Gradient Tape with random number about reversediff.jl HOT 1 CLOSED

This is actually easier than I thought, but I post the answer anyway:
The two following functions generate a random number of the same type as 'a'

ReverseDiff.@forward function randfor(a) # uniform random number generator
        rand(a)
end
ReverseDiff.@forward function randnfor(a) # Normal random number generator
        StatsFuns.norminvcdf(rand(a))
end

If I want to sample from a multivariate normal distribution with diagonal covariance, I can execute

r = @. m+s*randnfor(m)

This works well when the code is interpreted by ReverseDiff or ForwardDiff, but fails otherwise, as rand does not accept a Float64 input. This can be solved by setting

Base.rand(::Float64) = rand()

This works also for acceptance/rejection sampling algorithms. Random indexing seems to be trickier though...

from reversediff.jl.