--- Log opened Tue Jul 10 00:00:17 2012 | ||
@sonney2k | blackburn, mgiht be - I presented at europython and chris at scipy... | 00:22 |
---|---|---|
blackburn | sonney2k: will try to send a proposal then may be.. | 00:22 |
CIA-18 | shogun: Evgeniy Andreev master * red3c4b6 / src/shogun/lib/Map.h : added support for threads - http://git.io/dIx6tA | 00:23 |
CIA-18 | shogun: Soeren Sonnenburg master * re10bc96 / src/shogun/lib/Map.h : Merge pull request #633 from gsomix/trace_mallocs - http://git.io/Gst4cA | 00:23 |
blackburn | Sep, 28 is due | 00:23 |
@sonney2k | ...lots of time | 00:24 |
blackburn | sonney2k: shouldn't we provide not C1 and C2 but C selection too? | 00:33 |
@sonney2k | blackburn, I really have to fixup CTime - the timings gsomix sent where not helpful | 00:33 |
@sonney2k | blackburn, I wanted to do it differently - have an SGVector for C | 00:33 |
@sonney2k | and if that vector has length 1 -> single C | 00:33 |
blackburn | cool | 00:33 |
blackburn | got it | 00:33 |
blackburn | 2 C1 C2 | 00:33 |
@sonney2k | length 2 -> -1 / +1 | 00:33 |
@sonney2k | and otherwise full | 00:34 |
blackburn | N different Cs | 00:34 |
blackburn | for each example | 00:34 |
@sonney2k | but didn't find the time | 00:34 |
@sonney2k | and we didn't have modsel with vectors back tehn | 00:34 |
@sonney2k | but puffin did it just recently ... so | 00:34 |
blackburn | we should ask heiko probably | 00:34 |
blackburn | argh I press tab to complete words sometimes | 00:35 |
blackburn | :D | 00:35 |
blackburn | and ctrl + n as well | 00:35 |
blackburn | sonney2k: ah one more thing | 00:36 |
blackburn | we should support nested xval probably | 00:36 |
@sonney2k | blackburn, problem with the modsel framework is that we need a 1:1 mapping between variables and parameters | 00:39 |
@sonney2k | if we want some *scalar* C | 00:39 |
@sonney2k | but want to support Cn / Cp | 00:39 |
@sonney2k | and a vector | 00:39 |
@sonney2k | the most general formulation is a vector - so we probably need to store it that way | 00:40 |
@sonney2k | it is (from the user perspective) stupid/weird having to write sth like [numpy.array([1.0])] for some C | 00:40 |
@sonney2k | but anyways | 00:41 |
blackburn | but it is bad to force user | 00:41 |
@sonney2k | blackburn, nested x-val yes - I think only heiko and puffin are deep enough into that code currently | 00:41 |
@sonney2k | blackburn, ? | 00:41 |
blackburn | sonney2k: I am getting too | 00:41 |
@sonney2k | k | 00:41 |
blackburn | to force user to search | 00:41 |
blackburn | over | 00:41 |
blackburn | C1 and C2 | 00:42 |
@sonney2k | sure | 00:42 |
@sonney2k | most common case is just C | 00:42 |
@sonney2k | search | 00:42 |
blackburn | sonney2k: how should logistic regression output look like | 00:42 |
blackburn | ? | 00:42 |
@sonney2k | BinaryLabels ? | 00:42 |
blackburn | no I mean probability score | 00:43 |
@sonney2k | ??? | 00:43 |
blackburn | 1/(1+exp(-sign(out)*out)) | 00:43 |
blackburn | this way? | 00:43 |
blackburn | damn | 00:44 |
blackburn | I am confused with it | 00:44 |
blackburn | sonney2k: I need probability outputs based on given <w,x>+b | 00:44 |
@sonney2k | I don't know out of the top over my head | 00:44 |
@sonney2k | google for it | 00:44 |
@sonney2k | blackburn, ahh not logistic regression | 00:44 |
@sonney2k | just sigmoid you mean? | 00:44 |
blackburn | probably | 00:44 |
@sonney2k | https://en.wikipedia.org/wiki/Sigmoid_function | 00:45 |
@sonney2k | heh | 00:45 |
blackburn | I understand | 00:45 |
blackburn | but | 00:45 |
blackburn | to support probabilities in outputs | 00:45 |
blackburn | hmm | 00:45 |
blackburn | I think I should subtract 0.5 and multiply by 2? | 00:46 |
@sonney2k | https://www.cs.cornell.edu/courses/cs678/2007sp/platt.pdf | 00:47 |
@sonney2k | I think that is what you should read | 00:47 |
@sonney2k | should be platt's original method | 00:47 |
blackburn | thanks | 00:47 |
@sonney2k | there is some paper from platt about it | 00:51 |
blackburn | sonney2k: but I do not need fitting | 00:51 |
blackburn | I fitted it with logistic regression - but outputs | 00:51 |
@sonney2k | http://www.csie.ntu.edu.tw/~htlin/paper/doc/plattprob.pdf | 00:51 |
@sonney2k | blackburn, but then you already have probabilities | 00:52 |
blackburn | <w,x>+b? | 00:52 |
wiking | mmm | 00:53 |
