Open in new window / Try shogun cloud
--- Log opened Thu Nov 01 00:00:17 2012
-!- wiking_ [] has joined #shogun00:17
-!- wiking [~wiking@huwico/staff/wiking] has quit [Ping timeout: 264 seconds]00:19
-!- ptizoom [~christian@] has quit [Quit: Ex-Chat]03:24
-!- ptizoom_ [~christian@] has joined #shogun03:24
-!- n4nd0 [] has joined #shogun07:09
-!- ptizoom_ is now known as ptizoom08:04
blackburnn4nd0: hey you were writing to my webirc client at job - but I wasn't at job :)08:19
n4nd0blackburn: oops :P08:19
@sonney2kblackburn, do you even have the slightest idea what ptizoom is talking about?08:19
blackburnn4nd0: well my mistake to not close it08:19
blackburnsonney2k: NO~08:20
blackburnn4nd0: mas que nada!08:20
@sonney2kalright, let me continue the gsoc billing08:21
blackburnsonney2k: big boss counting money?08:21
blackburnsonney2k: no news from carol?08:22
@sonney2kblackburn, is it one week yet?08:25
@sonney2kblackburn, yeah counting money but heiko will receive it :D08:25
blackburnsonney2k: one day more and yes I guess :)08:25
@sonney2kwell sorting bills lets say08:25
@sonney2keven worse08:25
ptizoomoye, ptizoom, was doing metaphors and poetry, as sure as cosacks can ride...killing time on ics. but anyway forget it...09:20
sonne|workptizoom: we appreciate some daily poem from ptizoom on shogun planet :D09:22
-!- n4nd0 [] has quit [Ping timeout: 240 seconds]09:24
ptizoomI was also praising in the dojo of ML.09:24
ptizoomanyway enough unflattering flaf...I still do not understand why ...09:26
ptizoom*** Testing in ../data/classifier/*.m09:26
ptizoom../data/classifier/LibLinear_0017_1en05_1_True.m                ERROR09:26
ptizoomLibLinear - yesclassified: 5.644953e-02, bias: 9.286755e-04 <--- accuracy: 1.000000e-0409:26
ptizoomsonne*: shall this test be spot on, below accuracy .0001 ?09:27
ptizoomso my perl stuff is very dodgy indeed...09:28
blackburnptizoom: but it is .m? how perl?09:30
ptizoomI checked memory with debugger of vectorisation and matrification in/out and it shall be ok, but ... there must be a miss match in the engine...09:30
sonne|workblackburn: man thats the test09:30
blackburnsonne|work: matlab/octave test right?09:30
sonne|workblackburn: no09:30
ptizoomblackburn... yes09:31
blackburnsonne|work: it tests perl?09:31
sonne|workblackburn: this is just some file with the output09:31
blackburnhmm okay09:31
sonne|workptizoom: it suggests that sth is wrong with vector/matrix typemaps09:32
sonne|workptizoom: check if sth basic like the gaussian kernel works09:32
ptizoomI was slowly but surely perlifying the tests in /tests/unit...09:37
ptizoomthat would tell me where are the deficiencies in the swig perlmodular...09:39
ptizoombut hey, thanks for the tip.09:39
ptizoomfor info, not a single .m test works within accuracy ...09:41
ptizoomho gobble...09:42
sonne|workptizoom: well do sth very simple09:43
sonne|workcreate some matrix09:43
sonne|workfrom that x=RealFeatures(matrix)09:43
sonne|workand then x.get_feature_matrix() (or what it is called)09:43
sonne|workand compare if results match09:43
sonne|workif that works do k=GaussianKernel(x,x)09:44
ptizoomnoted, I will come to that. thanks09:46
sonne|workyou can compare the result w/ what you get on python_modular09:51
ptizoomI know... but that would be to easy, that would be cheating!09:51
-!- n4nd0 [] has joined #shogun10:12
sonne|workptizoom: I prefer cheating^H^H^H^H^H^H^H^Hwinning11:34
-!- n4nd0 [] has quit [Ping timeout: 256 seconds]12:24
-!- n4nd0 [] has joined #shogun12:50
-!- romi_ [~mizobe@] has joined #shogun13:30
-!- n4nd0 [] has quit [Ping timeout: 260 seconds]14:52
-!- Arya [bc9fd512@gateway/web/freenode/ip.] has joined #shogun15:19
AryaHi,I'm getting assertion failed when trying to read binary machines for multiclass classification. the line is : "CSVM *the_svm = (CSVM *)svm->get_machine(i);" which the_svm is Null15:22
