Home
Uni-Logo
 

Open Source Software

Binaries/Code Datasets Open Source Software
libsvmtl ImageJ-Plugins XuvTools Presto-Box iRoCS U-Net

LIBSVMTL - a Support Vector Machine Template Library

LIBSVMTL is a highly customizable C++ Support Vector Machine library based on Chih-Jen Lin's LIBSVM. The goals of splitting the library into template classes were:

  • let user decide wether to use sparse or dense storage for feature vectors
  • easier extention with new algorithms via a more object oriented design, e.g. in libsvmTL one kernel (which is the kernel function, its name, commandline parameters, loading/saving its parameters, description of commandline parameters) is represented by one object. Same for two-class and multi-class algorithms
  • integration into own programs (e.g., using own FeatureVector classes, using your own data file format, etc.)
  • easier development and maintainance of interfaces to higher level languages (e.g. matlab, python, etc)
  • making the library parts as independent as possible, to allow multiple developers to contribute
  • speed improvements

Besides that, the low-level training algorithms and structure is as identical to original libsvm from Chih-Jen-Lin as possible (only some minor changes to include the templated feature vectors and kernel functions), which allows easy integration of further improvements in libsvm and ensures correct numerics.

Table of contents