wiking | actually an issue: sgvector and sgsparsevector only have referencedata as common parent. thus if i want to support both (i.e. densefeatures and sparsefeatures) in my implementation i must define the vector format as a template. were there any plans for having SGVector as an abstract class and SGDenseVector and SGSparseVector as derived classes? | 00:55 |
@sonney2k | wiking, rather use the concept of dotfeatures! | 00:56 |
wiking | sonney2k: the problem is that my current api looks like this for example: SGVector<float64_t> get_psi(...) | 00:57 |
@sonney2k | blackburn, ok but then https://en.wikipedia.org/wiki/Logistic_regression | 00:57 |
blackburn | sonney2k: do you think I am banned by google? :D | 00:58 |
@sonney2k | blackburn, otherwise would you be asking :D | 00:58 |
@sonney2k | wiking, not a good idea | 00:58 |
@sonney2k | wiking, use the dot features API | 00:58 |
@sonney2k | it will allow you to even have strings or whatever as features | 00:58 |
@sonney2k | and mixtures of sparse dense, string features | 00:59 |
wiking | sonney2k: but that assumes that i can do a dot product somehow over those features... | 00:59 |
wiking | i mean actually among psi vectors i could | 00:59 |
@sonney2k | wiking, look carefully which operations you need | 00:59 |
@sonney2k | wiking, I suspect that dotfeatures have them all | 00:59 |
wiking | the main problem now that get_psi returns a simple vector | 01:00 |
@sonney2k | (maybe it needs some latent variable in there ...) | 01:00 |
@sonney2k | wiking, free your mind from *explicitly* getting this vector | 01:00 |
@sonney2k | this is usually not needed | 01:00 |
@sonney2k | wiking, actually no svm needs it | 01:00 |
@sonney2k | no solver I mean | 01:00 |
@sonney2k | neither liblinear nor newtonsvm nor sgd nor ... | 01:01 |
* sonney2k ZZZzzzz... | 01:02 | |
wiking | mmm | 01:02 |
wiking | that still doesn't help me | 01:02 |
wiking | but ok | 01:02 |
wiking | let's talk about this tomorrow? | 01:02 |
wiking | or blackburn up for some consulting? :)) | 01:02 |
blackburn | I am pretty stupid today | 01:02 |
wiking | no worries | 01:02 |
blackburn | or all the way | 01:02 |
blackburn | no idea | 01:02 |
blackburn | :D | 01:02 |
blackburn | okay so what is it? | 01:03 |
wiking | so the current way i generate feature vectors is that i have an api call: SGVector<float64_t> get_psi(index_t index_for_a_feature) | 01:03 |
wiking | basically it does the following z = PSI(x,h) | 01:04 |
wiking | where x is the input features and h is the latent variable | 01:04 |
wiking | the problem is with the current way that i force that z can only be a dense vector | 01:04 |
wiking | but actually it would be good to have support for z being a sparse vector | 01:05 |
blackburn | wiking: what do you do with z later? | 01:05 |
wiking | blackburn: feed it to a solver | 01:05 |
blackburn | what does solver do with z? | 01:06 |
wiking | blackburn: solves a minimization problem | 01:07 |
blackburn | what is the solver btw/ | 01:07 |
wiking | currently svmocas | 01:07 |
wiking | of course what i could do is | 01:09 |
wiking | change the api | 01:09 |
wiking | that it returns the whole feature set | 01:09 |
wiking | at once... as CDotFeatures | 01:09 |
blackburn | I mean what can you do | 01:10 |
blackburn | is to provide operations involving z | 01:10 |
blackburn | that's COFFIN spirit :D | 01:10 |
wiking | so they are calculated on the fly you mean? | 01:15 |
blackburn | probably you don't have to have z explicitly | 01:15 |
blackburn | but you only need some operations like | 01:16 |
blackburn | z+x | 01:16 |
blackburn | or z*x | 01:16 |
blackburn | so if you provide it you may specialize to sparse or any other features | 01:16 |
wiking | mmm | 01:17 |
* wiking feels stupid now :D | 01:17 | |
blackburn | wiking: if you need to add z to x | 01:17 |
blackburn | you don't have to know z explicitly | 01:17 |
blackburn | you may just call some virtual function | 01:17 |
blackburn | that's how it works in shogun with sparse and dense | 01:18 |
blackburn | they just provide different routines for doing the same thing | 01:18 |
wiking | btw: loool? http://theoatmeal.com/blog/charity_money | 01:22 |
wiking | mmm ok | 01:23 |
wiking | well let's see what i can come up with then | 01:23 |
blackburn | hmm doesn't load | 01:23 |
wiking | blackburn: guess it was just suddenly slashdoted or something | 01:24 |
wiking | noup.... worksforme (still) | 01:24 |