blackburnArya: hi15:35
blackburnArya: how did you learn that "svm"?15:35
blackburnwhat is the type?15:36
blackburnArya: my first guess is that binary machine could be of CLinearMachine class15:46
Aryayes libsvmm , I'm running the classifier_multiclasslibsvm.cpp example on my dataset15:47
Aryawhen the code reaches svm->apply() it happens15:48
blackburnshould be a bug when15:49
blackburnArya: my other guess is that could happen if you have more classes in test data than in train data15:50
Aryano, I test it on train data, but the same happens15:51
blackburnArya: can you please check number of machines after training? method 'get_num_machines'15:52
Aryait's correct, n(n-1)/215:53
Aryaint mc_libsvm(const char* ds, float64_t c, float64_t width, int32_t split) { init_shogun_with_defaults(); std::stringstream path; path<< "/home/arya/datasets/"<<ds << ".libsvm.train."<<split;  uint32_t n=0,d=0; std::string train_file=path.str(); const char* train_file_name=train_file.c_str(); get_size(train_file, &n, &d); SGMatrix<float64_t> X(d,n); SGVector<float64_t> y(n);  read_data(train_file_name, X, y); CLabels* train_15:53
Aryacan I send you my function?15:53
blackburnArya: pastebin maybe15:54
blackburnor mail me blackburn91 at gmail.com15:55
blackburnArya: gtg now - please email me I'll take a look15:59
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has quit [Quit: Page closed]15:59
-!- Arya [bc9fd512@gateway/web/freenode/ip.] has quit [Quit: Page closed]16:05
-!- romi_ [~mizobe@] has quit [Quit: Leaving]16:25
-!- arya [05c7817b@gateway/web/freenode/ip.] has joined #shogun16:30
-!- arya [05c7817b@gateway/web/freenode/ip.] has quit [Quit: Page closed]16:54
-!- n4nd0 [] has joined #shogun16:57
-!- blackburn [~blackburn@] has joined #shogun17:13
-!- wiking_ is now known as wiking17:58
-!- wiking [] has quit [Changing host]17:58
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun17:58
-!- audy [~audy@unaffiliated/audy] has quit [Remote host closed the connection]18:28
-!- n4nd0 [] has quit [Ping timeout: 252 seconds]19:55
-shogungit:#shogun- [shogun] lisitsyn pushed 1 new commit to master:
-shogungit:#shogun- shogun/master a55b25e Sergey Lisitsyn: Fix for 3-class training of multiclass libsvm20:04
@sonney2kblackburn, any clues on the bug?20:33
-!- n4nd0 [] has joined #shogun20:55
blackburnsonney2k: which bug?21:07
n4nd0blackburn: haha I see in the news Oksana is getting into shogun21:10
n4nd0well done! :)21:10
blackburnwell I've spent a while trying to understand the error and she helped21:11
blackburnso credit goes to21:11
@sonney2kblackburn, what is the  SG_PRINT("Num classes = %d\n",num_classes); in your patch?21:17
blackburnsonney2k: oops juk21:17
blackburnhere you go21:18
-shogungit:#shogun- [shogun] lisitsyn pushed 1 new commit to master:
-shogungit:#shogun- shogun/master 5c03620 Sergey Lisitsyn: Removed junk debug print21:19
blackburnn4nd0: # vectors1000021:21
blackburnCovertree construction took 7.880000 seconds21:21
blackburnCovertree-based neighbors search took 21.230000 seconds21:21
n4nd0blackburn: what about the na?ve strategy?21:22
n4nd0what is the speedup with covertree wrt it?21:22
blackburnn4nd0: partial_sort?21:22
n4nd0blackburn: yeah, the other one that was at the beginning21:22
blackburnmuch slower I bet21:22
blackburnn4nd0: before it was 14.94s for 1000021:24
blackburn23s now21:24
blackburnso a room for improvement still :D21:24
n4nd0so, that covertree is slower?21:25
blackburnn4nd0: probably21:25
blackburnJL is faster but it you are the only guy that managed to get it working21:25
n4nd0I still don't understand how is it possible that it can be slower than sorting!21:25
blackburnn4nd0: it is impossible :D21:27
n4nd0but it is like that, isn't it?21:30
blackburnn4nd0: ??21:31
n4nd0that covertree is slower than na?ve sorting21:31
