How to run libsvm in MATLAB?

Question

How to get started with libsvm under MATLAB?

I've downloaded the library, and extracted it in C:\Program Files\MATLAB\R2012a\toolbox\, but then I don't know how to use it in MATLAB.

Solution 1

Download and extract libsvm in a directory of your choosing, say C:\libsvm

As described in the C:\libsvm\matlab\README file, first you have to make sure a supported C/C++ compiler is installed. Note that on 64-bit systems, you need the correct 64-bit version of the compiler (e.g. Windows SDK is needed for Visual Studio Express edition)

>> mex -setup

Once you have selected a compiler, you need to compile the MEX-files:

>> cd('C:\libsvm\matlab')
>> make

Finally add the folder with the generated binaries to the MATLAB search path:

>> addpath('C:\libsvm\matlab')

Test the library with a simple example (fake data):

>> labels = double(rand(10,1)>0.5);
>> data = rand(10,5);
>> model = svmtrain(labels, data, '-s 0 -t 2 -c 1 -g 0.1')

Note that the current version of libsvm includes pre-compiled 64-bit MEX-files for Windows. The binaries are located in C:\libsvm\windows\*.mexw64 (copy those to the matlab subfolder from above)

Solution 2

For me I didn't need to recompile the libsvm files(it did cause some problems with the .net framwork and windows SDK) I only used the already compiled files and added them to a new folder by following the steps mentioned here minus the make step.

So to summarize:

1- I think you need to create "libsvm" folder under "C:\Program Files\MATLAB\R2014b\toolbox\".

2- Then copy the *.mexw64 files from the "libsvm-3.21\windows" folder to the new folder.

3- finally add the libsvm folder you just created to matlab bath by clicking the set path button in home and adding the new folder with the path "C:\Program Files\MATLAB\R2014b\toolbox\libsvm"

Comments

[Khalid Usman]

Thanks alot, it worked for me for Windows OS. Can you please help me to run the same thing for Mac OS?

[kAmol]

@Amro Add some more details about testing, coz I'm facing problems with option -b i.e. probability. The result is showing same prob distribution among 5 classes for all tests.