blackburn | works | 01:24 |
blackburn | loaded | 01:24 |
blackburn | woah | 01:24 |
wiking | :D | 01:24 |
blackburn | a lot of money | 01:24 |
n4nd0 | ok, time to switch off | 01:32 |
n4nd0 | good night guys! | 01:32 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 01:32 | |
-!- blackburn [~blackburn@81.28.187.0] has left #shogun [] | 02:22 | |
shogun-buildbot | build #17 of nightly_none is complete: Failure [failed compile] Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_none/builds/17 | 03:01 |
-!- zxtx [~zv@adsl-75-50-156-109.dsl.lsan03.sbcglobal.net] has joined #shogun | 03:03 | |
-!- zxtx [~zv@adsl-75-50-156-109.dsl.lsan03.sbcglobal.net] has quit [Ping timeout: 255 seconds] | 03:54 | |
-!- av3ngr [av3ngr@nat/redhat/x-trmcrnwzrrwvhlmy] has joined #shogun | 04:22 | |
-!- av3ngr [av3ngr@nat/redhat/x-trmcrnwzrrwvhlmy] has left #shogun [] | 04:22 | |
-!- puffin444 [62e3926e@gateway/web/freenode/ip.98.227.146.110] has joined #shogun | 06:01 | |
-!- gsomix [~gsomix@88.200.242.178] has joined #shogun | 07:30 | |
@sonney2k | puffin444, one question - if I understood correctly you wrote vector / matrix support for model selection? | 07:42 |
puffin444 | Yes. But I found another way to calculate sum gradients, so it's not used. | 07:43 |
puffin444 | but the code is still in there | 07:43 |
@sonney2k | puffin444, yeah but we need it so it is good that you wrote it | 07:53 |
@sonney2k | puffin444, (for setting svm-C as vector not just scalar) | 07:53 |
puffin444 | Did someone change the CMap code today? | 07:53 |
puffin444 | okay | 07:53 |
@sonney2k | puffin444, yes | 07:53 |
@sonney2k | puffin444, why? | 07:53 |
puffin444 | It seems to put my code in infinite recursion. It may be my fault though. | 07:53 |
@sonney2k | gsomix is also awake so if you have a backtrace... | 07:54 |
gsomix | good morning | 08:03 |
@sonney2k | morning gsomix | 08:05 |
@sonney2k | gsomix, please continue with the python buffer stuff | 08:05 |
@sonney2k | gsomix, I will fix up CTime and then run the benchmark again | 08:05 |
gsomix | sonney2k, I really hate my implementation of map. :) I think I should rewrite it far far later on trees as in STL. | 08:06 |
@sonney2k | gsomix, nothing to hate there | 08:08 |
@sonney2k | hashing is still the fastest one can do and thread safety never was a requirement | 08:09 |
-!- pluskid [~pluskid@111.120.75.182] has joined #shogun | 08:42 | |
gsomix | Wikipedia in Russian will be closed on 10th of July because of Russian parliament’s debating of amendments to the law «About information» that could create real censorship of the internet — a blacklisting and filtering of internet sites. | 08:54 |
gsomix | fine | 08:54 |
@sonney2k | gsomix, each contry the same sh!t | 09:05 |
@sonney2k | gsomix, we have had a law passed that basically allows companies to access all your data on request - and guess when it was passed? during the semi-final when germany was playing against italy | 09:06 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 09:46 | |
-!- uricamic [~uricamic@2001:718:2:1634:ed46:10d7:8a90:f39b] has joined #shogun | 09:59 | |
-!- pluskid [~pluskid@111.120.75.182] has quit [Ping timeout: 264 seconds] | 10:03 | |
-!- pluskid [~pluskid@202.130.113.141] has joined #shogun | 10:03 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Ping timeout: 244 seconds] | 10:10 | |
-!- puffin444 [62e3926e@gateway/web/freenode/ip.98.227.146.110] has quit [Ping timeout: 245 seconds] | 10:13 | |
-!- uricamic [~uricamic@2001:718:2:1634:ed46:10d7:8a90:f39b] has quit [Ping timeout: 245 seconds] | 10:46 | |
-!- uricamic [~uricamic@cmpgw-27.felk.cvut.cz] has joined #shogun | 10:47 | |
-!- pluskid [~pluskid@202.130.113.141] has quit [Ping timeout: 255 seconds] | 10:47 | |
-!- pluskid [~pluskid@202.130.113.141] has joined #shogun | 11:10 | |
-!- heiko [~heiko@host86-181-154-192.range86-181.btcentralplus.com] has joined #shogun | 11:11 | |
CIA-18 | shogun: Heiko Strathmann master * r45cb8f0 / src/shogun/mathematics/Statistics.cpp : fixed error - http://git.io/KWBUJQ | 11:19 |
CIA-18 | shogun: Heiko Strathmann master * rab7ddc2 / (2 files): optimization only possible when lapack is there - http://git.io/ryW7jQ | 11:19 |
CIA-18 | shogun: Heiko Strathmann master * rfa071cb / (3 files in 2 dirs): Merge pull request #635 from karlnapf/master - http://git.io/FovxPA | 11:19 |
