Open cortner opened 5 years ago
looks like the Zygote performance with StaticArrays is still terrible. Need to look at this more closely.
just to continue recording activity -> I'm starting with a generic AD hessian implementation that uses ForwardDiff.jl
for now; first prototype is working well.
2019 wasn't the time, but now it is. With Chainrules we can now do a huge amount. But we need to think about how it will interact with ACE, the new molecular simulation eco-system, e.g. Molly etc.
It looks like
Zygote.jl
is already very good and with some hacks can come very close to hand-coded gradients:https://discourse.julialang.org/t/zygote-performance/29028
It is probably time to have default AD gradients for all potentials and benchmark whether hand-coded gradients are still worth our while...