--- Log opened Tue Jan 03 00:00:19 2012 | ||
-!- ishaanmlhtr [~ishaan@115.240.71.113] has quit [Quit: Leaving] | 00:07 | |
-!- blackburn [~qdrgsm@109.226.69.156] has quit [Quit: Leaving.] | 01:17 | |
-!- ishaanmlhtr [~ishaan@115.241.218.168] has joined #shogun | 07:17 | |
-!- ishaanmlhtr [~ishaan@115.241.218.168] has quit [Ping timeout: 240 seconds] | 09:28 | |
-!- ishaanmlhtr [~ishaan@115.241.190.226] has joined #shogun | 09:41 | |
-!- ishaanmlhtr [~ishaan@115.241.190.226] has quit [Quit: Leaving] | 09:47 | |
-!- in3xes [~in3xes@49.249.163.29] has joined #shogun | 09:51 | |
-!- in3xes [~in3xes@49.249.163.29] has quit [Ping timeout: 252 seconds] | 10:25 | |
-!- in3xes [~in3xes@49.249.160.171] has joined #shogun | 10:37 | |
-!- in3xes [~in3xes@49.249.160.171] has quit [Ping timeout: 252 seconds] | 11:24 | |
-!- ishaanmlhtr [~ishaan@115.241.190.226] has joined #shogun | 12:42 | |
-!- ishaanmlhtr [~ishaan@115.241.190.226] has quit [Ping timeout: 260 seconds] | 13:50 | |
-!- ishaanmlhtr [~ishaan@115.240.126.161] has joined #shogun | 14:29 | |
-!- ishaanmlhtr [~ishaan@115.240.126.161] has quit [Client Quit] | 14:32 | |
-!- ishaanmlhtr [~ishaan@115.240.126.161] has joined #shogun | 14:53 | |
-!- blackburn [~qdrgsm@109.226.69.156] has joined #shogun | 15:18 | |
ishaanmlhtr | blackburn : there? | 15:34 |
---|---|---|
blackburn | ishaanmlhtr: yes | 15:35 |
ishaanmlhtr | i just trained the LibSVMMulticlass classifier and applied it to my test data | 15:36 |
ishaanmlhtr | the accuracy came out to be 0.121666666667 | 15:36 |
ishaanmlhtr | can it be improved somehow? | 15:37 |
ishaanmlhtr | took C=1 | 15:37 |
ishaanmlhtr | and width =2.1 | 15:37 |
ishaanmlhtr | is this accuracy good enough? or something can be done about it> | 15:38 |
blackburn | no, it is bad for sure | 15:38 |
blackburn | what are sizes of train/test set? | 15:38 |
ishaanmlhtr | 600 each with 28*28 | 15:39 |
blackburn | how many examples of each class? | 15:40 |
ishaanmlhtr | I would have to find a way to see that. | 15:40 |
ishaanmlhtr | I'll tell u in a while | 15:41 |
blackburn | for label in set(labels): | 15:41 |
blackburn | print label, labels.count(label) | 15:41 |
blackburn | something like that | 15:41 |
ishaanmlhtr | ya,ok | 15:41 |
blackburn | 12% is pretty useless classifier, something goes wrong here | 15:43 |
ishaanmlhtr | ok..at least 50 of each are there | 15:47 |
blackburn | ishaanmlhtr: you should try different Cs and widths | 15:51 |
ishaanmlhtr | ok, any particular range to try in? | 15:51 |
blackburn | something like C: 1e-2, 1e-1, 1, 10, 100, 1000 | 15:52 |
ishaanmlhtr | blackburn : i am getting the current prediction as all belonging to class with most examples | 15:52 |
blackburn | how many example have this class? | 15:53 |
blackburn | examples* | 15:53 |
blackburn | has* | 15:53 |
blackburn | :D | 15:53 |
ishaanmlhtr | 79 | 15:53 |
ishaanmlhtr | rest none has above 65 | 15:53 |
blackburn | ah *all* predictions& | 15:54 |
blackburn | ? | 15:54 |
ishaanmlhtr | blackburn : i din't get you. | 15:54 |
blackburn | are all predictions of svm identical? | 15:55 |
ishaanmlhtr | blackburn : ya , all belonging to class 1 | 15:55 |
ishaanmlhtr | which has around 79 examples in the trained set | 15:55 |
blackburn | is it a first class? | 15:56 |
ishaanmlhtr | the 2nd | 15:56 |
ishaanmlhtr | first is the one with label =0 having 58 examples | 15:56 |