-!- heiko [~heiko@host86-181-154-192.range86-181.btcentralplus.com] has quit [Ping timeout: 245 seconds] | 11:29 | |
-!- foo___ [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has joined #shogun | 11:29 | |
foo___ | hi all | 11:40 |
@sonney2k | foo___, bar! | 11:44 |
foo___ | hello again | 11:52 |
foo___ | I pull last update this morning and I got smthg strange with last modification | 11:52 |
foo___ | I tried to rerun basic multiclass examples and program dont stop anymore | 11:53 |
foo___ | I used multiclass c++ example from shogun | 11:53 |
-!- gsomix_ [~gsomix@80.234.29.105] has joined #shogun | 11:53 | |
-!- gsomix [~gsomix@88.200.242.178] has quit [Read error: Connection reset by peer] | 11:53 | |
foo___ | and just replace multiclassliblinear with CGMNPSVM and linear kernel | 11:54 |
foo___ | you can reproduce the bug easily | 11:54 |
foo___ | up: the example I used is "evaluation_cross_validation_multiclass.cpp" and I just replace multiclassliblinear with CGMNPSVM and linear kernel | 12:00 |
-!- heiko [~heiko@host86-181-154-192.range86-181.btcentralplus.com] has joined #shogun | 12:08 | |
foo___ | hi again, lets resume : with last shogun commit, I got infinite loop. You can reproduce the error using "evaluation_cross_validation_multiclass.cpp" and replacing CMulticlassliblinear machine with CGMNPSVM and linear kernel. | 12:11 |
foo___ | It would be nice to know if anyne else got the same issue. Thanks. | 12:11 |
-!- heiko [~heiko@host86-181-154-192.range86-181.btcentralplus.com] has quit [Ping timeout: 245 seconds] | 12:12 | |
-!- blackburn [~blackburn@81.28.187.0] has joined #shogun | 12:41 | |
-!- pluskid [~pluskid@202.130.113.141] has quit [Quit: Leaving] | 13:37 | |
wiking | foo___ letmecheck for you | 13:39 |
wiking | foo___: it ran and stopped for me! | 13:40 |
wiking | but then again this is not the latest commit | 13:40 |
wiking | let me rebase and check it again for you | 13:40 |
-!- zxtx [~zv@adsl-75-50-156-109.dsl.lsan03.sbcglobal.net] has joined #shogun | 13:45 | |
foo___ | wiking: thx a lot | 13:46 |
wiking | foo___: ok the basic example runs | 13:47 |
wiking | let me check for changing it to GMNPSVM | 13:47 |
foo___ | wiking: Multiclasslibsvm stop also | 13:47 |
foo___ | wiking: crossvalid seems not working well with other machine than multiclassliblinear | 13:48 |
wiking | foo___: just a sec i'm trying now... :) | 13:50 |
wiking | foo___: u've tried with linear kernel right? | 13:50 |
foo___ | wiking: yes | 13:52 |
foo___ | wiking: CLinearKernel* ker = new CLinearKernel(features,features); CMulticlassLibSVM* svm=new CMulticlassLibSVM(svm_C, ker, labels); | 13:53 |
wiking | foo___: ok | 13:58 |
wiking | same here | 13:58 |
foo___ | arf | 13:59 |
wiking | but afaik some time ago we already had problems with multiclass cv | 13:59 |
wiking | gsomix_: have you been working on CMap lately? | 14:01 |
gsomix_ | wiking, yes | 14:02 |
gsomix_ | problems? :( | 14:03 |
foo___ | wiking: what's CMap ? | 14:03 |
wiking | gsomix_: i dont know yet... i just see from gdb that it is called all the time | 14:03 |
wiking | so now i'm wondering what could be the problem | 14:03 |
wiking | foo___: well it is storing some stuff while you do crossvalidation | 14:03 |
wiking | so i'm just trying to figure out what's causing this loop | 14:04 |
gsomix_ | wiking, keep me up to date. I 'm ready to fix it. .____. | 14:07 |
wiking | gsomix_: heheh first i need to see why it's really happening | 14:07 |
wiking | oh yeah | 14:09 |
wiking | it seems we have an infinte loop within CKernelMulticlassMachine::store_model_features | 14:09 |
blackburn | infinite loop? | 14:12 |
blackburn | cool | 14:12 |
wiking | blackburn: yeah | 14:12 |
wiking | never gets out of that func call apparently | 14:13 |
wiking | i need -g flag for the compiler :) | 14:13 |
blackburn | hmm | 14:13 |
blackburn | lets try something | 14:14 |
blackburn | do any of you have reproducing example? | 14:14 |
blackburn | wiking: which method is called infinitely? | 14:15 |
foo___ | blackburn: just use "evaluation_cross_validation_multiclass.cpp" and replace CMulticlassliblinear machine with CGMNPSVM and linear kernel. | 14:15 |
wiking | blackburn: just trying now | 14:15 |
wiking | ok | 14:16 |
wiking | found it | 14:16 |
blackburn | wiking: found what? | 14:16 |
wiking | where it gets stuck | 14:16 |
blackburn | okay nice | 14:16 |
wiking | m_machines->get_num_elements() | 14:16 |
