[FEA] Add precomputed kernels to SVM

[tfeher]

Sklearn defines a precomputed option SVM's kernel parameter. In this case, instead of passing the training vectors, we pass the Gram matrix that contains kernel values between all training examples.

This can be considered as a big kernel cache which is initialized by the user, and could be implemented by adapting the KernelCache class accordingly.

[emildi]

Hello,

Is this functionality being considered for future implementation?

This would be quite useful as sklearn implementation is quite slow (being single threaded) and I find ThunderSVM quite buggy and outdated.

Thanks,
Emil