Open in new window / Try shogun cloud
--- Log opened Thu Aug 09 00:00:17 2012
shogun-buildbot_build #290 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: iglesias <>00:14
shogun-buildbot_build #291 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular]  Build details are at  blamelist: Sergey Lisitsyn <>00:26
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]01:50
shogun-buildbot_build #48 of nightly_none is complete: Failure [failed compile]  Build details are at
shogun-buildbot_build #54 of nightly_default is complete: Failure [failed compile]  Build details are at
shogun-buildbot_build #44 of nightly_all is complete: Failure [failed compile]  Build details are at
-!- zxtx [] has joined #shogun04:28
-!- pluskid [~pluskid@] has joined #shogun05:45
pluskidhi sonney2k05:46
@sonney2kpluskid, hi05:46
@sonney2kpluskid, not sure if you have seen but we are still fighting with your patch to work/compile on all interfaces05:47
pluskidsonney2k: oh, the buildbot seems flooding me by sending lots of emails, so I didn't read them all05:47
pluskidmaybe I should have a look at the latest buildbot output05:48
@sonney2kyeah well for a reason - all your fault ;-)05:48
@sonney2kpluskid, no seriously you did add a python_modular example that didn't work05:48
@sonney2kso I tried to fix this05:48
@sonney2kwhich creates problems in java_modular (of course :(05:48
pluskidsonney2k: this is the latest result right?05:48
@sonney2kpluskid, look at waterfall05:48
@sonney2kpluskid, but what I wanted to ask - you have a TreeMachine in multiclass that is templated05:49
@sonney2kis that really needed05:49
@sonney2kpluskid, but maybe let me explain a bit more first05:49
pluskidah, that would be good05:49
@sonney2kmany modular interfaces dont' support templates at all05:49
@sonney2k(python, ruby, ...)05:50
@sonney2kso that means for them one has to add an extra line in the corresponding swig .i interface file %template ...05:50
@sonney2kso if templates are not *required* or avoid code duplication don't use them05:50
pluskidhmm, I see05:51
@sonney2kif we have objects / structs that are not derived from SGObject we currently only have a workaround solution to get java and csharp modular to compile05:52
pluskidmaybe I should take a look at how to convert the tree-machine into non-template05:52
@sonney2kwhy? because we support serialization of native shogun objects05:52
@sonney2kand so we need the functions in SGObjects to be available basically in all classes/structs - which of course they are not05:53
pluskidso if the template base class is ignored by SWIG, the inheritance chain will break, right?05:54
@sonney2kso we currently have a workaround macro in an .i file where we add SERIALIZABLE_DUMMY(struct)05:55
@sonney2kpluskid, for swig to work we need to include all .h's from the inheritance chain (in the right order!)05:55
pluskidsonney2k: ah, I see05:56
@sonney2kpluskid, if one is a template in between we have to add %template to make it visible05:56
pluskidsonney2k: I'll try to remove the template. Luckily there are only two tree machines05:57
@sonney2kpluskid, anyway maybe we can resolve this quickly - but I need your help:05:57
@sonney2kI added these lines to Multiclass.i:05:57
@sonney2k%include <shogun/multiclass/tree/BalancedConditionalProbabilityTree.h>05:57
@sonney2k%include <shogun/multiclass/tree/ConditionalProbabilityTree.h>05:57
@sonney2k%include <shogun/multiclass/tree/RandomConditionalProbabilityTree.h>05:57
@sonney2k%include <shogun/multiclass/tree/RelaxedTree.h>05:57
@sonney2k%include <shogun/multiclass/tree/RelaxedTreeNodeData.h>05:57
@sonney2k%include <shogun/multiclass/tree/RelaxedTreeUtil.h>05:58
@sonney2k%include <shogun/multiclass/tree/TreeMachine.h>05:58
@sonney2k%include <shogun/multiclass/tree/TreeMachineNode.h>05:58
@sonney2k%include <shogun/multiclass/tree/VwConditionalProbabilityTree.h>05:58
@sonney2kpluskid, do we really need all of them to be visible from the interfaces? or are some of them just for internal use?05:58
pluskidsonney2k: VwConditionalProbabilityTree cannot work properly because it seems the Vw in shogun isn't working properly, maybe I should remove this one05:59
