Open in new window / Try shogun cloud
--- Log opened Sun Aug 26 00:00:17 2012
CIA-52shogun: Sergey Lisitsyn master * r9429c0e / src/shogun/evaluation/CrossValidationMKLStorage.cpp : Added handling of multiclass MKL machine in MKL storage -
yooblackburn: allright, going to test it00:10
CIA-52shogun: Sergey Lisitsyn master * rb35dbe8 / src/shogun/evaluation/CrossValidationPrintOutput.cpp : Added MKL multiclass handling in CV print output -
yooblackburn: could you explain me again why not using the Evaluation interfaces such as MulticlassAccuracy and MulticlassOVR ?00:13
blackburnyoo: what's the difference?00:13
yooblackburn: yes thats it: it is exactly the same right ?00:14
blackburnyeah, but it already have to do the same as MulticlassOVR00:14
blackburnso I just put binary evaluations there00:14
blackburnargh so we need a matrix to store accuracies too?00:15
yooisnt already here ? SGVector<float64_t> m_evaluations_results;00:16
blackburnno it is for binary00:16
yooah ok then we need multiclass acc (and confusion matrices)00:17
blackburnoh gosh00:17
blackburnconfusion matrices?00:17
yoothats why I ask00:17
yoomulticlassOVR and multiclass Acc00:17
yoopermits to choose what you want00:17
yooand not compute everythg00:18
yooI liked the way modelselecion_output worked00:18
yoomsout.add(mc_acc) or msout.add(mc_ovr(roc))00:18
blackburnconfusion matrices have to be handled in special way still00:19
yooie ?00:19
blackburnyou wanted to store it, right?00:19
blackburnokay I will add it as option00:19
yoothen we will have all the mc evaluation possibly stored00:20
-!- gsomix [~gsomix@] has quit [Ping timeout: 246 seconds]00:23
blackburnokay I will add three options00:23
blackburnto constructor00:23
blackburnROC, PRC and confusion matrices00:24
blackburnI am a little tired with that code so I will make it constant :)00:24
blackburnI mean not possible to change after calling constructor00:24
blackburnit requires special handling00:25
yooI begin to read easily the shogun code then I will be able to help with dirty task if you have some to give00:25
blackburnoh I will think about some task to get into development00:26
yoobut you can call several options in the constructor right ?00:26
yooand you check with dirty dynamic_cast ?00:26
blackburncheck what?00:26
yooROC or PCR : the BinaryEval type00:26
yoothe Multiclass Eval type I mean :p00:27
blackburnah in dynamic object array?00:27
blackburnwell I assume nobody breaks something wrong into that00:28
blackburnkind of ugly code that is00:30
yoo[100%] Built target shogun00:34
yoo =)00:34
yoowell I have to work with cpack for installation, didnt have manage the interfaces .. but this will probably work in the future00:34
CIA-52shogun: Evgeniy Andreev master * rab15b5f / examples/undocumented/python_modular/ : added example for DirectorDotFeatures -
CIA-52shogun: Evgeniy Andreev master * rab6a200 / (7 files in 3 dirs): protocols for CustomKernel -
CIA-52shogun: Sergey Lisitsyn master * r65ad845 / (8 files in 4 dirs): Merge pull request #757 from gsomix/buffer_protocol -
CIA-52shogun: Sergey Lisitsyn master * r27c7dd2 / (2 files): Added confidence matrix computation abilities to CV MC storage and -
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]01:16
-!- yoo [] has quit [Quit: Ex-Chat]02:12
shogun-buildbot_build #468 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Evgeniy Andreev <>, Sergey Lisitsyn <>02:42
shogun-buildbot_build #77 of nightly_default is complete: Failure [failed test]  Build details are at
-!- emrecelikten1 [] has quit [Ping timeout: 244 seconds]04:24
-!- emrecelikten [] has joined #shogun04:37
-!- in3xes [~in3xes@] has quit [Ping timeout: 245 seconds]05:27
-!- in3xes [~in3xes@] has joined #shogun06:18
-!- in3xes [~in3xes@] has quit [Ping timeout: 264 seconds]06:24
CIA-52shogun: Soeren Sonnenburg master * rf6aaf00 / examples/undocumented/python_modular/ : return none when numpy version etc is not sufficiently new -
-!- sr___ [u5548@gateway/web/] has quit [Quit: Connection closed for inactivity]07:25
shogun-buildbot_build #469 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Soeren Sonnenburg <>07:36