wiking | fuck | 14:16 |
wiking | SGVector<int32_t> sv_idx(all_sv.get_num_elements()); | 14:17 |
wiking | for (index_t i=0; i < sv_idx.vlen; ++i) | 14:17 |
wiking | sv_idx[i] = all_sv.get_element(i); | 14:17 |
wiking | so here | 14:17 |
wiking | this loop never stops :) | 14:17 |
wiking | it puzzles me why | 14:17 |
wiking | coz in the current example i'm trying sv_idx.vlen = 34 | 14:17 |
blackburn | interesting | 14:18 |
wiking | ok | 14:18 |
wiking | i see | 14:18 |
wiking | all_sv.get_element(i); | 14:19 |
wiking | never returns | 14:19 |
blackburn | gsomix_: ! | 14:19 |
blackburn | :) | 14:19 |
wiking | heheh yep it's gsomix_ | 14:19 |
gsomix_ | fck | 14:19 |
wiking | dont' we want a configure flag that adds -g flag for gcc? | 14:19 |
wiking | like --debug | 14:19 |
wiking | or seomthing | 14:19 |
wiking | no shit | 14:20 |
wiking | it's there | 14:20 |
blackburn | wiking: --enable-debug? | 14:20 |
wiking | blackburn: no apparently the -g flag is actually there | 14:20 |
wiking | then why does gdb still bitching | 14:20 |
wiking | ... | 14:20 |
blackburn | even without debug? | 14:20 |
wiking | adsf!@# | 14:20 |
wiking | i must have something wrong on my side | 14:20 |
wiking | gsomix_: void CKernelMulticlassMachine::store_model_features() | 14:21 |
wiking | in machine/KernelMulticlassMachine.cpp | 14:21 |
wiking | basically the second for loop will never stop running as all_sv.get_element(i); where CMap<int32_t, int32_t> all_sv; | 14:21 |
wiking | get_element never returns | 14:21 |
gsomix_ | hah, lol | 14:22 |
gsomix_ | my fault | 14:22 |
gsomix_ | just a moment | 14:22 |
wiking | int32_t added_index=add(key, T()); | 14:23 |
wiking | result=get_node_ptr(added_index); | 14:23 |
wiking | soemwhere here it dies... | 14:24 |
gsomix_ | >> get_element never returns | 14:24 |
gsomix_ | fixed | 14:24 |
wiking | Map.h:197-198 | 14:24 |
wiking | ok | 14:24 |
wiking | somebody pull it | 14:24 |
wiking | :) | 14:24 |
blackburn | yeah yeah | 14:24 |
CIA-18 | shogun: Sergey Lisitsyn master * rdceddcf / src/shogun/classifier/mkl/MKL.cpp : Fixed MKL training double free - http://git.io/kqUgcg | 14:24 |
wiking | heheh no that one :D | 14:25 |
wiking | *not | 14:25 |
gsomix_ | one moment | 14:25 |
wiking | ah ok | 14:25 |
gsomix_ | need to check | 14:25 |
wiking | where's you repo?:) | 14:25 |
wiking | gsomix_: u can send me the fix | 14:26 |
wiking | before the commit | 14:26 |
wiking | or just write paste here the diff | 14:27 |
wiking | and i'll do the changes by hand | 14:27 |
gsomix_ | blackburn, pr, pr | 14:28 |
CIA-18 | shogun: Evgeniy Andreev master * r6aae77b / src/shogun/lib/Map.h : fixed some bugs - http://git.io/wQTjww | 14:28 |
CIA-18 | shogun: Sergey Lisitsyn master * r731fafe / src/shogun/lib/Map.h : Merge pull request #636 from gsomix/trace_mallocs - http://git.io/OezbTw | 14:28 |
gsomix_ | do not laugh, hehe :( | 14:28 |
blackburn | here | 14:29 |
gsomix_ | wiking, done | 14:29 |
wiking | ok let me check | 14:29 |
wiking | compiling... | 14:30 |
blackburn | is there any other hot bugs? | 14:31 |
CIA-18 | shogun: Sergey Lisitsyn master * r80adbb4 / src/shogun/modelselection/ModelSelectionParameters.h : Fixed num steps computation of model selection - http://git.io/208CGQ | 14:32 |
wiking | foo___: ok | 14:34 |
wiking | foo___: update your code to the latest... | 14:34 |
CIA-18 | shogun: Sergey Lisitsyn master * r2aba715 / (3 files in 2 dirs): Made feature block logistic regression outputs use sigmoid - http://git.io/1mWTIw | 14:34 |
-!- heiko [~heiko@host86-174-150-246.range86-174.btcentralplus.com] has joined #shogun | 14:34 | |
wiking | it works | 14:34 |
wiking | but the results are shocking :DDD | 14:34 |
blackburn | wiking: shocking like? | 14:35 |
wiking | ok wait | 14:35 |
wiking | maybe i'm misinterpreting something | 14:35 |
foo___ | wiking: compiling too. | 14:35 |
foo___ | wiking: you said smthg wrong with the xval ? | 14:36 |
wiking | [DEBUG] result of cross-validation run 9 is 0.271930 | 14:36 |
blackburn | heiko: build values was wrong :) | 14:36 |
blackburn | wiking: what is evaluation? | 14:36 |
wiking | is that loss? | 14:36 |
blackburn | wiking: it is result of evaluation you do use | 14:36 |
wiking | [DEBUG] result of cross-validation run 9 is 0.271930 | 14:36 |
wiking | fuck | 14:36 |
heiko | blackburn, what exactly in there? | 14:36 |