pluskidto make the template work, maybe all of them are needed06:00
pluskidexcept RelaxedTreeUtil.h06:00
@sonney2kpluskid, heh - RelaxedTreeUtil.h is to simple to cause trouble :D06:01
pluskidsonney2k: so what's the trouble here?06:02
pluskidI cannot understand the java_modular compile error06:02
pluskidsonney2k: shall we expand with %template the superclass for RelaxedTree and ConditionalProbabilityTree for SWIG?06:03
-!- naywhayare [] has quit [Ping timeout: 252 seconds]06:03
@sonney2kpluskid, I am just compiling it locally here...06:03
@sonney2kpluskid, which of the classes above are not derived from SGObject?06:04
pluskidall derived from SGObject06:05
-!- naywhayare [] has joined #shogun06:05
pluskidsonney2k: maybe have a look at TreeMachineNode.h ?06:05
pluskidit seemed I got some compile error of the class_list so I play a trick in the syntax of class derivation to avoid it being detected by the script06:06
@sonney2kpluskid, if you don't prefix your class with C then it will not be included in class list - also you can add the IGNORE_IN_CLASS_LIST (or so define)06:08
@sonney2kpluskid, RelaxedTreeUtil is not derived from sgobject06:08
@sonney2kpluskid, why not?06:09
pluskidsonney2k: It is just a util, like a struct06:09
@sonney2kok then we %ignore it06:09
@sonney2kpluskid, ok so now we have these06:13
@sonney2k../../shogun/multiclass/tree/ConditionalProbabilityTree.h:43: Warning 401: Nothing known about base class 'CTreeMachine< ConditionalProbabilityTreeNodeData >'. Ignored.06:13
@sonney2k../../shogun/multiclass/tree/ConditionalProbabilityTree.h:43: Warning 401: Maybe you forgot to instantiate 'CTreeMachine< ConditionalProbabilityTreeNodeData >' using %template.06:13
@sonney2k../../shogun/multiclass/tree/RelaxedTree.h:34: Warning 401: Nothing known about base class 'CTreeMachine< RelaxedTreeNodeData >'. Ignored.06:13
@sonney2k../../shogun/multiclass/tree/RelaxedTree.h:34: Warning 401: Maybe you forgot to instantiate 'CTreeMachine< RelaxedTreeNodeData >' using %template.06:13
@sonney2kso I guess it needs TreeMachine to be included first and node data too06:13
pluskidsonney2k: so to make it work, we will use %template here, right?06:13
pluskidyes, TreeMachine, and NodeData and then use %template06:14
pluskidsonney2k: btw: CondProbTreeNodeData is in the ConditionalProbabilityTree.h file06:14
@sonney2kpluskid, any ideas for a name?06:15
@sonney2kpluskid, so06:17
@sonney2k    %template(TreeMachineWithConditionalProbabilityTreeNodeData) CTreeMachine<ConditionalProbabilityTreeNodeData>;06:17
@sonney2k    %template(TreeMachineWithRelaxedTreeNodeData) CTreeMachine< RelaxedTreeNodeData>;06:17
pluskidsonney2k: yeah, does it work now?06:18
@sonney2kpluskid, of course not o_O06:20
pluskidsonney2k: why?06:21
@sonney2kpluskid, ohh it is getting better06:22
@sonney2k../../shogun/multiclass/tree/ConditionalProbabilityTree.h:43: Warning 401: Nothing known about base class 'CTreeMachine< ConditionalProbabilityTreeNodeData >'. Ignored.06:22
@sonney2k../../shogun/multiclass/tree/ConditionalProbabilityTree.h:43: Warning 401: Maybe you forgot to instantiate 'CTreeMachine< ConditionalProbabilityTreeNodeData >' using %template.06:22
pluskidstrange, we've instantiated it06:24
pluskidsonney2k: oh, maybe we should put  ConditionalProbabilityTreeNodeData in a separate header file06:24
@sonney2kpluskid, maybe order of includes is wrong06:25
* sonney2k checks06:25
pluskidand include that header file06:25
pluskidthen use %template06:25
pluskidthen include conditionalprobabilitree.h06:25
@sonney2kif I move it up to relaxedtreenode include06:25
@sonney2kthen I get06:25
@sonney2k../../shogun/multiclass/tree/ConditionalProbabilityTree.h:43: Warning 401: Base class 'CTreeMachine< ConditionalProbabilityTreeNodeData >' undefined.06:25
@sonney2kMulticlass.i:73: Warning 401: 'CTreeMachine< ConditionalProbabilityTreeNodeData >' must be defined before it is used as a base class.06:25
pluskidsonney2k: sorry, the family is calling me for the lunch, I'll see you later06:26
@sonney2kpluskid, kids are waking up here so I will have to stop too ... I will commit what I have maybe you can get it to work...06:27
