KernelFunctions.jl
KernelFunctions.jl copied to clipboard
Make Zygote work for Wiener kernel
#116
I have defined kernelmatrix for wiener kernel. This avoid the "slow method" ~~but gives NaNs in the Zygote gradient~~.
The NaN values in the Zygote gradient seem to be caused by the the pairwise(Euclidean(), x, y). I am not sure why this is happening.
Can you try the Zygote master branch locally? Maybe it was fixed by https://github.com/FluxML/Zygote.jl/pull/787.
Can you try the Zygote master branch locally? Maybe it was fixed by FluxML/Zygote.jl#787.
Thanks! The Zygote and ForwardDiff tests pass on Zygote#master. However, ReverseDiff fails, probably requires nan checking.
There's a new Zygote version which contains the fix mentioned above.
It is weird that Zygote on Gabor kernel tests are failing on Mac without any changes related to it. https://travis-ci.com/github/JuliaGaussianProcesses/KernelFunctions.jl/jobs/386076575#L369
This looks to be stale, so I'm closing.