blackburn | wiking: accuracy | 14:36 |
wiking | 105 CMulticlassAccuracy* eval_crit = new CMulticlassAccuracy (); | 14:36 |
wiking | yes | 14:36 |
blackburn | heiko: num steps - I fixed | 14:37 |
heiko | blackburn, good that finally somebody checks all this :) | 14:37 |
blackburn | heiko: 0.0,0.5 with step 0.1 produced 0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0 before | 14:37 |
heiko | it worked for my examples but I already has the suspicion that there were some bugs in there | 14:37 |
heiko | oh ok | 14:37 |
blackburn | heiko: yeah time for xval to work | 14:37 |
wiking | [DEBUG] result of cross-validation run 9 is 0.826316 | 14:38 |
wiking | so this is the problem | 14:38 |
blackburn | heiko: I'd like to add nested xval support | 14:38 |
heiko | that is? | 14:38 |
wiking | blackburn: CGMNPSVM with linear kernel (0.271930) vs CMulticlassLibLinear (0.826316) | 14:39 |
foo___ | wiking: yes for me too | 14:39 |
blackburn | wiking: aham | 14:39 |
blackburn | heiko: sorry what is the question? :) | 14:39 |
wiking | foo___: so i guess there's still a bug somewhere | 14:39 |
heiko | what is nested xval? what do you mean by that? | 14:39 |
blackburn | heiko: that one described here http://stackoverflow.com/questions/2595176/when-to-choose-which-machine-learning-classifier | 14:40 |
blackburn | wiking: GMNP is ok I am sure | 14:41 |
foo___ | wiking: indices of chosen train/testing data seems good .. | 14:41 |
blackburn | I'd say something is wrong with normalization may be | 14:41 |
wiking | blackburn: i guess then there's something wrong with KernelMachine....? | 14:42 |
wiking | as the difference between the 2 accuracy is really huge | 14:42 |
foo___ | blackburn: why do you need several shuffle fction in xval ? | 14:42 |
heiko | blackburn, mmmh, I dont see the advantage of this approach over normal xval?, what is it? | 14:42 |
wiking | :S | 14:42 |
blackburn | heiko: I am not the expert here unfortunately | 14:42 |
blackburn | but ckwidmer told me it is a common way | 14:42 |
heiko | kk | 14:43 |
blackburn | wiking: try larank please | 14:45 |
wiking | blackburn: segfault :D | 14:48 |
blackburn | nice | 14:48 |
wiking | blackburn: larank_kcache_query_row (self=0x0, i=0, len=18) | 14:49 |
blackburn | more | 14:49 |
blackburn | self=NULL? | 14:49 |
wiking | dunno why | 14:50 |
wiking | that supposed to be the kernel :) | 14:50 |
blackburn | wiking: I am curious why python works | 14:59 |
wiking | blackburn: :DDDDD | 14:59 |
wiking | really? :)))) | 14:59 |
blackburn | wiking: yes python larank works smoothly | 14:59 |
wiking | and what's the multiclass accu for cv? | 15:00 |
wiking | btw: larank it self worksforme but only in this setup with mc+cv it fails... | 15:01 |
blackburn | ahh | 15:02 |
wiking | mmm | 15:03 |
wiking | i think i might see the problem | 15:03 |
blackburn | good | 15:03 |
wiking | i guess in the meanwhile | 15:03 |
wiking | the kernel gets freed | 15:03 |
wiking | so i guess we need an SG_REF on the kernel in | 15:03 |
wiking | void LaRankOutput::initialize (CKernel* kfunc, int64_t cache) | 15:03 |
blackburn | wiking: fix it then ;) | 15:04 |
wiking | trying | 15:04 |
wiking | it was just a hunch | 15:05 |
wiking | noup | 15:05 |
wiking | doh | 15:08 |
wiking | well the whole LaRankOutput gets destroyed and then it gets called by train_machine -> null pointer exception | 15:12 |
wiking | problem is that i have no idea about larank | 15:12 |
blackburn | wiking: okay nevermind | 15:13 |
blackburn | :) | 15:13 |
blackburn | I will check a little later | 15:13 |
-!- heiko1 [~heiko@host86-180-42-236.range86-180.btcentralplus.com] has joined #shogun | 15:18 | |
-!- heiko [~heiko@host86-174-150-246.range86-174.btcentralplus.com] has quit [Ping timeout: 245 seconds] | 15:19 | |
-!- blackburn [~blackburn@81.28.187.0] has quit [Remote host closed the connection] | 15:28 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 15:28 | |
foo___ | wiking: I get others false result using mc cv. have indentify what the pb is ? | 15:29 |
wiking | foo___: hehehe not really | 15:35 |
wiking | foo___: but there's definitely something wrong with cv for kernel machines | 15:35 |
-!- alexlovesdata_ [c25faeba@gateway/web/freenode/ip.194.95.174.186] has joined #shogun | 15:38 | |
foo___ | wiking: ok it reassures me a liitle bit .. =) Do you know if som1 is going to investigate this in the near future ? ^^ | 15:42 |