CIA-18shogun: Soeren Sonnenburg master * r4b11e44 / (5 files in 2 dirs): add a couple of missing %templates to Multiclass.i -
pluskidsonney2k: shall I compile with java_modular to test this?06:38
pluskidsonney2k: is there something wrong with the git repo?07:08
pluskidsonney2k: I get a very long list of applying many patches that are not created by me when I rebase on upstream/master07:08
shogun-buildbot_build #292 of deb3 - modular_interfaces is complete: Failure [failed test java_modular]  Build details are at  blamelist: Soeren Sonnenburg <>07:09
@sonney2kpluskid: see the new buildbot error running java tests.08:41
pluskidsonney2k: I'm getting some trouble with the git repo08:41
@sonney2kthe eEuclidean example needs to be renamed08:42
@sonney2kpluskid: what trouble ?08:43
pluskidsonney2k: I run rebase against upstream/master08:43
pluskidthen it shows a huge amount of patches being applied08:43
pluskidI guess something is wrong here08:43
-!- n4nd0 [] has joined #shogun08:43
pluskidhi n4nd008:44
n4nd0hey pluskid08:44
n4nd0that was fast :)08:44
pluskidn4nd0: will you fetch and rebase the latest shogun git repo08:44
pluskidto see do you get any trouble?08:44
n4nd0pluskid: sure08:44
pluskidI'm seeing something strange here, but not sure it is bad or ordinary08:45
n4nd0pluskid: what kind of thing?08:45
n4nd0any interface in particular?08:45
pluskidn4nd0: I run git rebase08:45
@sonney2kpluskid: no idea..  maybe someone forgot to rebase and someone merged such pr08:45
pluskidit says a large amount of patches are being made08:45
pluskidusually it only apply patches that I made locally08:46
n4nd0pluskid: aham08:46
n4nd0I could rebase without problems08:46
pluskidn4nd0: OK08:46
pluskidthen I'll check my repo08:46
pluskidmaybe the easiest thing is to do a new clone from github08:47
n4nd0no idea08:47
n4nd0pluskid: have you googled the error?08:47
pluskidn4nd0: no error, just some strange things08:48
n4nd0sonney2k: btw, I asked you yesterday for any reference on risk functions for hm-svm08:49
n4nd0do you know something about it?08:49
n4nd0so far I have just found a Ph.D. thesis: Regularized bundle methods for large-scale learning problems with an08:49
n4nd0application to large margin training of hidden Markov models08:49
@sonney2kn4nd0: it is the right part in rhetoric unconstrained formulation08:51
@sonney2kscratch rhetoric08:51
@sonney2ktyping on mobile here08:52
-!- uricamic [~uricamic@2001:718:2:1634:3078:70c3:e281:1376] has joined #shogun08:54
pluskidsonney2k: I fixed the java euclidian error. Does the output of the buildbot of java mean we are getting tree machine OK there?08:56
n4nd0sonney2k: sorry I didn't get it, should I look for rhetoric unconstrained formulation? scratch unconstrained?08:56
@sonney2kpluskid: there is still a warning08:57
n4nd0yeah that's how I reacted when I read08:59
n4nd0<@sonney2k> n4nd0: it is the right part in rhetoric unconstrained formulation08:59
n4nd0<@sonney2k> scratch rhetoric08:59
@sonney2kI said scratch rhetoric09:00
@sonney2kremove this word09:00
n4nd0oh shit09:00
n4nd0I am retarded09:00
n4nd0sonney2k: sorry09:00
pluskidsonney2k: ah, I finally find it, I will try to compile it locally09:00
@sonney2ktouchpad typing on mobile is not exactly easy :-/09:02
n4nd0yeah, I understand. My bad I didn't understand09:03
@sonney2kpluskid: thanks09:13
shogun-buildbot_build #55 of nightly_default is complete: Success [build successful]  Build details are at
-!- nickon [] has joined #shogun09:35
-!- nickon [] has quit [Client Quit]09:35
n4nd0sonney2k: around?10:07
-!- n4nd0 [] has quit [Ping timeout: 240 seconds]10:20
-!- n4nd0 [] has joined #shogun10:22
-!- n4nd0 [] has quit [Client Quit]10:26
-!- n4nd0 [] has joined #shogun10:26
CIA-18shogun: Chiyuan Zhang master * re1a50de / (2 files): renaming java modular: euclidian => euclidean -
CIA-18shogun: Chiyuan Zhang master * r490741f / (4 files in 2 dirs): fix template class of CPT in java modular. -
CIA-18shogun: Soeren Sonnenburg master * rd699fc5 / (6 files in 3 dirs): Merge pull request #698 from pluskid/multiclass -
