Yes that is a good point. It would be cool to have it implemented say for all subclasses of KernelMachine. All it really does is changing the basis set to represent the kernel function using landmark points rather than the training data. If you want to impement it in this manner, this would be a very welcome contribution. However, doing this in general is difficult, and e.g. SVMs will have a different implementation as the solver itself will be changed, so definitely checking out sklearn would help here for abstractions. Any ideas how to go ahead with this? H
Am Mo., 23. März 2020 um 15:06 Uhr schrieb sai_ng via shogun-list < shogun-list@shogun-toolbox.org>: > Hi again, > It's me Nanubala Gnana Sai. I was checking out SKlearn implementation of > Nystrom and tried comparing with our own implementation. I was wondering, > why is the given approximation technique bounded to a specific method ( > for example: KRR), it should be implemented as different class atleast > that's what's done in Sklearn. There is an implementation for RFF which > works in a similar fashion, so I thought it's only logical to have Nystrom > implementation as well. Looking forward to hearing from you ! :D >
