you know when/if Sergey or Soeren will be on IRC? | 18:36 |
PhilTillet | I have no idea :p | 18:36 |
PhilTillet | I think they are eagerly waiting for the number of slots to be announced =D | 18:36 |
blackburn | puffin444: I'm partially available | 18:37 |
blackburn | I am first one ;) | 18:37 |
puffin444 | So blackburn it looks like I broke the build :( | 18:38 |
blackburn | I've seen the mail | 18:38 |
blackburn | first of S heh | 18:38 |
blackburn | let me check logs | 18:39 |
blackburn | puffin444: does it work on your machine? | 18:39 |
puffin444 | yes | 18:40 |
puffin444 | That's what I don't understand | 18:40 |
blackburn | any errors detected by valgrind? | 18:42 |
-!- PhilTillet [~Philippe@npasserelle10.minet.net] has quit [Ping timeout: 245 seconds] | 18:45 | |
puffin444 | There are no memory leaks. However, it does complain about freeing memory. This happens to all the examples though. | 18:46 |
puffin444 | In Valgrind | 18:46 |
blackburn | puffin444: do you copy the covariance matrix when you return it? | 18:47 |
puffin444 | Hmm... maybe not. | 18:48 |
blackburn | I mean is it a member? | 18:48 |
puffin444 | No. | 18:48 |
blackburn | ah no | 18:48 |
puffin444 | But the problem is that I can't reproduce this error on my system, so I don't really know how to fix it. | 18:49 |
blackburn | yeah | 18:49 |
blackburn | let me try to compile and test that | 18:49 |
blackburn | puffin444: actually the problem is some matrix is free'd two times | 18:51 |
puffin444 | Aah I see | 18:51 |
puffin444 | blackburn, I am extremely sorry to do this, but I need to leave the IRC immediately | 18:51 |
blackburn | puffin444: no problem | 18:51 |
puffin444 | I'll check and find this double-freed matrix. | 18:51 |
-!- puffin444 [230bf329@gateway/web/freenode/ip.35.11.243.41] has quit [Quit: Page closed] | 18:51 | |
blackburn | yeah and I'll try to reproduce | 18:52 |
-!- blackburn [5bde8018@gateway/web/freenode/ip.91.222.128.24] has quit [Ping timeout: 245 seconds] | 18:56 | |
-!- flxb [~cronor@fb.ml.tu-berlin.de] has joined #shogun | 18:57 | |
flxb | are kernels in shogun centered? | 18:58 |
-!- PhilTillet [~Philippe@157.159.42.154] has joined #shogun | 19:00 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 19:04 | |
flxb | or is it normalized in any way if I use custom kernels? | 19:07 |
-!- harshit_ [~harshit@182.68.45.192] has quit [Ping timeout: 260 seconds] | 19:11 | |
-!- harshit_ [~harshit@182.68.45.192] has joined #shogun | 19:17 | |
-!- blackburn [5bde8018@gateway/web/freenode/ip.91.222.128.24] has joined #shogun | 19:26 | |
blackburn | flxb: we have kernel normalizers for that | 19:27 |
-!- flxb [~cronor@fb.ml.tu-berlin.de] has quit [Quit: flxb] | 19:28 | |
wiking | nyapp | 19:39 |
-!- blackburn [5bde8018@gateway/web/freenode/ip.91.222.128.24] has quit [Ping timeout: 245 seconds] | 19:40 | |
-!- blackburn [5bde8018@gateway/web/freenode/ip.91.222.128.24] has joined #shogun | 19:51 | |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun | 19:52 | |
-!- harshit_ [~harshit@182.68.45.192] has quit [Ping timeout: 245 seconds] | 19:54 | |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has quit [Ping timeout: 276 seconds] | 19:58 | |
-!- PhilTillet [~Philippe@157.159.42.154] has quit [Ping timeout: 245 seconds] | 20:00 | |
-!- harshit_ [~harshit@182.68.45.192] has joined #shogun | 20:08 | |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has joined #shogun | 20:11 | |
wiking | blackburn: here? | 20:15 |
-!- blackburn [5bde8018@gateway/web/freenode/ip.91.222.128.24] has quit [Ping timeout: 245 seconds] | 20:18 | |
-!- harshit_ [~harshit@182.68.45.192] has quit [Ping timeout: 252 seconds] | 20:40 | |
-!- harshit_ [~harshit@182.68.45.192] has joined #shogun | 20:49 | |
-!- PhilTillet [~Philippe@157.159.42.154] has joined #shogun | 21:09 | |
@sonney2k | hey therre | 21:42 |
n4nd0 | hi | 21:42 |
@sonney2k | ohh the man with the fast covertree :) | 21:43 |
n4nd0 | haha not yet :) | 21:43 |
n4nd0 | I won't say until we can use it within shogun | 21:43 |
n4nd0 | ;) | 21:44 |
@sonney2k | hmmh does anyone know when the business day of pacific time is over | 21:45 |
@sonney2k | looks like in 5-6 hours but I guess at google they work a bit longer so lets say 8 hrs ;-) | 21:45 |
n4nd0 | I am around the channel in case I get something, I have seen you there too | 21:47 |