-!- gsomix [~gsomix@] has joined #shogun08:06
-!- gsomix [~gsomix@] has quit [Ping timeout: 268 seconds]08:18
-!- n4nd0 [] has joined #shogun08:29
n4nd0wiking: around?08:30
-!- n4nd0 [] has quit [Quit: leaving]10:10
-!- blackburn [~blackburn@] has joined #shogun10:11
CIA-52shogun: Sergey Lisitsyn master * r84fe14c / tests/regression/r_static/preprocessor.R : Fixed naming in R regression preprocessor test -
CIA-52shogun: Sergey Lisitsyn master * r4a9f485 / (3 files in 2 dirs): Fixed memory handlnig in CV MC storage -
CIA-52shogun: Sergey Lisitsyn master * rd1d57a1 / src/shogun/lib/slep/slep_mc_tree_lr.cpp : Removed scaling in multiclass tree guided LR -
CIA-52shogun: Sergey Lisitsyn master * rad6f6dc / (3 files): Marked string kernels failing on regression tests as unstable -
shogun-buildbot_build #470 of deb3 - modular_interfaces is complete: Failure [failed test python_modular]  Build details are at  blamelist: Sergey Lisitsyn <>11:12
-!- in3xes [~in3xes@] has joined #shogun11:33
shogun-buildbot_build #471 of deb3 - modular_interfaces is complete: Success [build successful]  Build details are at
-!- blackburn [~blackburn@] has left #shogun []12:26
-!- yoo [575b08cb@gateway/web/freenode/ip.] has joined #shogun12:29
-!- yoo [575b08cb@gateway/web/freenode/ip.] has quit [Quit: Page closed]12:54
-!- n4nd0 [] has joined #shogun13:06
n4nd0someone out there :)?13:08
-!- yoo [575b08cb@gateway/web/freenode/ip.] has joined #shogun13:09
-!- yoo [575b08cb@gateway/web/freenode/ip.] has quit [Client Quit]13:09
n4nd0emrecelikten: hehe ok13:11
n4nd0do you happen to know about structured output learning algorithms?13:12
emreceliktenNo :/13:12
n4nd0ok, no problem ;)13:12
n4nd0I was just wondering what's the state of the art13:12
n4nd0someone will be able to answer me around here sooner or later, there are some experts13:12
-!- bern4rd [1f22c5b5@gateway/web/freenode/ip.] has joined #shogun13:37
n4nd0hey bern4rd13:43
n4nd0I don't think people are around here right now today13:43
bern4rdah ok, no problem13:44
n4nd0bern4rd: tomorrow I am starting pattern recognition :)13:45
n4nd0bern4rd: are you taking that course at the end?13:45
bern4rdme too13:45
-!- Marty28 [9eb54d46@gateway/web/freenode/ip.] has joined #shogun13:47
n4nd0bern4rd: I'll update you when you come ;)13:47
bern4rdnice :) Anyway 29 I end the internship so i'll be available13:48
Marty28Can you tell me where I can find a documentation for the data in shogun-data-0.3 ?13:49
n4nd0Marty28: for the release of that version?13:50
n4nd0to tell the truth I am not sure there's documentation for the data part of the project ...13:50
n4nd0at least I don't know about it existence13:50
n4nd0Marty28: what do you want to know in any case? I might be able to help :D13:51
Marty28It is about /asp13:51
Marty28the folder13:52
Marty28seems to be splicing data13:52
Marty28yet no inline documentation13:52
Marty28legacy, probably13:53
n4nd0let's check in the commit messages13:54
n4nd0nothing there13:54
Marty28seems to be truncated anyway13:57
Marty28says %asplicer definition file version: 1.0  %acceptor splice acc_splice_b=-2.867314e+00 acc_splice_order=22 acc_splice_window_left=60 acc_splice_window_right=79 acc_splice_alphas=[1.408119e+00, 7.82051 acc_splice_svs=[13:57
Marty28I will ask R?tsch, I bet it is their data.13:58
n4nd0yes I think so13:59
n4nd0have you taken a look to msplice?13:59
n4nd0maybe it is related to it13:59
n4nd0sorry mgene14:00
-!- in3xes [~in3xes@] has quit [Remote host closed the connection]14:03
-!- n4nd0 [] has quit [Quit: leaving]14:04
-!- bern4rd [1f22c5b5@gateway/web/freenode/ip.] has quit [Quit: Page closed]14:06
-!- Marty28 [9eb54d46@gateway/web/freenode/ip.] has quit [Quit: Page closed]14:07
-!- blackburn [~blackburn@] has joined #shogun16:21
-!- K0stIa [] has joined #shogun16:42
-!- K0stIa [] has left #shogun []16:42
--- Log closed Sun Aug 26 17:39:59 2012
--- Log opened Sun Aug 26 17:51:28 2012
-!- shogun-toolbox [] has joined #shogun17:51
-!- Irssi: #shogun: Total of 11 nicks [1 ops, 0 halfops, 0 voices, 10 normal]17:51