n4nd0the one you are testing21:32
n4nd0you pasted ^21:32
blackburnyou misunderstood me21:32
n4nd0oh I see, sorry21:32
blackburnit was a covertree as well21:32
-!- travis-ci [] has joined #shogun21:32
travis-ci[travis-ci] it's Sergey Lisitsyn's turn to pay the next round of drinks for the massacre he caused in shogun-toolbox/shogun:
-!- travis-ci [] has left #shogun []21:32
blackburntravis-ci you are drunk go away21:33
n4nd0I love that guy, he has style21:34
n4nd0you know logs in21:34
n4nd0blame some people and go away21:34
@sonney2kblackburn, hurray - I finished the billing stuff :)22:38
@sonney2kJust submitted to google22:38
@sonney2klets hope it works...22:38
blackburnsonney2k: what is it? what you had to do?22:39
@sonney2kblackburn, collect bills write some itemized billing and submit22:41
blackburnno idea22:42
blackburnwe just steal money here and stash it under a pillow22:43
@sonney2kyeah I guess that is what you do in rossia22:43
@sonney2kwiking, any ideas about
@sonney2kblackburn, do you know whether wiking is busy with some paper or sth?22:44
blackburnsonney2k: he should be probably yes22:44
@sonney2kI need to have a second dinner22:45
blackburnare you growing up?22:45
@sonney2knot volume wise - but yes I guess doing the bills did cost a lot of energy ;)22:46
@sonney2kyay pizza is waiting for me :)22:49
@sonney2kyummy yummy yummy22:50
n4nd0damn it, that made me hungry22:51
@sonney2kn4nd0, just join me22:52
* sonney2k hands over a slice of pizza to n4nd0 22:52
n4nd0sonney2k: danke22:52
@sonney2kn4nd0, how is it going?22:53
blackburnhaha superfriendly sonney2k now22:53
* sonney2k hands another slice to blackburn 22:53
n4nd0sonney2k: good, what about you?22:53
@sonney2kn4nd0, long time no code contributions? I guess you are pretty busy currently...22:54
@sonney2kn4nd0, I am just glad that I managed to write this blogpost somehow22:54
@sonney2kI suck at blogging22:55
blackburnsonney2k: I employed him22:55
blackburnmy scout reached him by email22:55
n4nd0sonney2k: yeah, I am trying to put KNN into the EDRT22:55
blackburnyou know I have scout team22:55
n4nd0hopefully I will have something ready by tom night22:55
@sonney2kI see, you hijacked him22:55
blackburnsonney2k: you have JMLR pubs we don't22:56
@sonney2kblackburn, does it make me feel better in any way?22:56
blackburnsonney2k: you are the only who knows22:56
@sonney2kthen guess what the answer is22:57
blackburnwe are young and want to try various dangerous things haha22:57
n4nd0hehe good answer for a good question I think22:57
n4nd0sonney2k: I want to put the word recognizer application inside though22:57
@sonney2kn4nd0, that would be cool22:57
n4nd0I think I will go a little bit slow with it but I'll do it22:57
@sonney2kn4nd0, could we make this some live thing?22:58
blackburnsonney2k: come on, all the EDRT stuff will be included to shogun anyway22:58
@sonney2kblackburn, I always wanted to have KNN :P22:58
blackburnsonney2k: it is already in shogun22:58
@sonney2kblackburn, I know :) I wrote it22:58
n4nd0sonney2k: I have done a live one that worked fine but the problem is the data22:59
blackburnsonney2k: I mean with covertree not sort22:59
n4nd0the one I did worked fine with my and bernard's voices22:59
n4nd0but it will probably explode with other's22:59
blackburnsonney2k: NNs is a bottleneck for me so I hired a best specialist in JL's covertree around :)22:59
@sonney2kblackburn, s/hired/hijacked :)23:00
@sonney2kn4nd0, well then we need more people :)23:00
n4nd0blackburn, sonney2k :
n4nd0open speech recognition23:00
@sonney2kn4nd0, speaking of bernard - is he doing fine?23:00
@sonney2ksince he wrote the website he basically disappeared23:00
blackburnn4nd0: impressive CV I'd hire him too23:01
n4nd0sonney2k: yeah, he is23:01
n4nd0sonney2k: we are doing together a CV course now and he is starting to look for a company where to do his master thesis23:01