@sonney2k | no I will go to bed need to recover a bit | 21:47 |
n4nd0 | we'll know tomorrow then | 21:49 |
CIA-64 | shogun: Soeren Sonnenburg master * rc80508e / src/shogun/mathematics/Math.cpp : Fix compile error in twonorm() when lapack is not available - http://git.io/zvw5Qg | 21:50 |
* gsomix have returned from dance floor. | 21:51 | |
PhilTillet | hey sonney2k | 21:51 |
@sonney2k | hey | 21:51 |
wiking | oops sonney2k sorry about that bug :( | 21:52 |
@sonney2k | yeah luckily we have that buildbot (last year we didn't and that was really causing us quite some headache | 21:55 |
wiking | dooooh | 21:58 |
wiking | javac] symbol : constructor RealFeatures(org.jblas.DoubleMatrix) | 21:58 |
wiking | [javac] location: class org.shogun.RealFeatures | 21:58 |
wiking | how come? | 21:58 |
@sonney2k | got to leave train | 21:59 |
wiking | was working till now | 21:59 |
wiking | doh these days... | 21:59 |
@sonney2k | cu | 21:59 |
wiking | cya | 21:59 |
gsomix | good night | 22:14 |
@sonney2k | nite gsomix | 22:16 |
@sonney2k | wiking, what is the exact error? | 22:16 |
@sonney2k | seems a little unspecific... | 22:16 |
n4nd0 | good night gsomix | 22:16 |
wiking | sonney2k: i think i had some trouble with detecting jblas | 22:16 |
wiking | now i'm recompiling and i guess it'll work now | 22:17 |
@sonney2k | not in CLASSPATH ? | 22:17 |
@sonney2k | hmmhh | 22:17 |
wiking | it was actually | 22:17 |
@sonney2k | wiking, are you using java interface of shogun? | 22:17 |
wiking | i mean the problem was with | 22:17 |
wiking | that javac was defined to be gcj | 22:18 |
wiking | and that had some troubles detecting the jblas (it's at default location) | 22:18 |
@sonney2k | hmmhh maybe | 22:18 |
wiking | but now trying with sun's javac | 22:18 |
wiking | ok now it's fixed | 22:18 |
wiking | only thing now is | 22:18 |
wiking | no that is my error | 22:19 |
wiking | ok now compiled fine | 22:19 |
wiking | yeah i'm using sometimes java interface | 22:19 |
@sonney2k | wiking, so are you using the java interace? | 22:19 |
@sonney2k | ahh ok | 22:19 |
wiking | yep yep | 22:19 |
wiking | need to | 22:19 |
wiking | i have some features coming from hadoop | 22:19 |
@sonney2k | I haven't used it except for crafting some examples | 22:20 |
@sonney2k | so all feedback welcome | 22:20 |
wiking | and i don't want to do constantly a transforming the features into some other formats | 22:20 |
@sonney2k | wiking, btw you should be even able to serialize a shogun object from java and load it later from python :) | 22:20 |
wiking | so in order to be able to simply use hadoop created sequence file i just use java interface | 22:20 |
wiking | let's see multi class linear machine | 22:21 |
wiking | with homo kernel mapping :)) | 22:21 |
@sonney2k | heh | 22:23 |
wiking | mm i wonder what's happening | 22:26 |
wiking | still calculating :))) | 22:26 |
wiking | i thought it's a fast linear machine | 22:27 |
wiking | can it be that this linear machine is slower than a kernel machine with linear kernel? | 22:31 |
@sonney2k | wiking, liblinear? | 22:36 |
@sonney2k | no should be very fast | 22:37 |
-!- PhilTillet [~Philippe@157.159.42.154] has quit [Ping timeout: 252 seconds] | 22:37 | |
wiking | mmm | 22:43 |
wiking | wtf | 22:43 |
wiking | Exception in thread "main" java.lang.NullPointerException | 22:43 |
wiking | at bme.iclef.shogun.Classify.classify(Classify.java:412) | 22:43 |
wiking | :)))) | 22:43 |
wiking | ahhahaha | 22:43 |
wiking | Exception in thread "main" java.lang.NullPointerException | 22:43 |
wiking | at bme.iclef.shogun.Classify.classify(Classify.java:412) | 22:43 |
wiking | mmm not htis | 22:43 |
wiking | ok again my bug | 22:44 |
wiking | ok i'm gonna time this one now :) | 22:50 |
n4nd0 | time to sleep today, good night guys! | 22:53 |
wiking | gnite | 22:53 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 22:53 | |
-!- puffin444 [230bf329@gateway/web/freenode/ip.35.11.243.41] has joined #shogun | 23:36 | |
-!- puffin444 [230bf329@gateway/web/freenode/ip.35.11.243.41] has quit [Client Quit] | 23:36 | |
-!- muddo [~muddo@gateway/tor-sasl/muddo] has left #shogun ["Leaving"] | 23:51 | |
-!- blackburn [~qdrgsm@83.234.54.186] has joined #shogun | 23:58 | |
--- Log closed Thu Apr 12 00:00:19 2012 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!