-
-
Save jonpsy/cf0222c5b7eb5813478c881a1d220a5c to your computer and use it in GitHub Desktop.
width = exp(log_width*2.0) * 2.0
this seems not right why is there a factor or 2 inside the exp?
in shogun, you can compare against the gaussian kernel computed with shogun, no need to implement your own
and the same for sklearn actually.
Because this still doesnt test whether the parametrization is the same you know. The results have to match when passing the same log_width to GaussianKernel
and the RFF embedding
About the bandwidth, I've seen np.log(100)
in our ipynb notebooks so I went with that(for ex: Gaussian Kernel in Classification.ipynb).
Next, here's what "get_width" gives return std::exp(m_log_width * 2.0) * 2.0
.
that is probably for high dimensional data then?
Ok thanks for the formula, seems ok then. Still you want to compare against shogun's kernel not your own
Will be updated, be assured
Looks like it's done!
Looks good now :)
The bandwidth is a bit too large to be realistic