David Widmann
David Widmann
I just replaced your model with ```julia function U_smooth(t, fr, delta, beta, tau) a = tanh(5 * (t - tau)) return exp(-delta * t) * (beta - fr) + fr...
> With a limit what would one be able to do? I would assume you could use it as a proper initial value. > Also the other observable of the...
As far as I can see right now, we only know how ```julia @. (1-Θ)*u0 + Θ*u1 ``` should be defined for the continuous variables `u0.x` and `u1.x`. IMO everything...
I had another look at the logs, and actually the problem seems to start earlier: the pullback for `cholesky` in ChainRules is not picked up it seems (https://github.com/FluxML/Zygote.jl/runs/7124557589?check_suite_focus=true#step:6:1540). I suspect...
Let's see what CI tests tell us. At least it seems there's a clear improvement locally, before 20 tests threw an error whereas it seems now only 8 fail and...
DynamicPPL errors on the current HEAD with Julia 1.8 are fixed by updates of Tracker and Bijectors.
FYI I opened a PR to StatsFuns a while ago (https://github.com/JuliaStats/StatsFuns.jl/pull/106) but it has not been approved (yet) even though the package already depends on ChainRulesCore indirectly via SpecialFunctions. Maybe...
I made a PR for AbstractFFTs: https://github.com/JuliaMath/AbstractFFTs.jl/pull/58
I just checked https://github.com/TuringLang/DistributionsAD.jl/pull/198 (the CR1 version) locally and it still fails with the same error messages ("adjoint for constructor ..."), even with this PR.
Ah sorry, I misunderstood your comment. Unfortunately, the example is not fixed either.