Type stability for GPs

[kaandocal]

It seems like AbstractGPs is not always type stable in the way we'd expect when calling posterior:

using AbstractGPs

gp = GP(Matern32Kernel())

X1 = ColVecs(rand(1, 10))
Y1 = rand(10) 

gp_post = posterior(gp(X1, 0.01), Y1)

X2 = ColVecs(rand(1, 10))
Y2 = rand(10)

typeof(posterior(gp_post(X2, 0.01), Y2)) == typeof(gp_post) # returns false

[devmotion]

Shouldn't you use

typeof​(​posterior​(​gp​(X2, ​0.01​), Y2))

in your comparison?

BTW your comparison doesn't guarantee type stability, you would want to check if the type can be inferred correctly. This can be done e.g. with @code_warntype and @inferred.

[kaandocal]

My fault, I forgot the second typeof (the result doesn't change however). Just to clarify, by type stability I meant that adding more observations to a GP changes its type in this case!

[devmotion]

I can't check it right now but I assume that the problem is a missing vcat definition for ColVecs, so the type changes when concatenating X1 and X2. Does it work with regular vectors for X1 and X2?

[kaandocal]

That is indeed the problem! For 1-dimensional vectors X1 and X2 it works. I'll create a PR in KernelFunctions.jl.

[willtebbutt]

@kaandocal is this now resolved as #291 is merged?

[devmotion]

I'll take that as a yes 🙂