-!- Irssi: Join to #shogun was synced in 7 secs17:51
-!- shogun-buildbot [] has joined #shogun17:52
n4nd0hey blackburn17:52
blackburnhey n4nd017:52
n4nd0how are you doing?17:52
n4nd0any news with the segmentation application?17:52
blackburnfine, and you?17:52
blackburnno i didn't make any progress yet17:52
blackburnI did implement director model17:53
blackburnhave you seen?17:53
n4nd0I didn't17:53
n4nd0but cool :)17:53
blackburnit works17:53
blackburnI added code and python example yesterday17:53
n4nd0let me check17:53
blackburnperfomance wise it sucks17:54
blackburnbut still17:54
n4nd0very different from the C++ one?17:55
blackburnwell multiclass generic training is not efficient already17:56
blackburnbecause of big sparse vectors17:56
blackburnand swig stuff makes it pretty slow as well17:57
n4nd0how is it possible that you didn't have to define get_joint_feature_vector?17:57
n4nd0aham there is no need for it in the BMRM algorithm I guess17:58
blackburnI made that explicitly in17:58
blackburnBMRM needs only risk17:58
blackburnand risk needs only argmax17:58
n4nd0I see17:58
n4nd0I am not sure if BMRM uses argmax too17:58
blackburnand that's actually nice17:58
n4nd0apart from inside risk17:58
blackburnBMRM uses only risk17:58
blackburnbut you would need joint feature vectors when computing gradients17:59
n4nd0gradients where?18:00
blackburnn4nd0: in risk18:05
blackburnto be correct18:05
n4nd0blackburn: aham ok, but thanks to generic risk one doesn't need to define it explicitily either18:08
blackburnn4nd0: what should I define to use HM-SVM?18:11
blackburnwhat do matrix features contain in means of HM ?18:12
n4nd0the matrix features contain the observations18:13
blackburnso if I have observations18:14
blackburn[0.3, 0.2, 0.1]18:14
blackburn[0.5, 0.5, 0.9]18:14
blackburnwhat would be the matrix?18:14
blackburn^ or transposed ^?18:14
n4nd0ok let's see18:15
n4nd0you should first tell me how many features you have18:15
n4nd0and time steps18:15
n4nd0from what you have written there18:15
blackburnokay see what I mean18:16
blackburnwhen we are segmenting image18:16
n4nd0I would guess for 1 feature, 4 time steps and two feature vectors18:16
blackburnwe go through all image18:16
blackburnobservations are 9 number18:16
blackburnpixel values18:16
blackburnof the pixel and its neighborhood18:16
blackburnsee what I mean?18:16
n4nd0so for every pixel you get 9 values?18:17
blackburnyes exactly18:17
blackburnassume image is square18:17
blackburnit  requires N*N timesteps18:17
blackburngot it?18:17
n4nd0aham ok18:17
n4nd0but the point is that with just an image18:18
n4nd0I don't really see where is the time dimension18:18
blackburnn4nd0: you just iterating over all pixels18:18
n4nd0ok got it18:21
blackburnso all I need is to define that matrix and sequence18:22
n4nd0wait a sec18:22
n4nd0what matrix and seq? :D18:23
blackburnobservation matrix and sequence of {foreground,background}18:23
blackburnoh nice18:23
blackburnokay will be done soon then18:23
n4nd0the sequence represent the labels18:24
blackburnis there an interface to construct labels?18:24
blackburnin python18:24
n4nd0do you want to discuss something more about the matrix features?18:24
blackburnn4nd0: last thing not clear for me is18:24
blackburnwhat is #rows and #cols18:24
n4nd0blackburn: to construct the labels just create an HMSVMLabels instance18:24
n4nd0you should be able to do it giving an SGVector in the constructor, i.e. an nparray18:25
n4nd0for python I meant18:25
blackburnso I would need to put18:25
blackburn0 and 118:25
n4nd0ok, now about the matrix features18:26
n4nd0the matrix features internally is an SGMatrixList, a list of matrices18:26
n4nd0each matrix in this list represents the observations of an image, ok?18:26
n4nd0if you want to train your HM-SVM using 100 images, your matrix features will be composed of 100 matrices18:27
blackburnright sure18:27
n4nd0all right18:27
blackburnthat's perfect clear18:27
n4nd0now, for each matrix, what are the rows and columns?18:27
n4nd0the number of columns is the same dimension of the sequence18:28
