:nothing initialised in FastChain

[MatisPatel]

Hey I'm having some issues running the base examples from the docs. The problem seems to be FastChain initialising some parameters as :nothing. MWE:

using DiffEqFlux
U = FastChain(
    FastDense(2,5),   
    FastDense(2,5))
@show intitial_params(U)

# initial_params(U) = Union{Nothing, Float32}...

[KirillZubov]

oh, it's the same problem #594

[ChrisRackauckas]

@DhairyaLGandhi has this been fixed?

[DhairyaLGandhi]

The fix for the nothings is in FluxML/Flux.jl#1670 and ref #592

However, the error seen in https://github.com/SciML/DiffEqFlux.jl/pull/592/checks?check_run_id=3128353069 comes from this function which was from FluxML/Flux.jl#1636 and that would still break the Newton tests and maybe others.

[DhairyaLGandhi]

The nothing issue should be fixed on Flux master

[MatisPatel]

Hey sorry for the novice question but can I get julia use master branch rather than release? add Flux gives me v0.12.5

Edit: worked it out add Flux#master

[ChrisRackauckas]

You shouldn't have to grab master. Why isn't Flux just doing a release?

[DhairyaLGandhi]

I was just about to schedule a release before I saw activity here :)

As noted, it doesn't fix all the tests that fail yet.

[ChrisRackauckas]

What's the deal with the Newton test? What adjoint is the one throwing the error/warning?

[DhairyaLGandhi]

Close?