Is there a C++ SVM library that runs on the GPU with CUDA 6.5, without Matlab (cuSVM). I tried with libSVM but doesn't work with CUDA 6.5 (the last version works just with cuda 5.5, ) and with WU-SVM but I have problems with the installation. CuSVM works with Matlab but I don't need to use Matlab. I'm working with Linux, C ++, and OpenCV. OpenCV have SVM functions but don't run in the GPU, just in CPU. I'm looking for a SVM library easy to use that runs with CUDA and GPU.
I already did it. MKL just works until CUDA 5.5 version. I am working with CUDA 6.5.
Hello José García and Hossein Pejman
SVMs rely on finding hyperplanes that split the space Rn where the data vectors are contained. Training the network means to effectively calculate the hyperplanes and then use these to obtain a NN performing reasonably well on the data.
If your work is restricted to SVM, there is nothing essentially new to be tried.
But if you are free to change paradigms there is another hyperplane based procedure, much more reliable and fast than SVM, to train NNs.
Go to
https://www.noisebridge.net/wiki/Polyhedrons_and_Perceptrons
There you will find a recently posted .cpp file ready to build an executable that calculates the desired NN. There are also data files for trials, instructions and links to some theoretical papers explaining the procedure.
Summing up, you will have the full source code, theory, data sets and instructions.
Pleased to be useful and with best regards
Daniel Crespin
Any news on this ? I'm in the same situation (except I'm working with CUDA 9.1)
- Badges
- Science topic