Compatibility with Zygote AD

[Uroc327]

It seems that only ForwardDiff is supported for ADing sampling. Is it possible to implement rules for Zygote as well?

julia> ForwardDiff.derivative(x -> sum(rand(Normal(x, 10), 10)), 0.)
10.0

julia> Zygote.gradient(x -> sum(rand(Normal(x, 10), 10)), 0.)
(nothing,)

[trahflow]

Is it possible to implement rules for Zygote as well?

It is possible, e.g. (just a quick hack):

function ChainRulesCore.rrule(::typeof(rand), d::Normal{T}, n::Integer) where {T}
   vals = rand(d, n)
   function rand_pullback(rand_bar)
       d_bar = Tangent{Normal{T}}(;μ=n, σ=sum((vals .- d.μ)) / d.σ)
       return NoTangent(), d_bar, NoTangent()
   end
   return vals, rand_pullback
end

But I wonder why a dedicated rule is necessary at all. Looking at the the definition for rand for Normal, this should be easily differentiable. I don't really see why Zygote returns nothing here

[devmotion]

It is due to https://github.com/JuliaDiff/ChainRules.jl/blob/158ca756ef99ccf3f1dde2e66b5855e8e68e0363/src/rulesets/Random/random.jl#L23-L25. It is a deliberate design decision to mark rand methods as non-differentiable in ChainRules (which is used by Zygote). They were made more restrictive to explicitly not cover e.g. rand(Normal()) (see e.g. the discussion in JuliaDiff/ChainRules.jl#262). However, I assume the problem with the example above is that sampling multiple samples in Distributions is done via pre-allocating the output and sampling with rand! - and that one is marked non-differentiable, regardless of the type of the arguments. See also JuliaDiff/ChainRules.jl#603.

[trahflow]

sampling multiple samples in Distributions is done via pre-allocating the output and sampling with rand!

Ah, I missed that. Then everything makes senes.

Sort of a meta question: Where would one put AD rules for such things?
Here in the ext/DistributionsChainRulesCoreExt module?
Also found TuringLang/DistributionsAD.jl#123 but not sure if that Package is still meant to be used?