blackburnOMFG I have to do people detector23:02
blackburngood you did remind me;23:02
blackburnehm sleep reduction technique anyone?23:02
@sonney2kn4nd0, we are trying to do some live demos on the website with shogun as backend and django & d3js etc as frontend. maybe he would be interested in playing with this?23:02
blackburnyeah we will consult him with hyperplanes23:03
n4nd0sonney2k: I'll ask him right now23:03
blackburnn4nd0: is he your flatmate? :D23:03
n4nd0not really23:03
@sonney2kn4nd0, basically stuff like we have in the graphical examples + e.g. the ocr demo23:03
@sonney2kbut face recognition would be even cooler :D23:04
@sonney2k+ gender detection23:04
@sonney2kbut quite some effort23:04
n4nd0sonney2k: he said he is sorry but no :(23:04
blackburnsonney2k: haha I came up with very offensive app23:04
@sonney2kn4nd0, too bad.23:04
n4nd0sonney2k: face detection would be awesome but we need adaboost for that!23:04
@sonney2kn4nd0, no23:04
blackburnso you enter the website it takes image from your camera23:04
n4nd0that reminds me when I started to come around here, adaboost23:04
blackburnand says you are a man23:05
@sonney2kn4nd0, I would rather just use opencv's FD23:05
@sonney2kit works OKish23:05
@sonney2kand on top of it we put the gender classifier23:05
blackburnopencv is a heap of api crap23:05
@sonney2kwho cares23:05
@sonney2kit does the job :D23:06
n4nd0I know they use Haar features23:06
n4nd0is it with adaboost too?23:06
blackburnit has god damn WORST HOG thing ever23:06
blackburnthey should rather call it humandetectorarraycomparator :D23:06
@sonney2kn4nd0, some cascades - decision trees maybe?23:06
@sonney2k(wildly guessing)23:06
n4nd0maybe a cascade of strong classifiers with adaboost23:07
n4nd0that's what we did in our course actually23:07
@sonney2kno boosting I think23:07
blackburnn4nd0: why do you think adaboost is so required here?23:07
@sonney2kblackburn, it really is very well suited for that task23:07
blackburnfor haar yes23:08
@sonney2kand other object recognition23:08
n4nd0sorry for that ugly link...23:08
n4nd0but it looks like it is adaboost23:08
n4nd0blackburn: I think it is a pretty awesome algorithm23:09
blackburnn4nd0: dependsssssss23:09
n4nd0blackburn: and also, it is the only thing I have done for face detection :D idk if there are other strategies23:09
blackburneven LBP + linear svm works23:09
blackburnwith HKM even better I think23:10
@sonney2kblackburn, but your traffic sign stuff would also be very cool23:10
n4nd0that link was definitely nicer :S23:10
@sonney2kblackburn, it would actually be very easy to put it on the website once you put the thing into /applications23:10
blackburnsonney2k: draw me how do you see it23:11
@sonney2kblackburn, just some image upload button -> name or what do you mean?23:12
blackburnsonney2k: n4nd0: [BENCHMARK] Randomized eigendecomposition took 0.57 seconds23:12
blackburnsonney2k: n4nd0: 15K x 15K eigenproblem23:12
blackburnresult is a little approximate though but come on it is linear23:13
blackburnsonney2k: I am a good composer of stolen methods :D23:13
@sonney2kblackburn, its an art!23:14
@sonney2kstealing is believing :)23:14
@sonney2klooks like they have one based on LBP23:16
blackburnsonney2k: yes they do23:16
@sonney2kprobably better these days23:16
blackburnLBP is as well good for faces23:16
blackburnHOG+LBP is unbeatable :D23:17
@sonney2kblackburn, don't tell that yann lecun :D23:17
blackburnsonney2k: well I respect his vision that we should learn features not models23:18
@sonney2kalright I guess when I find the time I will do some shogun demoing23:19
@sonney2kbut for now23:19
@sonney2kbed time23:19
n4nd0time to sleep here too, good night23:24
blackburnn4nd0: night23:24
-!- n4nd0 [] has quit [Quit: leaving]23:24
--- Log closed Fri Nov 02 00:00:17 2012