-!- n4nd0 [] has quit [Ping timeout: 268 seconds]10:53
-!- n4nd0 [] has joined #shogun10:55
-!- n4nd0 [] has quit [Quit: leaving]11:09
shogun-buildbot_build #293 of deb3 - modular_interfaces is complete: Success [build successful]  Build details are at
-!- pluskid [~pluskid@] has quit [Quit: Leaving]11:34
-!- blackburn [~blackburn@] has joined #shogun11:39
@sonney2kshogun-buildbot_, hurray!11:49
shogun-buildbot_What you say!11:49
wikingsonney2k: apart from the minor changes u suggested any other requests?12:15
@sonney2kwiking, well make sure that current svm ocas gives the same result as the updated one12:20
wikingsonney2k: i'll try to define a test for it :D12:22
shogun-buildbot_build #45 of nightly_all is complete: Success [build successful]  Build details are at
wikingwhat was the class name for generating random data by heiko?12:33
wikingi thought that it's in statistics...12:34
shogun-buildbot_build #49 of nightly_none is complete: Success [build successful]  Build details are at
blackburnwiking: it is in features IIRC12:38
wikingmmm let me check12:42
wikingmmm i guess it's going to be DataGenerator.h12:45
wikinglet's see how it works :)12:46
wikingsonney2k: any ideas how svmocas would deal with an input of two normally distributed data?12:54
wikingi mean standard distrib..12:54
@sonney2kwiking, ?13:27
@sonney2kit will separete the data :D13:27
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]13:36
-!- blackburn [~blackburn@] has joined #shogun13:41
wikingyeah 'm just checking it's performance now13:47
CIA-18shogun: Sergey Lisitsyn master * r5e63792 / (8 files in 3 dirs): Added multitask ROC evaluation -
blackburnwiking: my self-education made me to dense sift :D14:01
blackburnI actually do not think it could be a way for huge data14:02
wikingany ideas how i could unit test this14:36
wikingbecause with ROC i get value of 0.4168 -> 0.553214:37
wikingimho i should create the vectors, otherwise with random ones i won't be able to be deterministic enough14:37
@sonney2kwiking, what?14:40
@sonney2kI don't get it14:40
@sonney2kthough for random stuff you should seed the rng14:41
wikingsonney2k: well i've used the datagenerator of heiko14:42
wikingthen i've basically divided the generated matrix into two equally sized feature set14:42
wikingi.e. train/test14:42
wikingthe same amount of each vector in both of the sets14:42
wikingthen i've ran training and test on those features with ocas14:43
wikingand now i'm using ROCEvaluation for checking the classification14:43
wikingand the result varies a lot14:43
wikingso i was wondering how else i could test svmocas14:44
wikingsince this doesn't seemed to be too deterministic14:44
@sonney2kwiking, well you should seed the rng14:44
@sonney2kotherwise it cannot be deterministic14:44
@sonney2kand also use *two* gaussians that don't have same mean14:45
@sonney2kotherwise you will just get about 50% accuracy14:45
wikingas far as i understood14:45
wikingdatagenerator will generate that14:45
wikingTakes each m samples from two distributions p and q, where each element * is standard normally distributed, except for the first dimension of q, * where the mean is shifted by a specified value.14:46
wikingthat's what generate_mean_data(...) does14:46
wikingheh ok this should work14:52
@sonney2kwiking, how many dims?14:55
@sonney2kif this is low dim you should get 100% acc14:55
@sonney2kroc 1.014:55
wikingsonney2k: i've tried with dim = 1014:58
wikingit's now 0.768414:58
wikingwith a given seed14:58
blackburnwhat are means of gaussians?15:00
wikinghere's a utest for SVMOcas15:00
wikinglet's see how u like it :))15:01
blackburnwiking: two nonoverlap gaussians should have roc 1.0..15:01
wikingblackburn: ok check the code where i've done then a mistake :)15:02
blackburnwiking: generate_mean_data does some strange thing15:04
blackburnpretty strange15:05
blackburnwiking: in matlab it would be15:05
blackburnX = randn(dim,2*m)15:05
blackburnX(0,m:2*m) += mean_shift15:05
wikingblackburn: can u try to run this unittest on your machine15:07
wikingjust to see if it's the same at least :)15:07
blackburnwiking: I am not surprised15:07
