Avik Pal
- [x] Add a formatter - [ ] Add common SciML workflows - [x] General Housekeeping for Linear Curve Fit (if needed) - [ ] Interface with NonlinearSolve.jl for Nonlinear...
I am using this to accumulate all the fixes needed for the project, so let's hold off on merging. - Fixes SteadyState Adjoint for OOP jacobian
The Nested AD used in DiffEqFlux is not ideal because it calls `ForwardDiff.gradient/jacobian` and Zygote overrides them to compute the Hessian before doing the HVP. [BatchedRoutines.jl](https://github.com/LuxDL/BatchedRoutines.jl/blob/main/src/api.jl) has routines to do...
Current version ```julia @concrete struct NeuralODE{M
```julia # using ArrayInterface, NNlib, Strided, Enzyme, FastBroadcast using NNlib, Enzyme, Strided, FastBroadcast @inline function __fast_broadcast!(f::F, x, args...) where {F} # if ArrayInterface.fast_scalar_indexing(x) if maximum(length, (x, args...)) > 200_000 #...
The current implementation for https://lux.csail.mit.edu/stable/api/Lux/contrib#Map-over-Layer was written when there was no such functionality in Functors. We can rewrite the `@layer_map` to use a custom Walk type and KeyPath. Of course,...
- [ ] #588 for specifying constants directly - [ ] Special syntax for providing functions that are evaluated in `initialparameters` (maybe `@initfn` annotation) - [ ] **(Very Speculative)** Rewrite...
Currently we drop all of the init functions from the Lux models. However, for SciML Applications it is quite critical to support those. Proposed solution: Store the lux model along...
- CPU Support: We have tests covering several models. - [x] PolyAlg selecting the correct broadcast mechanism for `fast_activation!!` fails https://github.com/EnzymeAD/Enzyme.jl/issues/1408 (Fixed upstream needs to be verified here) - [x]...