@kernel macro for creating kernel

[theogf]

So I started creating a first version of the @kernel macro which for now follows this syntax:

@kernel scale*Kernel(keywords arguments of the kernel) <keywordargument/argument for a transform>

So the transform is given either by a keyword l=2.0, l=[2.0,3.0] etc or directly by a parameter
For the kernel parameters they are restricted to be in the Kernel constructor

Edit : Left to do :

• Add docs
• Add tests

[willtebbutt]

Am busy this week. Will take a look early next week.

[willtebbutt]

Unable to review at the minute as the @kernel macro gives the error:

julia> @kernel 5 * Matern32Kernel() l=5.0
ERROR: LoadError: UndefVarError: kw not defined
Stacktrace:
 [1] macro expansion at ./show.jl:562 [inlined]
 [2] @kernel(::LineNumberNode, ::Module, ::Expr, ::Any) at /Users/willtebbutt/.julia/dev/KernelFunctions/src/kernels/kernel_macro.jl:8
in expression starting at REPL[13]:1

[theogf]

Sorry about that! it should be fixed now

[willtebbutt]

This seems good:

julia> (@kernel Matern32Kernel() l=[5.0, 2.0]) + (@kernel Matern52Kernel() l=[4.0, 3.0])
Sum of 2 kernels:
	- (w=1.0) Matern32Kernel
			- ARDTransform{Float64,2}([5.0, 2.0])
	- (w=1.0) Matern52Kernel
			- ARDTransform{Float64,2}([4.0, 3.0])

(the printing is already quite cool, nice work!)

This I found surprising:

julia> @kernel (Matern32Kernel() + Matern52Kernel()) l=[5.0, 4.0]
ERROR: MethodError: no method matching transform(::KernelSum, ::Array{Float64,1})
Closest candidates are:
  transform(::BaseKernel, ::AbstractArray{T,1} where T) at /Users/willtebbutt/.julia/dev/KernelFunctions/src/kernels/transformedkernel.jl:22
Stacktrace:
 [1] top-level scope at REPL[26]:1

Maybe it's just the case that transform's type constraints need to be relaxed to Kernel from BaseKernel?

Also, this doesn't seem right:

julia> @kernel Matern32Kernel() l=5.0
Matern32Kernel
	- Scale Transform s=5.0

edit: I just found this last example confusing because I was confused as to what `ScaleTransform` is doing -- it's got a very similar name to `ScaledKernel`. It might be better for change how some of this prints, but that's for a separate PR I would imagine.

[theogf]

(the printing is already quite cool)

This I found surprising:

julia> @kernel (Matern32Kernel() + Matern52Kernel()) l=[5.0, 4.0]
ERROR: MethodError: no method matching transform(::KernelSum, ::Array{Float64,1})
Closest candidates are:
  transform(::BaseKernel, ::AbstractArray{T,1} where T) at /Users/willtebbutt/.julia/dev/KernelFunctions/src/kernels/transformedkernel.jl:22
Stacktrace:
 [1] top-level scope at REPL[26]:1

Maybe it's just the case that transform's type constraints need to be relaxed to Kernel from BaseKernel?

Yeah somehow I got fixed in the idea that Transform could only be applied on a BaseKernel I will extend the transform to all kernels.

Also, this doesn't seem right:

julia> @kernel Matern32Kernel() l=5.0
Matern32Kernel
	- Scale Transform s=5.0

I don't understand why? Should we start consider the parameter l as the actual lengthscale ? (1/l)

[willtebbutt]

I don't understand why? Should we start consider the parameter l as the actual lengthscale ? (1/l)

Sorry, no, I just misunderstood what is going on. I think it might be better to modify how this is printed a little bit, but that can be addressed in a separate issue. Please ignore my third point :)

[theogf]

In terms in UI design, is it preferable to force to have a keyword, e.g. @kernel SEKernel() l = 3.0 or is it okay to throw in any argument to make it an equivalent of transform, e.g. @kernel SEKernel 3.0 ?

[devmotion]

In terms in UI design, is it preferable to force to have a keyword, e.g. @kernel SEKernel() l = 3.0 or is it okay to throw in any argument to make it an equivalent of transform, e.g. @kernel SEKernel 3.0 ?

Initially I thought @kernel SEKernel() 3.0 would be too cryptic. However, I'm not so sure anymore since l = ... or t = ... is also not a completely intuitive syntax. That got me thinking if we need @kernel at all if one can just write transform(SEKernel(), 3.0) instead (but arguably the explicit syntax is completely user-friendly). At least it seems it's not needed to handle the scale with @kernel since the same syntax can be used outside of @kernel by writing 3 * @kernel SEKernel() l = 3.0 instead?

[theogf]

This is true I can completely remove the scaling. If one put it inside of outside would not matter.
@kernel is indeed an alias for transform, but should be more intuitive. But there is no absolute best solution here. I tend to prefer l=3.0 but I guess it is personal.

[theogf]

Goodbye little one!