gamma ignored in fast kernel classifier

[amueller]

Right now it looks to me as if gamma is completely ignored, which is surprising.
I think it would be nice if we could replicate the behavior of SVC. If gamma doesn't make sense for the other kernels we can use bandwidth there, but it's weird to have a gamma parameter and then have it ignored.

[Alex7Li]

You're right. I previously was switching between using gamma and the bandwidth depending on the kernel, but after the potential confusion was pointed out, I changed it to use only bandwidth for all the kernels. I just forgot about removing gamma from the parameter list. I will go and remove it

[amueller]

that will make it tricky for someone who's using SVC to replace it by this model, though.

[Alex7Li]

Do you think it would be better to use only gamma and ignore bandwidth? That might be more consistent with sklearn. Also, if we are trying to be similar to sklearn, should we have a string for the gamma/bandwidth parameter like 'scale' or 'auto'? I'm not super knowledgeable about the theory but I can talk with siyuan/misha about how well the svm heuristics apply/creating some other heuristic.

[amueller]

I think being consistent will be good if it makes sense and is possible. Also, supporting the same strings, at least for the rbf kernel, would be nice.

[Alex7Li]

Ok. Is it better to make a new pull request or add this fix to the one that is already open #33?

[amueller]

I would open a different PR as these are independent issues.