blackburn | try C=1000 | 15:56 |
ishaanmlhtr | ok, | 15:56 |
blackburn | and width 1,10,100,1000,10000 | 15:57 |
blackburn | do you normalize images? is it in range 0-255? | 15:57 |
blackburn | I would suggest to normalize it with dividing / 255 | 15:58 |
ishaanmlhtr | i dint do any normalization as such . but ya, they were in the range 0-255 | 15:58 |
ishaanmlhtr | ok | 15:58 |
ishaanmlhtr | i would try all these options. | 15:58 |
blackburn | it should be better this time | 15:59 |
ishaanmlhtr | ok. | 15:59 |
ishaanmlhtr | i'll try and get back to you. | 15:59 |
blackburn | ok | 15:59 |
ishaanmlhtr | getting 87.83% | 16:10 |
ishaanmlhtr | blackburn : i am getting 87.83% | 16:11 |
blackburn | ishaanmlhtr: much better now :) | 16:11 |
ishaanmlhtr | C=100,1000 width =100 | 16:11 |
ishaanmlhtr | yaa..:) | 16:11 |
ishaanmlhtr | but not able to improve beyond it by changing whatever values of C and width | 16:11 |
ishaanmlhtr | How did that normalization help out btw? | 16:12 |
blackburn | ishaanmlhtr: well it is a common practice | 16:13 |
blackburn | numerical issues, etc | 16:13 |
blackburn | and width of kernel depends on normalization too | 16:13 |
ishaanmlhtr | ok. and this changing of the values of C and width - do we have a better way of cross validation | 16:14 |
ishaanmlhtr | instead of just changing these values and testing | 16:14 |
ishaanmlhtr | ? | 16:14 |
blackburn | ishaanmlhtr: yes | 16:15 |
blackburn | it was a project of Heiko Strathmann this summer :) | 16:15 |
ishaanmlhtr | ok..the documentation is available? | 16:16 |
blackburn | I'm afraid no good documentation yet | 16:16 |
blackburn | check modelselection_* examples | 16:17 |
ishaanmlhtr | ok,fine | 16:17 |
ishaanmlhtr | and what else can i do now except training more and more classifiers on this MNIST database | 16:18 |
ishaanmlhtr | ? | 16:18 |
ishaanmlhtr | can you suggest me something?again for learning purposes only | 16:18 |
blackburn | well you can clean up your application and get it into shogun/applications | 16:18 |
blackburn | or yes I could come with something | 16:19 |
ishaanmlhtr | the MNIST one? | 16:19 |
ishaanmlhtr | ok,i could do that!..:) | 16:20 |
blackburn | unfortunately I guess we are not able to include MNIST dataset | 16:21 |
blackburn | so you would have to write a downloader too | 16:21 |
ishaanmlhtr | ok,that means automatically the files get downloaded from the source page when i call upon my application? | 16:22 |
ishaanmlhtr | and then the processing starts? | 16:23 |
blackburn | well yes, but once | 16:23 |
ishaanmlhtr | hmm..alright,i could give it a try for sure. | 16:25 |
-!- ishaanmlhtr [~ishaan@115.240.126.161] has quit [Quit: Leaving] | 16:46 | |
-!- ishaanmlhtr [~ishaan@115.240.126.161] has joined #shogun | 17:12 | |
-!- ishaanmlhtr [~ishaan@115.240.126.161] has quit [Client Quit] | 17:12 | |
CIA-1 | shogun: Sergey Lisitsyn master * r87ba5fe / src/shogun/classifier/svm/SVMOcas.cpp : Fixed memsetting at SVM OCAS - http://git.io/ENoHpg | 19:42 |
shogun-buildbot | build #92 of nightly_default is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_default/builds/92 | 21:19 |
shogun-buildbot | build #106 of nightly_all is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_all/builds/106 | 21:32 |
--- Log closed Wed Jan 04 00:00:19 2012 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!