-
### Motivation and description
When trying to precompile function calls to gradients of `pairwise`, a mutation error is sometimes thrown. According to [this discourse thread](https://discourse.juli…
-
I'm trying to use Zygote.jl together with StaticArrays.jl, but am getting an error message I don't understand. Here's a very minimal example
```julia
julia> gradient(n->SMatrix{1,1}(n)[1], 1)
ERR…
-
Since checkpointing can be implemented as a sensitivity,
it may very well belong here.
See:
- Nabla: https://github.com/invenia/Nabla.jl/pull/171
- Zygote: https://fluxml.ai/Zygote.jl/dev/ad…
-
Cool project! I think it would be cool if simulation-based methods could be supported here by hooking into DifferentialEquations.jl and its DiffEqFlux fast adjoint schemes (https://diffeqflux.sciml.ai…
-
PR #467 implemented the bi-Heyting algebra of sub-C-sets, which is the propositional fragment of the logic of a presheaf topos. We should upgrade it to predicate logic, which would involve implementin…
-
Hi @Ceyron
I love the videos!
Wrote up the derivation for the ODE adjoints and wrote a derivation for getting gradients for an optimal controller u(t) that influences the ODEs.
I'm working on th…
-
Looking at the description of the arguments of `diffeqsolve` closely, I realized that irrespective of choosing the solver to be `diffrax.ReversibleHeun`, the backward pass by default is `diffrax.Recur…
-
This came up in #366.
And it has proved tricky to debug.
AFAICT there were two issues, one from ChainRules directly (I forgot to put the adjoint for `abs(::Complex)` back in, fixed now).
But second…
-
I think it is better to pass in `f_p`, `g_p` and `g_u` explicitly for adjoint sensitivity analysis instead of relying on the parameter `p` in `f(du,u,p,t)` and `dg(out,u,p,t)`. Because `p` can be used…
-
e.g. https://github.com/JuliaLang/julia/pull/34896
I think there's also a per-module `@nospecialize` trick. Zygote's compiler itself doesn't need to be fast, just the code and adjoints it generates…