blackburnwith 0.715:07
blackburnbecause it is the same gaussian15:07
blackburnexcept first variable15:07
blackburnjust check code of data generator generate mean data15:08
blackburnsee what I mean?15:09
blackburnso the only mistake is to use that method :)15:11
CIA-18shogun: Sergey Lisitsyn master * r4286ace / (5 files): A few doc improvements for task api -
CIA-18shogun: Sergey Lisitsyn master * r443f89f / doc/tutorial : Updated tutorial submodule -
-!- gsomix [~gsomix@] has joined #shogun15:18
gsomixgood day15:18
wikingblackburn: ideas?15:22
blackburnwiking: thou shalt use da second method15:23
blackburnand actually you should compare not the auc15:24
blackburnbut the solution15:24
blackburnit should coincide15:24
blackburnwiking: ah generate_sym_mix_gauss is random mixture :D15:58
CIA-18shogun: Sergey Lisitsyn master * r10a0961 / (5 files): A few doc improvements for malsar based multitask algorithms -
wikingblackburn: is there any simple generator in shogun? or shall we add one? :)))16:12
blackburnwiking: no, datagenerator is the only16:12
wikingmaybe then i should extend it with a new fucntion16:13
wikingto be able to generate n gaussians16:13
-!- heiko1 [] has joined #shogun16:23
blackburnheiko1: hey16:24
heiko1blackburn, hi16:24
blackburnheiko1: have you ever heard of success of applying svms to boolean domain?16:25
heiko1boolena domains?16:25
heiko1like boolean functions?16:25
blackburnyeah, features are {0,1}^d16:25
heiko1not direct16:25
heiko1but perceptrons have been studien well in that context16:26
heiko1actually, I think except for xor and the complent, perceptrons can already learn all boolean functions16:26
blackburnheiko1: what is easy to google is a dnf kernel - it appears pretty often16:26
blackburnbut latest is 200516:26
blackburndoesn't sound like a success story :D16:27
heiko1I mean this topic is like ultra old16:27
heiko1so what do you want to do?16:27
blackburnheiko1: I have bit features16:27
blackburnnow I am learning just a linear svm16:27
heiko1I see16:28
blackburnbut linear is not enough I guess16:28
heiko1you can linearly discriminate all boolean functions  but xor16:28
heiko1so you weould need a kernel16:28
heiko1a kernel that for example allows to distinguish the xor set (plus all others) is (x1, x2, x1*x2)16:29
blackburnheiko1: yeah, dnf kernel does that16:29
heiko1but then you are fine16:29
heiko1in this feature space all booleans are linearly distinguishable16:29
blackburnsuper slow, I thought of better way16:29
heiko1I see16:29
blackburndataset is like 40K16:30
blackburnroad signs that is, again16:30
heiko1and the how long does liblinera take?16:32
blackburnheiko1: well kind of 5 minutes16:32
blackburnheiko1: one problem I don't know how to approach16:33
blackburnit is defined as -1 + 2^same1(x,y)16:33
heiko1but you dont have kernel parameters16:34
heiko1isnt 5 mins fine then=?16:34
heiko1I suspect this is training time16:34
heiko1what is apply time?16:34
blackburnheiko1: yeah but liblinear is linear16:34
blackburnwith kernel it is kind of infinity16:34
blackburnor I am doing something wrong16:34
heiko1what if you dont use a kernel but manually construct new features? What dimension are you in?16:35
blackburnheiko1: 512 bits16:35
heiko1mmh, too many16:36
blackburn13407807929942597099574024998205846127479365820592393377723561443721764030073546976801874298166903427690031858186486050853753882811946569946433649006084096 is pretty long16:36
heiko1what about Gaussian kernel? :)16:36
heiko1on original features16:36
blackburnwell I guess it would be the same slow16:36
heiko1or waht about feature selection before that, I guess the important things are happening in a few dimensions only16:36
blackburnheiko1: any idea how to normalize -1+2^bignumber?16:37
blackburnor rather how to avoid overflow16:37
blackburnheiko1: does it mean with kernel mapping to linearly separable space I don't have to use C at all?16:38
heiko1ehm what?16:39
blackburnyou said boolean functions are linearly separable16:39
heiko1What I meant is that you probably dont need all these dimensions16:39
blackburnno, before16:39
heiko114/16 are linearly separable in  2 dimenions16:39
