stackoverflow when running inference with Vector{Any} observations.

[albertpod]

I am getting cryptic stackoverflow when running inference with Vector{Any} observations.

@model function example_bug()
    y = datavar(Vector{Any})
    x ~ MvNormalMeanCovariance(zeros(2), diageye(2))
    y ~ MvNormalMeanCovariance(x, diageye(2))
end

result = infer(model=example_bug(), data=(y = Any[1, 2.0], ),)

Running inference yields:

ERROR: StackOverflowError:
Stacktrace:
     [1] mean(itr::PointMass{Vector{Any}})
       @ Statistics ~/.julia/juliaup/julia-1.10.0+0.x64.apple.darwin14/share/julia/stdlib/v1.10/Statistics/src/Statistics.jl:44
     [2] mean(fn::typeof(identity), distribution::PointMass{Vector{Any}})
       @ BayesBase ~/.julia/packages/BayesBase/ZObVB/src/densities/pointmass.jl:26
--- the last 2 lines are repeated 39990 more times ---
 [79983] mean(itr::PointMass{Vector{Any}})
       @ Statistics ~/.julia/juliaup/julia-1.10.0+0.x64.apple.darwin14/share/julia/stdlib/v1.10/Statistics/src/Statistics.jl:44

To circumvent this error, we need to change the data var and data input as follows, which is fine, but the error should be handled better.

@model function example_bug()
    y = datavar(Vector{Float64})
    x ~ MvNormalMeanCovariance(zeros(2), diageye(2))
    y ~ MvNormalMeanCovariance(x, diageye(2))
end

result = infer(model=example_bug(), data=(y = [1, 2.0], ))

[bvdmitri]

I just checked out this problem (sorry completely forgot about it), thinking it would be easy to fix, but it's not that simple. Basically, we can sort out the mean, but the cov/std/pdf/logpdf methods are trickier. And fixing only the mean is not really a solution because I suppose the rules will call cov as well. The problem is they need one(T) and zero(T) to be defined, which isn't the case when T = Any. So, in the code, we've specifically said T has to be a type of number (Real), but that's causing a confusing error because the method for T != Real doesn't exist. Maybe we should just prevent making a PointMass if T isn't a real number. What do you think?

[bvdmitri]

Perhaps we should just do a generic fallback

BayesBase.mean(p::PointMass) = p.point