Streaming, multi-resolution, adaptive, kernel solver
Cite the code:
The StreaMRAK folder contains the StreaMRAK code, while the MRFALKON folder contains the LP-KRR code. Experiments for the StreaMRAK paper are found in the Experiments folder. To run the experiments, it is first neccessary to generate the datasets using the code in the Experiments/Datasets folder. The MRFALKON code corresponds to the LP-KRR in the StreaMRAK paper.
When running the MRFALKON experiments, it is neccessary to set the correct value for the lowlim_ntrp parameter in the MRFALKONmainConfig_mrfalkon configuration file. This file can be found in the folder StreaMRAK/StreaMRAKconfig.
For the varsin experiments lowlim_ntrp = 110000
For the varsin experiments lowlim_ntrp = 180000
For the varsin experiments lowlim_ntrp = 390000
The falkonSolver.py implementation in the StreaMRAK folder is a python adaptation that we have made of the MATLAB code in https://github.com/LCSL/FALKON_paper.git with the affiliated paper [1].
[1] A. Rudi, L. Carratino, L. Rosasco, FALKON: An Optimal Large Scale Kernel Method, in: Advances in Neural Information Processing Systems Vol. 30, 2017, pp. 3889-3899.