Working with arbitrary multivariate mean

[parikshit-pareek]

I am attempting to define a arbitrary multivariate mean GP as:

function g(X)  
    x = X[:,1]; z = X[:,2]; 
    x +z
end

 f = stretch(GP(X->g(X),SEKernel()),1 / l);
a = randn(2,N_train)
X = ColVecs(a) # Converting the matrix into a multi-column object

However, I am not able to optimise the hyper parameters and getting the following error:

BoundsError: attempt to access 2-element view(::Matrix{Float64}, :, 1) with eltype Float64 at index [1:2, 2]

Where can I find any example of doing this?

[willtebbutt]

Thanks for opening this @parikshit-pareek . Could you please provide a full copy of your stack trace so that I can get a better idea of where the problem might be coming from?

[parikshit-pareek]

@willtebbutt Take a look at this code : Here

Also, I discovered that if the mean function is expressed in terms of matrix operations, it works fine.

 function m(X)  
          X'*X
 end

[willtebbutt]

Ahhh It's an AD issue.

The way to solve it is to implement _map_meanfunction(::CustomMean{typeof(g)}, ::ColVecs) using something like the matrix operations that you've described above. e.g.

function AbstractGPs._map_meanfunction(::CustomMean{typeof(g)}, x::ColVecs)
    X = x.X
    return <some operations on X that returns a vector of length `length(x)`>
end

This is the most important operation on a mean function anyway, so it's a good idea to implement this if you at all care about performance.