foo___ | wiking: thks again for checking this bug from the beginning. | 15:43 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Read error: Operation timed out] | 15:49 | |
-!- gsomix_ [~gsomix@80.234.29.105] has quit [Remote host closed the connection] | 15:59 | |
-!- foo___ [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has quit [Quit: Page closed] | 16:21 | |
-!- blackburn [~blackburn@81.28.187.0] has joined #shogun | 16:43 | |
-!- blackburn [~blackburn@81.28.187.0] has quit [Quit: Leaving.] | 16:51 | |
-!- foo__ [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has joined #shogun | 17:30 | |
-!- puffin444 [62e3926e@gateway/web/freenode/ip.98.227.146.110] has joined #shogun | 17:54 | |
-!- zxtx [~zv@adsl-75-50-156-109.dsl.lsan03.sbcglobal.net] has quit [Ping timeout: 248 seconds] | 17:57 | |
-!- uricamic [~uricamic@cmpgw-27.felk.cvut.cz] has quit [Quit: Leaving.] | 18:10 | |
foo__ | hi again | 18:35 |
wiking | yo | 18:36 |
foo__ | CMulticlasslabels needs sgvector<double> in constructor and MKLMulticlass: Multiclass Labels must be integers (sgcector<int>). | 18:37 |
foo__ | whats the right way to do that, I dont get it | 18:37 |
foo__ | perhaps we need mclabels constructor with sgvector<int> ? | 18:39 |
foo__ | ok my mistake, it seems that my label 8 was approx to 7.82676 .. | 18:43 |
foo__ | sorry for that | 18:43 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 18:46 | |
-!- zxtx [~zv@cpe-75-83-151-252.socal.res.rr.com] has joined #shogun | 18:51 | |
CIA-18 | shogun: Heiko Strathmann master * re1b4a40 / (2 files): added optimization of combined kernel weights - http://git.io/VfNKfA | 18:56 |
CIA-18 | shogun: Heiko Strathmann master * rb8b8336 / (2 files): added example for combined kernel weights optimization - http://git.io/U02UvQ | 18:56 |
CIA-18 | shogun: Heiko Strathmann master * r90c4bca / (4 files in 2 dirs): Merge pull request #637 from karlnapf/master - http://git.io/ipBdEg | 18:56 |
n4nd0 | heiko1: hey! how is it going with the opt. problem? | 18:59 |
-!- foo__ [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has quit [Quit: Page closed] | 19:00 | |
heiko1 | n4nd0, hej! | 19:08 |
heiko1 | its already committed :) | 19:09 |
heiko1 | although I just found a strange inconsistency with matlab | 19:09 |
heiko1 | but thats a minor thing | 19:09 |
n4nd0 | good | 19:09 |
n4nd0 | heiko1: do you happen to know some Swedish? I am just wondering because you said hej ;) | 19:10 |
heiko1 | ehm. | 19:10 |
heiko1 | no :) | 19:10 |
heiko1 | not a single word ;) | 19:10 |
heiko1 | I somehow got into the habit of writing that | 19:10 |
heiko1 | (you can do that in German too) | 19:10 |
n4nd0 | aaah, ok :) | 19:11 |
n4nd0 | well you know at least a word then, hej :D | 19:11 |
heiko1 | yeah :) | 19:11 |
heiko1 | How are things going at yours? | 19:12 |
n4nd0 | fine | 19:12 |
n4nd0 | I am reading some code now | 19:12 |
n4nd0 | I have to understand another implementation of HM-SVM in order to compare my results with theirs | 19:12 |
heiko1 | ah, nice :) | 19:12 |
heiko1 | are you also using the libqp framework btw? | 19:12 |
n4nd0 | no, I couldn't use it because they do not cover the constraints I needed for my problem | 19:13 |
n4nd0 | I think uricamic used it though | 19:13 |
heiko1 | theres one thing I wonder: Why does one have to specify a feasible starting point | 19:14 |
n4nd0 | in libqp? | 19:15 |
heiko1 | yes | 19:15 |
heiko1 | I use the constraints a' * x = 1 | 19:15 |
heiko1 | so its easy | 19:15 |
heiko1 | but I wonder why it isnt dont automatically, and if I am missing something | 19:16 |
n4nd0 | I have no idea why | 19:16 |
n4nd0 | do they give a reference to the opt. method they use? | 19:16 |
heiko1 | yes, good idea Ill check that | 19:16 |
heiko1 | n4nd0, Ill leave now, got a headache and hav eto turn off the computer for some time, it has been on all day ;) | 19:18 |
heiko1 | bye | 19:18 |
-!- heiko1 [~heiko@host86-180-42-236.range86-180.btcentralplus.com] has left #shogun [] | 19:18 | |
n4nd0 | bye | 19:19 |
-!- blackburn [~blackburn@81.28.187.0] has joined #shogun | 19:22 | |
-!- emrecelikten [~emre@213.153.218.163] has joined #shogun | 19:44 | |
-!- alexlovesdata_ [c25faeba@gateway/web/freenode/ip.194.95.174.186] has quit [Quit: Page closed] | 19:45 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 20:46 | |
wiking | blackburn: ping | 20:47 |
