GPU-SVM has modules for multi-class SVM training and classification, and also supports probabilistic output. We are currently upgrading GPU-SVM to support SVM regression.
The associated paper of this source code is: Wen, Zeyi, et al. "MASCOT: fast and highly scalable SVM cross-validation using GPUs and SSDs." 2014 IEEE International Conference on Data Mining.
This software is licensed under Apache Software License v2.0.
CUDA 7.5 or later; g++ 4.8 or later
-
How can I use the source code?
A: Download the repository and then issue "make" command under the fold where our Makefile is located. After the command is completed, you will see a binary file named "mascot" in the "bin" folder. To start playing with it, run the "run.sh" script. The datasets shown in run.sh are available here in LibSVM site. -
What is the format of the input file?
A: The file format is the same as the format of files in LibSVM site. -
Does this version support the Windows OS?
A: No. However, the code should work on Windows OS, although we haven't tested it on Windows. -
Do I have to install an SSD if I want to use GPU-SVM?
A: No. GPU-SVM works fine with HDDs, although SSDs would help improve the efficiency. -
What are the meanings of the options?
A: -b is for SVM with/without probability output; -o is for setting the task type (e.g. training or cross-validation); -g is for setting the gamma value; -c is setting C for regularisation; -e is for setting the test dataset name; -f is to let GPU-SVM know the data dimensionality (this argument is optional). -
I got "error while loading shared libraries: libcudart.so.6.0: wrong ELF class: ELFCLASS32", when I run the executable file "mascot".
A: Running the command ''sudo ldconfig /usr/local/cuda/lib64'' should resolve the problem..