Open in new window / Try shogun cloud
--- Log opened Tue Jan 03 00:00:19 2012
-!- ishaanmlhtr [~ishaan@] has quit [Quit: Leaving]00:07
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]01:17
-!- ishaanmlhtr [~ishaan@] has joined #shogun07:17
-!- ishaanmlhtr [~ishaan@] has quit [Ping timeout: 240 seconds]09:28
-!- ishaanmlhtr [~ishaan@] has joined #shogun09:41
-!- ishaanmlhtr [~ishaan@] has quit [Quit: Leaving]09:47
-!- in3xes [~in3xes@] has joined #shogun09:51
-!- in3xes [~in3xes@] has quit [Ping timeout: 252 seconds]10:25
-!- in3xes [~in3xes@] has joined #shogun10:37
-!- in3xes [~in3xes@] has quit [Ping timeout: 252 seconds]11:24
-!- ishaanmlhtr [~ishaan@] has joined #shogun12:42
-!- ishaanmlhtr [~ishaan@] has quit [Ping timeout: 260 seconds]13:50
-!- ishaanmlhtr [~ishaan@] has joined #shogun14:29
-!- ishaanmlhtr [~ishaan@] has quit [Client Quit]14:32
-!- ishaanmlhtr [~ishaan@] has joined #shogun14:53
-!- blackburn [~qdrgsm@] has joined #shogun15:18
ishaanmlhtrblackburn : there?15:34
blackburnishaanmlhtr: yes15:35
ishaanmlhtri just trained the LibSVMMulticlass classifier and applied it to my test data15:36
ishaanmlhtrthe accuracy came out to be 0.12166666666715:36
ishaanmlhtrcan it be improved somehow?15:37
ishaanmlhtrtook C=115:37
ishaanmlhtrand width =2.115:37
ishaanmlhtris this accuracy good enough? or something can be done about it>15:38
blackburnno, it is bad for sure15:38
blackburnwhat are sizes of train/test set?15:38
ishaanmlhtr600 each with 28*2815:39
blackburnhow many examples of each class?15:40
ishaanmlhtrI would have to find a way to see that.15:40
ishaanmlhtrI'll tell u in a while15:41
blackburnfor label in set(labels):15:41
blackburn  print label, labels.count(label)15:41
blackburnsomething like that15:41
blackburn12% is pretty useless classifier, something goes wrong here15:43 least 50 of each are there15:47
blackburnishaanmlhtr: you should try different Cs and widths15:51
ishaanmlhtrok, any particular range to try in?15:51
blackburnsomething like C: 1e-2, 1e-1, 1, 10, 100, 100015:52
ishaanmlhtrblackburn : i am getting the current prediction as all belonging to class with most examples15:52
blackburnhow many example have this class?15:53
ishaanmlhtrrest none has above 6515:53
blackburnah *all* predictions&15:54
ishaanmlhtrblackburn : i din't get you.15:54
blackburnare all predictions of svm identical?15:55
ishaanmlhtrblackburn : ya , all belonging to class 115:55
ishaanmlhtrwhich has around 79 examples in the trained set15:55
blackburnis it a first class?15:56
ishaanmlhtrthe 2nd15:56
ishaanmlhtrfirst is the one with label =0 having 58 examples15:56
blackburntry C=100015:56
blackburnand width 1,10,100,1000,1000015:57
blackburndo you normalize images? is it in range 0-255?15:57
blackburnI would suggest to normalize it with dividing / 25515:58
ishaanmlhtri dint do any normalization as such . but ya, they were in the range 0-25515:58
ishaanmlhtri would try all these options.15:58
blackburnit should be better this time15:59
ishaanmlhtri'll try and get back to you.15:59
ishaanmlhtrgetting 87.83%16:10
ishaanmlhtrblackburn : i am getting 87.83%16:11
blackburnishaanmlhtr: much better now :)16:11
ishaanmlhtrC=100,1000 width =10016:11
ishaanmlhtrbut not able to improve beyond it by changing whatever values of C and width16:11
ishaanmlhtrHow did that normalization help out btw?16:12
blackburnishaanmlhtr: well it is a common practice16:13
blackburnnumerical issues, etc16:13
blackburnand width of kernel depends on normalization too16:13
ishaanmlhtrok. and this changing of the values of C and width - do we have a better way of cross validation16:14
ishaanmlhtrinstead of just changing these values and testing16:14
blackburnishaanmlhtr: yes16:15
blackburnit was a project of Heiko Strathmann this summer :)16:15
ishaanmlhtrok..the documentation is available?16:16
blackburnI'm afraid no good documentation yet16:16
blackburncheck modelselection_* examples16:17
ishaanmlhtrand what else can i do now except training more and more classifiers on this MNIST database16:18
ishaanmlhtrcan you suggest me something?again for learning purposes only16:18
blackburnwell you can clean up your application and get it into shogun/applications16:18
blackburnor yes I could come with something16:19
ishaanmlhtrthe MNIST one?16:19
ishaanmlhtrok,i could do that!..:)16:20
blackburnunfortunately I guess we are not able to include MNIST dataset16:21
blackburnso you would have to write a downloader too16:21
ishaanmlhtrok,that means automatically the files get downloaded from the source page when i call upon my application?16:22
ishaanmlhtrand then the processing starts?16:23
blackburnwell yes, but once16:23
ishaanmlhtrhmm..alright,i could give it a try for sure.16:25
-!- ishaanmlhtr [~ishaan@] has quit [Quit: Leaving]16:46
-!- ishaanmlhtr [~ishaan@] has joined #shogun17:12
-!- ishaanmlhtr [~ishaan@] has quit [Client Quit]17:12
CIA-1shogun: Sergey Lisitsyn master * r87ba5fe / src/shogun/classifier/svm/SVMOcas.cpp : Fixed memsetting at SVM OCAS -
shogun-buildbotbuild #92 of nightly_default is complete: Success [build successful]  Build details are at
shogun-buildbotbuild #106 of nightly_all is complete: Success [build successful]  Build details are at
--- Log closed Wed Jan 04 00:00:19 2012