CUDA-SVM

Example of Support Vector Machine (SVM) Classifier w/ Cross-Compilation for CUDA & CPU. The Pegasos method for sub-gradient optimization has been implemented.

How to Compile CUDA-SVM

As of the time of this writing, CUDA-SVM supports three compilation modes:

• CPU Version: By running the make command. If clang is available, it will be used as the default compiler. Otherwise, gcc will be adopted.
• CUDA GPU Version: By running the make cuda command.
• Debug Mode (Currently only supported in CPU): By running the make debug command.

Generating Dataset

In order to generate the datasets (i.e, either the small binary Mushrooms dataset, or the large multilabel RCV1), run the ./fetch_data.sh command followed by the argument mush or rcv1.

Note: As of right now, the RCV1 version has not been implemented.

Pegasos Method for Subgradient Descent Optimization of SVMs

• Original Paper
• MIT Slides on Pegasos
• MIT Slides on SVMs
• Good Reference Python Implementation