Remove _kernel, sort out kappa

[willtebbutt]

Addresses #87 .

Also no longer exports kappa, as it's now not needed as part of the public API.

I found a redundant function called _scale, so removed that.

Iteration over concrete subtypes of Kernel removed from generic as no longer required.

edit: the reason for quite a bit of code movement is that I originally hadn't realised I could implement (k::SimpleKernel)(x, y), and had to modify most kernel files, so sorted out spacing while I was there. I stand my the changes in formatting, even if they're not strictly the focus of this PR.

[devmotion]

I could implement (k::SimpleKernel)(x, y)

As I wrote in my comment above, I would be fine with this change but it requires us to drop support for Julia < 1.3 which we have been trying to avoid so far.

[willtebbutt]

Oh, fair point. For some reason I thought it didn't affect this case... my mistake. Personally I'm completely happy to drop < 1.3 support -- it's not like we have packages that depend on KernelFunctions that are restricted to 1.0 as we've not been around for long enough. Moreover, 1.5 is likely to be the next LTS, at which point it really makes sense to insist that 1.5 is the minimum version of Julia that we support.

@theogf thoughts on this?

[theogf]

I am also happy to drop support julia < 1.3. Given the new structure, it does not make sense anymore to keep it this way.

[willtebbutt]

All requested changes addressed. If either @theogf or @devmotion could re-review to approve / request more changes, that would be grand.

[devmotion]

Looks good to me, just left two minor comments. If they are resolved and tests pass, I'm fine with merging!

[theogf]

Ok for me too! Just need to make the tests pass

[willtebbutt]

All comments addressed. Will merge once tests pass.

[willtebbutt]

@theogf any idea what's going on with the test failures?

[theogf]

It's weird I just tried the branch locally and it's passing the tests

[willtebbutt]

Hmmm yeah, same. Presumably a dependency has broken somewhere in the pipeline. Maybe it's a Flux thing?

[devmotion]

It's weird I just tried the branch locally and it's passing the tests

It fails for me 🤔 Did you run ] up before running the tests?

[devmotion]

Presumably a dependency has broken somewhere in the pipeline. Maybe it's a Flux thing?

I compared https://travis-ci.com/github/JuliaGaussianProcesses/KernelFunctions.jl/jobs/323153454 with https://travis-ci.com/github/JuliaGaussianProcesses/KernelFunctions.jl/jobs/322480149, and it seems the main difference in the dependencies are Zygote 0.4.17 instead of 0.4.15 and CUDAdrv 6.2.3 instead of 6.2.2. The Flux version was 0.10.4 in both cases.

[willtebbutt]

This change in the v0.4.17 release looks relevant.

[devmotion]

Yep, seems someone else got similar issues with 0.4.17 (linked in the relevant PR). I just confirmed it locally, tests pass with 0.4.16.

[willtebbutt]

I've temporarily pinned Zygote to 0.4.16. @devmotion do you have the bandwidth to chase this up on the Zygote repo?