Using the noise parameters sol.W.W from an SDE solution throws an error when taking gradients

Specific Domains Machine Learning
, ,
2

I think the problem traces back to sol.W being nothing in the backward pass. Using ForwardDiff instead of Zygote to compute the gradients

@time gs = ForwardDiff.gradient(p) do p
	some_loss(p)
end

removes the error. If reverse-mode AD is needed, you could use pre-defined noise values and NoiseGrid from the DiffEqNoiseProcess package

  t = Array(tspan[0]:dt:tspan[1])
  Z = randn(length(t))
  Z1 = cumsum([0;sqrt(dt)*Z[1:end-1]])
  Zygote.ignore() do
     NG = NoiseGrid(t,Z1)
  end
  tmp_prob = remake(prob,p=p,noise=NG)

as a possible workaround. @ChrisRackauckas, do you have a better idea for using the noise values in a loss function with Zygote?