blackburnI mean if kernel maps to space where features are linearly separable16:40
blackburndo I have to regularize at all?16:40
heiko1I see16:40
heiko1if the data is 100% separable16:40
heiko1do dont really need to do that16:40
heiko1(since the test error is 0 anyway)16:40
heiko1but you still get a better margin with reg.16:41
heiko1so more robustness16:41
blackburnPROGRESS:       nan    -nan seconds remaining    -nan seconds total16:44
blackburnprobably I would have to wait a few lives more16:44
heiko1hehe :)16:44
blackburnwill try to learn on subset16:45
heiko1do some feature selection16:45
heiko1throw away pointless dimensions16:45
heiko1if that wont help, the problem is unsolvable anyway (which I guess it isnt since my brain can solve it :D)16:45
blackburnheiko1: well even dropping 50% of features will speed up it only twice16:46
heiko1sorry, I meant not features in shogun sense16:46
heiko1but dimensions of data16:46
heiko1with that number16:46
heiko1you can probably drop all but a few thousand or eeven hundred16:46
blackburnprobably my kernel is just wrong16:46
blackburnmy features are wrong :D16:47
blackburnI forgot to remove bitwise stuff16:47
blackburnso I learn on 4K dimension features with kernel = const16:47
heiko1ehm ,what? :)16:48
blackburnheiko1: I forgot to remove my map 01010101 -> [0,1,0,1,0,1,0,1]16:49
heiko1and now it works?16:51
blackburnno idea it is slow even after reducing :(16:51
heiko1try a PCA16:53
blackburnheiko1: will do16:55
heiko1and look at the Eigenspectrum, that will tell you how much it works16:55
blackburnheiko1: eigenspectrum of features?16:56
-!- gsomix [~gsomix@] has quit [Ping timeout: 245 seconds]16:56
blackburnheiko1: wait, how can I do PCA of boolean features16:56
heiko1the eigenspectrum of the covariance matrix16:57
blackburnmy kernel is -1.0 constantly16:57
heiko1lol :)16:57
blackburngood kernel16:57
-!- alexlovesdata [~binder@] has joined #shogun16:59
-!- alexlovesdata [~binder@] has left #shogun []17:00
-!- alexlovesdata [~binder@] has joined #shogun17:00
-!- Jo____ [c3b2a844@gateway/web/freenode/ip.] has joined #shogun17:00
Jo____Hello, trying to use the liblinear/libsvm module and have a question if anyone has time for it.17:00
blackburnJo____: sure, ask17:00
Jo____I'm doing document classification, like sports/economy/etc. I'm fully aware how to achieve this but there's a quirk I'm having issues with. probability output.17:01
blackburnwell we don't compute probabilities right there actually17:02
blackburnso you solve multiclass problems, right?17:02
Jo____So far I've used liblinear, this limits me to logic regression, in svm there's more options but not sure of "best practices".17:02
Jo____It's in essence a multiclass problem yes.17:03
blackburnprobabilistic output can be a problem there..17:03
Jo____Further down the pipe these values are needed though so not an option to eliminate it regardless of how good the classifier gets :(17:04
blackburnheiko1: what would you say there?17:05
heiko1Jo, so you just want to get probability for multiclass SVM right?17:06
Jo____Pretty much yes.17:06
heiko1we dont have this in shogun yet17:07
heiko1however, you can compute that by hand17:07
heiko1let me find you a paper where this is described, I have one somewhere17:07
Jo____ok, thanks!17:07
heiko1Milgram, J., Cheriet, M., Sabourin, R., and Others (2006). "One Against One" or "One Against17:08
heiko1All": Which One is Better for Handwriting Recognition with SVMs? Strategy17:08
heiko1you would have to fit sigmoids to all involved binary svms in order to get binary probabilities17:08
heiko1and with this paper you can combine these to get multiclass probabilities17:08
heiko1Its on my todo list to implement into shogun17:08
heiko1(at least the sigmoid fitting)17:08
heiko1is it urgent?17:09
Jo____Ok I'll give it an read for sure, think alot of people would apriciate that too.17:09
Jo____Isn't it always but not to worry will work around. In the original liblinear/libsvm package there's some support though right?17:10
heiko1sorry phone 2min17:12
Jo____np random people taking time to help me with my little issues, grateful.17:13
