David Widmann

[WIP] Run format check for PRs from forks as well

2

In principle, it should be possible to run format checks (and build the documentation) also when a PR is opened from a fork. The security settings on Github do not...

We could also implement this approach (Section 5) https://arxiv.org/pdf/1807.01750.pdf _Originally posted by @theogf in https://github.com/JuliaGaussianProcesses/KernelFunctions.jl/issues/245#issuecomment-766824569_

Make ARDTransform a subtype of LinearTransform?

4

As discussed in https://github.com/JuliaGaussianProcesses/KernelFunctions.jl/pull/106, the ARDTransform could be viewed as a special LinearTransform with a diagonal matrix. Do we still want to keep these separate types?

Add `erfinv` and `erfcinv` for `Float16` and generalize `logerfc` and `logerfcx`

2

As the title says :slightly_smiling_face: I was a bit unsure if computations in `erfinv` and `erfcinv` should be performed with `Float32` or `Float16` since the coefficients in Blair's paper are...

Use `RealDot.realdot` for more efficient `real(dot(x, y))`

2

This PR uses `RealDot.realdot` to perform computations of `real(dot(x, y))` more efficiently. [RealDot.jl](https://github.com/JuliaMath/RealDot.jl) is a very lightweight package that only defines `realdot` and only depends on LinearAlgebra: https://github.com/JuliaMath/RealDot.jl/blob/main/src/RealDot.jl

Handle arguments of 1 in `beta` and `logabsbeta`

4

This PR adds special handling for arguments of 1 in `beta` and `logabsbeta`. This fixes one of the issues in https://github.com/JuliaStats/StatsFuns.jl/pull/126 (I checked locally that the problem is fixed) but...

Derivatives of the incomplete gamma functions with respect to parameter `a` are missing

13

Copied from https://github.com/JuliaMath/SpecialFunctions.jl/pull/305: > I added ChainRules definitions for gamma(a, x), loggamma(a, x), and gamma_inc according to https://functions.wolfram.com/GammaBetaErf/GammaRegularized/introductions/Gammas/ShowAll.html . Similar to the Bessel functions, derivatives with respect to the first...

Use/export `LogExpFunctions.jl`?

14

The implementation of `logsumexp` in StatsFuns is quite optimized (see, e.g., https://github.com/JuliaStats/StatsFuns.jl/pull/97), it works with GPUs, is numerically more stable than the implementation in NNlib, and uses a one-pass algorithm....

Use existing implementations in StatsFuns

6

This PR removes the implementations of the softplus and the logistic function from this package and uses the implementations in StatsFuns instead. It goes back to https://github.com/FluxML/NNlib.jl/pull/98#issuecomment-476617160.

Skellam (special case) is broken with Zygote and SpecialFunctions >= 1

7

Currently, these distributions are broken due to https://github.com/FluxML/Zygote.jl/issues/873. As discussed in https://github.com/TuringLang/Bijectors.jl/pull/155, the corresponding tests are marked as broken. The issue will be fixed upstream. See also: https://github.com/TuringLang/DistributionsAD.jl/issues/145