n4nd0the time of this time series18:28
n4nd0you see what I mean?18:29
n4nd0it was not a very good description to tell the truth ....18:29
blackburnno it is good18:29
blackburnso in case I have 200x200 image18:29
blackburnI've got 40000 cols18:29
blackburneach containing say 9 observation values18:29
blackburnuh I will try on 50x50 first18:30
n4nd0I think that should be your feature dimension18:30
n4nd0got it?18:31
blackburnyeah I think I did18:31
blackburnlets try to code it18:31
n4nd0and let's hope it is a good model for this problem :)18:32
n4nd0it would be really cool if it works18:32
n4nd0I wonder how people find these things out18:33
blackburnCRF works for that very nice18:33
n4nd0have you taken the inspiration of the features from it?18:35
blackburnno I'll try pixels first18:35
blackburnit is easy to change features afterwards18:35
blackburnn4nd0: can state be unknown?18:40
blackburnkind of latent state18:40
blackburnwhen you have no idea whether it is a background or a foreground pixel18:41
n4nd0mmm no with the state model we use18:41
blackburnyes, I mean in theory18:41
n4nd0yeah in theory18:42
n4nd0I mean, you could add just one state there, unknown18:42
blackburnbut it would be predicted then18:42
n4nd0not really18:42
n4nd0you could do an state model18:43
n4nd0with this unknown state18:43
n4nd0you model the possibility of going from BG or FG to this unknown state18:43
n4nd0but you don't model the observations from this model18:43
n4nd0it is something similar to the start and stop states that are internally used in the TwoStateModel18:43
blackburnI see18:44
n4nd0but you don't model the observations from this state18:44
n4nd0I said "from this model" before, that was wrong18:44
blackburnyes I got it18:44
n4nd0but this should work fine with some noise in there18:44
n4nd0so I think you can still use the TwoStateModel and maybe it is not affected badly18:45
blackburnokay lets see what it would be18:45
n4nd0what kind of images are you using?18:47
blackburncat :D18:47
blackburnone image yet18:47
n4nd0isn't it a pain to "label" it?18:48
blackburnI asked my gf before18:48
n4nd0I mean to generate this background/foreground vector18:48
n4nd0did she do it?18:49
blackburnyes for one image18:49
blackburnit is not that hard18:49
n4nd0not hard but kind of pain in the ass to do it18:50
n4nd0I guess it depens on the method too18:50
blackburnno not really, takes 3-4 minutes18:51
n4nd0aham not that bad then18:51
n4nd0do you use something like a GUI?18:51
blackburnI am going to teach HM model on that image and predict18:52
blackburnand we will see what it models18:52
n4nd0just for an image?18:53
n4nd0let's see18:53
blackburnn4nd0: hmsvmlabels segfault on adding if default constructor is called18:55
blackburnone should be careful :)18:55
n4nd0probably it is because no memory is allocated for the DynamicObjectArray in StructuredLabels18:59
blackburnn4nd0: is loss relevant?19:00
n4nd0could be19:00
n4nd0we just have hinge loss this far though19:01
blackburnI see19:01
blackburnSystemError: Out of memory error, tried to allocate 7378945280 bytes using malloc.19:02
n4nd0really? just with one image19:03
n4nd0it looks weird to me19:03
blackburnsomething is wrong I believe ;)19:03
n4nd0yeah I think so19:03
n4nd0how big is the image?19:03
blackburn50 x 50 now19:04
n4nd0there shouldn't be a memory error for that indeed19:04
blackburnI get 019:05
n4nd0not good19:06
n4nd0what training algorithm?19:06
n4nd0send me the code and I can try to train it with PrimalMosekSOSVM19:06
blackburnagain -19:06
blackburn# of cols19:07
blackburn= time dimension19:07
n4nd0yes, let me check just in case19:07
n4nd0in this case it should be 250019:07
blackburnI do not use border pixels19:07
blackburnso feature matrix19:08
blackburn(2304, 9)19:08
blackburnis that correct?19:08
n4nd0number of columns is the 2nd dimension19:08
n4nd0shouldn't it be the other way?19:08
blackburnI am a little mixed19:09
n4nd0num_rows = 919:09
n4nd0num_cols = 230419:09
blackburnthat's how it should be?19:09
blackburnno it is now19:09
n4nd0the length of the state sequence19:10