-!- gsomix [~gsomix@] has joined #shogun17:14
heiko1yeah I know that this is needed :)17:15
heiko1blackburn, do you think its instantly possible to do this within the current framework?17:15
blackburnheiko1: I do not really know what's the way17:16
heiko1I mean computing apply() on all data and then access all svm scores (not just the sign) to estimate a probability which is then saved in the label17:16
heiko1this is first step17:16
blackburnproblem is the confidence17:16
blackburnis not a vector17:16
heiko1second step is (when this is done for all machines in OvO or OvR17:16
blackburnfor each feature vector I mean17:17
heiko1Currently, in apply() one can collect all svm scores right?17:17
heiko1for all data17:17
blackburnyes, it is done17:17
blackburnbut they are lost later17:18
heiko1once one has these17:18
heiko1(for a single svm)17:18
heiko1one can estimate a probability using this plat sigmoid fitting algo17:18
heiko1this is then stored per machine per datum17:18
blackburnyeah I think so17:18
heiko1and then once all this is done, these can easily combined to one multiclass probability for all data17:19
heiko1Ok, I will have a look later17:20
heiko1blackburn, did you see this guy wanting to have MKL weights in each cross-validation fold saved?17:20
-!- uricamic [~uricamic@2001:718:2:1634:3078:70c3:e281:1376] has quit [Quit: Leaving.]17:26
blackburn 28 ????????uint8_t* avec = ((CDenseFeatures<uint8_t>*)lhs)->get_feature_vector(i,alen,afree);17:34
blackburn 29 ????????uint8_t* bvec = ((CDenseFeatures<uint8_t>*)rhs)->get_feature_vector(i,blen,bfree);17:34
blackburnfind mistake :D17:34
heiko1no diea17:34
blackburnheiko1: i,i :(17:34
-!- gsomix_ [~gsomix@] has joined #shogun17:35
heiko1Is it working then now blackburn? :D17:37
-!- gsomix [~gsomix@] has quit [Ping timeout: 265 seconds]17:38
blackburnheiko1: it says 37 minutes remaining - that's why I hate kernel methods :D17:45
heiko1whats that for? 37 minutes :)17:49
-!- gsomix_ [~gsomix@] has quit [Ping timeout: 244 seconds]17:54
blackburnheiko1: to train on 4300 feature vectors :(17:56
blackburnsomething is totally wrong here17:56
blackburnah yes it is a brain related problem17:57
-!- puffin444 [62e3926e@gateway/web/freenode/ip.] has joined #shogun17:58
blackburnterrible brain damage of mine I'd say :D17:58
blackburnbut I must say one can cast DenseFeatures<float64_t> to DenseFeatures<uint8_t> and it works *somehow*17:59
alexlovesdata:D blackburn18:01
blackburnalexlovesdata: may be you know how to handle binary descriptors in recognition?18:01
alexlovesdatayou discover denoising by compression ?!18:02
blackburnalexlovesdata: no I mean it doesn't fail neither runtime nor compilation18:02
alexlovesdatawhta means: to handle binary descriptors?18:02
alexlovesdatainteger math?18:02
blackburnalexlovesdata: well I am trying to make FREAK work with road signs18:03
alexlovesdatathe normal way is to do convex relaxation from the hypercube edges to the linear domain\18:03
blackburnalexlovesdata: so I have 512 bit features18:03
blackburnalexlovesdata: earlier I was just mapping each bit to float (1.0 or 0.0) and learn a linear model18:04
alexlovesdatayou can encode the 512 bit in one float18:04
blackburnalexlovesdata: ehmm?18:04
alexlovesdataso a proper machine can use the value for learning18:05
blackburn8 floats you mean?18:05
alexlovesdataI mean 512 bit encode a numerical value range18:05
blackburnalexlovesdata: do you think it have a chance to work as feature?18:06
alexlovesdatayou can assign each possible number a float value18:06
alexlovesdatathen you have a one dim ... or N-dim feature18:06
alexlovesdatain which each dim captures one bit range18:06
alexlovesdataas you said eg N=512/32 or whatever18:07
alexlovesdataso that the range covered by one float is reasonable18:07
alexlovesdatait is one form of embedding the thing18:07
-!- gsomix [~gsomix@] has joined #shogun18:08
gsomixthunderstorm - horses 1:018:08
alexlovesdatawith N=512/32 each dim encodes a value range of 2^32 or so18:08
blackburnalexlovesdata: so you suggest to construct floats from bits18:08
blackburnand use it as features?18:08