-!- ckwidmer [8ca3fe9d@gateway/web/freenode/ip.140.163.254.157] has quit [Ping timeout: 245 seconds] | 20:47 | |
blackburn | wiking: pong | 20:48 |
wiking | any ideas about the cv? | 20:48 |
blackburn | larank, gmnp? | 20:49 |
wiking | yep | 20:49 |
wiking | :) | 20:49 |
blackburn | well I didn't check yeat | 20:49 |
blackburn | yet* | 20:49 |
blackburn | but will do | 20:49 |
-!- heiko [~heiko@host86-180-42-236.range86-180.btcentralplus.com] has joined #shogun | 21:17 | |
wiking | blackburn: we don't allow multiple inheritance in shogun right? | 21:20 |
blackburn | wiking: exactly | 21:20 |
wiking | :( | 21:20 |
-!- heiko [~heiko@host86-180-42-236.range86-180.btcentralplus.com] has left #shogun [] | 21:24 | |
blackburn | wiking: working on laranking | 21:28 |
blackburn | or may be | 21:29 |
blackburn | wiking: working oning laranking | 21:29 |
blackburn | not enough ing here :D | 21:29 |
-!- puffin444 [62e3926e@gateway/web/freenode/ip.98.227.146.110] has quit [Ping timeout: 245 seconds] | 21:32 | |
wiking | :> | 21:32 |
wiking | cool | 21:32 |
blackburn | wiking: interesting bug, kernel disappeared | 21:33 |
wiking | blackburn: i know that why | 21:34 |
wiking | basically after a train_machine() | 21:34 |
wiking | the whole structure (LaRankOutput) is destroyed | 21:34 |
wiking | by the destroy() on the end of train_machine | 21:34 |
wiking | so when the 2nd time it's called | 21:35 |
wiking | you'll get a null pointer reference | 21:35 |
blackburn | wiking: I do not understand why it modifies m_kernel | 21:36 |
wiking | i.e. train_machine is being called twice on the same machine | 21:36 |
wiking | blackburn: it doesn't modify m_kernel | 21:36 |
wiking | it modifies LaRankOutput | 21:36 |
wiking | LaRank->outputs | 21:38 |
wiking | within that you'll have a null pointer reference for some reason | 21:38 |
blackburn | argh | 21:41 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 21:52 | |
-!- romi_ [~mizobe@187.74.10.115] has joined #shogun | 21:57 | |
blackburn | I'd rather get back to it later | 22:00 |
wiking | xD | 22:06 |
@sonney2k | wiking, are you working on your new patch? | 22:12 |
wiking | sonney2k: yep | 22:14 |
wiking | trying to finish | 22:14 |
blackburn | hey guys | 22:17 |
blackburn | sonney2k: wiking: is there any way to compute low-rank approx of matrix without explicitly computing svd? | 22:18 |
blackburn | n4nd0: ^ | 22:18 |
n4nd0 | blackburn: no idea | 22:20 |
-!- emrecelikten [~emre@213.153.218.163] has quit [Quit: Leaving.] | 23:11 | |
@sonney2k | blackburn, please push Chris to submit your evaluation (he is the only one who didn't yet...) | 23:40 |
blackburn | hmm ok | 23:41 |
@sonney2k | and wiking you too dont' forget about it! | 23:41 |
wiking | sonney2k: about alex? | 23:41 |
wiking | sonney2k: we had a chat today about it he said he'll deal with it but he had to run | 23:41 |
@sonney2k | wiking, no. you have to submit your eval too! | 23:41 |
wiking | ah yeah | 23:42 |
wiking | ! | 23:42 |
blackburn | sonney2k: sent a mail to him | 23:42 |
@sonney2k | (I can see who has / hasn't) | 23:42 |
blackburn | what is due? | 23:42 |
@sonney2k | blackburn, you man when? | 23:42 |
@sonney2k | mean | 23:42 |
wiking | sonney2k: has eyes everywhere | 23:42 |
blackburn | yeah yeah | 23:42 |
blackburn | when | 23:42 |
@sonney2k | blackburn, still a few days | 23:42 |
blackburn | 13? | 23:42 |
@sonney2k | but I set the deadline for end of today | 23:42 |
@sonney2k | because it is *important* to not miss it | 23:43 |
blackburn | yeah would be bad | 23:43 |
blackburn | sonney2k: if you have any other way except email could you please push him too? | 23:45 |
blackburn | skype maybe? | 23:45 |
blackburn | I have no skype of him actually | 23:45 |
blackburn | July 13 19:00 UTC is due | 23:46 |
blackburn | okayy | 23:46 |
@sonney2k | blackburn, sure there still is time - I always play it safe | 23:47 |
@sonney2k | wiking, thx | 23:58 |
@sonney2k | new the PR and we are one happy family :D | 23:58 |
@sonney2k | wiking, btw did the COFFIN principle make sense to you now? | 23:59 |
@sonney2k | blackburn, I really wonder what rieck is doing... he is probably watching us and training one of his bots :D | 23:59 |
blackburn | sonney2k: he said one thing | 23:59 |
blackburn | 'it's been a long' | 23:59 |
--- Log closed Wed Jul 11 00:00:02 2012 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!