n4nd0should be equal to the second dimension of each element in matrix features19:10
blackburnnot zero now19:12
n4nd0what was it?19:13
blackburnhow to extract sequence now19:13
blackburnfrom a prediction19:13
n4nd0so apply gives you StructuredLabels right?19:14
blackburngot it19:14
n4nd0use HMSVMLabels::obtain_generic to get HMSVMLabels19:14
blackburnI don't have to19:14
n4nd0and later just do a get_label(0)19:14
blackburnI should rather obtain Sequence from generic19:14
blackburnget_label(0) works already - it is defined in structured labels19:15
n4nd0the frist obtain_generic is not really necessary19:15
n4nd0loss of time :D19:15
blackburntypemap didn't work19:15
blackburnI've got sgvector19:15
-!- in3xes [~in3xes@] has joined #shogun19:16
blackburnI will add a getter there19:17
n4nd0I did that once but didn't commit it19:18
n4nd0I thought I was doing something wrong in python19:18
n4nd0why the typemap doesn't work?19:18
blackburnno idea19:19
blackburnI hope it would be typemapped now19:22
blackburneverything is a background19:24
n4nd0too bad19:25
n4nd0but I think that the concept of training with one image is weird19:25
blackburnbut it should fit to the image, right?19:26
n4nd0if we map it to non-structured learning19:26
n4nd0I don't really know to tell the truth19:26
n4nd0from the point of view of the HM-SVM it looks like that19:26
n4nd0but if you think of the training algorithm19:26
blackburnif we train svm with one point19:26
n4nd0you just have one example19:27
blackburnit won't make an error19:27
n4nd0so it is weird19:27
n4nd0I don't think that is the right way of thinking of it ...19:27
blackburnoh I've got something19:27
blackburnsome strange pattern19:27
n4nd0in any case19:31
n4nd0I think that for one example it is not going to work pretty good19:31
n4nd0at least it gives me that feeling regarding the way in which the training algorithm works19:32
blackburnbut still19:32
n4nd0hehe you are hard to convince19:33
blackburnlambda changes the game19:33
blackburnwell I saw approximately correct result19:33
n4nd0we need some sort of model selection for that lambda very badly19:35
blackburnn4nd0: I'll try on that ^19:36
n4nd0it seems a different concept this of brackground maintenance19:37
n4nd0but probably we can use the data for our purpose?19:37
blackburnwhy different?19:48
blackburnsomething is kinda different yes19:49
blackburnnot that bad in general20:07
n4nd0just one image in training?20:08
n4nd0yes, not too bad20:08
n4nd0it will be better with more probably I guess20:08
blackburnand with better features20:08
n4nd0can you use the images there then?20:09
blackburnfrom microsoft research?20:09
blackburnI'll try later20:09
blackburnI need to learn blender, no way to get back to 3ds max I know pretty well :(20:14
blackburntime to update NEWS20:15
n4nd0for what do you want to use blender?20:16
n4nd0you don't like 3D max?20:16
blackburnto generate synthetic images20:16
blackburnI don't have windows20:16
n4nd0synthetic images?20:17
blackburnn4nd0: yes, for example some vase to segment :)20:17
blackburnor anything like that20:17
n4nd0in 3D?20:18
-!- in3xes [~in3xes@] has quit [Ping timeout: 256 seconds]20:18
n4nd0time for dinner, see you later20:18
-!- n4nd0 [] has quit [Quit: leaving]20:18
CIA-52shogun: Sergey Lisitsyn master * re0696f4 / src/interfaces/python_modular/DenseFeatures_protocols.i : Merge pull request #758 from Nightrain/testbranch -
CIA-52shogun: Sergey Lisitsyn master * r1468c7e / src/shogun/labels/MulticlassLabels.cpp : Changed get binary for class method -
shogun-buildbotbuild #475 of deb3 - modular_interfaces is complete: Failure [failed compile csharp_modular]  Build details are at  blamelist: Sergey Lisitsyn <>22:50
shogun-buildbotbuild #476 of deb3 - modular_interfaces is complete: Failure [failed compile csharp_modular]  Build details are at  blamelist: Sergey Lisitsyn <>23:06
CIA-52shogun: Sergey Lisitsyn master * r0e7695a / (4 files in 2 dirs): Fixed csharp crasher with making sequence data protected -
-!- zxtx [] has quit [Ping timeout: 248 seconds]23:56
--- Log closed Mon Aug 27 00:00:17 2012