alexlovesdatawhy not ?!18:09
-!- Jo____ [c3b2a844@gateway/web/freenode/ip.] has quit [Ping timeout: 245 seconds]18:09
alexlovesdataN should be large enough so that it is reasonably separable18:09
blackburnalexlovesdata: sounds like something crazy actually :)18:10
alexlovesdatacrazy things often come out of my mind ... O:-)18:10
alexlovesdatawe can talk about Pascal VOC later18:10
alexlovesdatamaybe mid of next week?18:11
alexlovesdataI need to code some experimental heatmap stuff18:11
blackburnalexlovesdata: yeah, I don't mind18:11
alexlovesdataright now18:11
blackburnI will try that 'embedding' now18:12
blackburnalexlovesdata: no that works worse than just bits as features18:29
heiko1blackburn, what about the online pdf of the tutorial, any news on that?18:30
blackburnheiko1: well no, I don't have any news18:30
heiko1sonney2k, ^18:31
blackburnalexlovesdata: 15% accuracy lossss18:32
blackburnalexlovesdata: no, with a few keypoints it is surprisingly pretty near18:38
blackburnalexlovesdata: I did 0/1 to kind of ++/-- mapping - works surprisingly well18:47
-!- alexlovesdata [~binder@] has quit [Ping timeout: 245 seconds]19:02
-!- alexlovesdata [~binder@] has joined #shogun19:06
alexlovesdatawhich kernel did you use?19:06
blackburnalexlovesdata: no kernel19:07
alexlovesdataI mean after the recoding19:07
alexlovesdataor what classification method?19:07
blackburnlinear crammer-singer19:07
alexlovesdataso linear19:07
blackburnyes sure19:07
alexlovesdatawhy linear???19:07
alexlovesdataI would say this is always the worst choice19:07
blackburnI won't be able to try anything with kernels19:08
alexlovesdatanot even explicit mappings?19:08
blackburntrainset is pretty large19:08
blackburnah with explicit mapping I should try19:08
heiko1blackburn, or anyoidy else, is there an online (latex) generator for doxygen comments? such as online latex renderes?19:08
blackburnheiko1: no I don't know any19:09
-!- alexlovesdata [~binder@] has left #shogun []19:22
CIA-18shogun: Heiko Strathmann master * r5832250 / src/shogun/mathematics/Statistics.h : fixed some warnings and documentation updates in doxygen latex -
CIA-18shogun: Heiko Strathmann master * r8948ff2 / src/shogun/statistics/TwoDistributionsTestStatistic.h : fixed a warning -
CIA-18shogun: Heiko Strathmann master * r7b2a2b8 / (2 files in 2 dirs): Merge pull request #701 from karlnapf/master -
-!- heiko1 [] has left #shogun []19:30
-!- n4nd0 [] has joined #shogun19:44
n4nd0nice, I see that the buildbot is healthy again19:46
n4nd0sonney2k, blackburn: should I do something else about the PR before it gets merged?19:48
blackburnn4nd0: I don't know , let me check19:48
n4nd0I updated it and rebased master yesterday IIRC19:49
blackburnn4nd0: oh so you did matrix list structure?19:52
blackburndid add19:52
n4nd0the captain told me to do it :)19:53
blackburnit is pretty huge to review ^(19:55
n4nd0I think that sonney2k has already reviewed it19:56
blackburnyeah sure19:56
n4nd0I am eager to see it merged :)19:57
blackburnit can't be merged :D19:57
blackburnneed to be rebased19:58
n4nd0whatever :P19:58
blackburnlet me know when it is rebased19:59
n4nd0I am still looking for some application of this I have done19:59
blackburnyeah would be nice to have some example20:00
-!- n4nd0 [] has quit [Quit: leaving]20:10
CIA-18shogun: Sergey Lisitsyn master * rc3843f9 / (2 files): Reformatted HKM -
shogun-buildbot_build #300 of deb3 - modular_interfaces is complete: Failure [failed test ruby_modular]  Build details are at  blamelist: Heiko Strathmann <>21:12
shogun-buildbot_build #301 of deb3 - modular_interfaces is complete: Failure [failed test ruby_modular]  Build details are at  blamelist: Sergey Lisitsyn <>21:51
CIA-18shogun: Sergey Lisitsyn master * rdf8cd27 / (14 files in 6 dirs): A few warning fixes -
shogun-buildbot_build #302 of deb3 - modular_interfaces is complete: Success [build successful]  Build details are at
CIA-18shogun: Sergey Lisitsyn master * r356cfcf / src/shogun/transfer/multitask/MultitaskCompositeMachine.cpp : Fixed MultitaskCompositeMachine train locked -
--- Log closed Fri Aug 10 00:00:17 2012