KernelFunctions.jl

KernelFunctions.jl copied to clipboard

Failures on master

4

- [x] Transform: #530 - [ ] Transform: new problem with AD and `AxisArrays` - [x] BaseKernels: #528 - [x] BaseKernels: #546 - [x] Deprecation warnings: #529 The BaseKernels failures...

simsurace

Optimize `pairwise` for `Sinus` and scalar inputs

1

**Summary** A call like `kernelmatrix(PeriodicKernel(), rand(1000))` produces a number of allocations of the same order as the size of the kernel matrix because `Sinus` has a vector parameter and the...

simsurace

[WIP] Try out scalar parameters

1

This is just an experiment to see what would break if we replace some vector parameters by scalars.

simsurace

https://github.com/JuliaGaussianProcesses/KernelFunctions.jl/pull/528 added a less efficient workaround for a Zygote-specific issue: https://github.com/FluxML/Zygote.jl/issues/1464 The PR can be reverted once the upstream issue is fixed.

devmotion

Make Test a weak dependency

3

I'm not sure if this is a good idea at all and if it works as intended - BUT I've encountered multiple repos in the last few months where people...

devmotion

We should use ChainRulesTestUtils to test the rules defined in this package.

simsurace

add enzyme to AD tests

2

**Summary** **Proposed changes** * ... **What alternatives have you considered?** **Breaking changes**

st--

Second derivative of Matern in zero is wrong

23

```julia julia> using KernelFunctions: MaternKernel julia> k = MaternKernel(ν=5) Matern Kernel (ν = 5, metric = Distances.Euclidean(0.0)) julia> import ForwardDiff as FD julia> kx(x,y) = FD.derivative(t -> k(x+t, y), 0)...

FelixBenning

Loosen tolerance on test

2

**Summary** Replaces an equality test with an approximate equality test to handle roundoff issues. Should resolve #494 **Proposed changes** Loosen a test tolerance very slightly **What alternatives have you considered?**...

willtebbutt

Documenting caveats of `SimpleKernel`

7

The [docs](https://juliagaussianprocesses.github.io/KernelFunctions.jl/stable/create_kernel/#SimpleKernel-for-kernel-functions-depending-on-a-metric) recommend `SimpleKernel` for building ones own kernel using Distances.jl. They should probably here also note that not every `PreMetric` yields a positive-definite kernel. In particular, as Theorem 1...

sethaxen