David Widmann

Isn't that the convention in Zygote and expected? In Zygote every undefined gradient is `nothing`, so it seems that it would quite breaking if in some cases this convention would...

Ideally, the convention in Zygote would be changed but I don't think this will happen anytime soon (and possibly never). But it's still not clear to me why Zygote being...

I'll take back my statement above. Apparently Zygote returns `NotImplemented`: https://github.com/FluxML/Zygote.jl/issues/1204#issuecomment-1098460916 Possibly the issue in #425 was that the pullback was incorrect (missing `NoTangent` for the function itself) - but...

I noticed that some recent changes in Distances make it necessary to analyze more carefully which methods are called. Distances.pairwise is not owned by Distances anymore. Instead it is owned...

I can reproduce the benchmarks. And replacing `Fix1` with a regular closure seems to fix the problem: ```julia julia> function kernelmatrix2(κ::KernelFunctions.SimpleKernel, x::AbstractVector) return map(d -> KernelFunctions.kappa(κ, d), KernelFunctions.pairwise(KernelFunctions.metric(κ), x)) end...

I defined custom matrix-valued kernels in https://github.com/devmotion/CalibrationErrors.jl/ up to version 0.3, but then replaced them with tensor product kernels on an extended input space (that encodes the desired dimension of...

> This is very common from what I can tell, and I think what @devmotion is talking about? No, it seems that's different from what I was using in my...

In this case the corresponding scalar-valued kernel is `k((x, s), (y, t)) = exp(-|x-y|^2) (delta(s, t) - (x_s - y_s) (x_t - y_t))`. It seems it can't be written as...

I think it wasn't me (actually I thought you implemented ColVecs and RowVecs initially @willtebbutt :smile: ) but in my opinion it is correct to return views here. I view...

I think `getindex` should satisfy the same properties as for `Vector`s of views: ```julia julia> A = rand(5, 4); julia> x = [view(A, :, i) for i in axes(A, 2)];...