Diffractor.jl

Diffractor.jl copied to clipboard

This ought to return a tuple; what it returns now works in some contexts but not https://github.com/JuliaDiff/ChainRules.jl/pull/644 . ```julia julia> using Diffractor, ChainRulesCore julia> ret = rrule_via_ad(Diffractor.DiffractorRuleConfig(), x -> log(exp(x)),...

mcabbott

Very WIP: Stage2 revival

2

Needs some corresponding changes in Cthulhu

Keno

Support getindex on non-numeric arrays

1

This came up while trying to run Flux models with `Chain{

ToucheSir

Add jacobian

4

This uses a slightly different definition of the jacobian than ForwardDiff. For jacobian(f:N×M->P) it will return a result of shape `P×N×M` while ForwardDiff will return `P×(N⋅M)`

oscardssmith

At present Diffractor may return thunks, but doesn't seem use them (or anything else) to accumulate effciently: ```julia julia> @btime gradient(x -> sum(x), $(rand(100, 100))) |> first |> typeof min...

mcabbott

Take testing seriously

5

This adds many tests from Zygote. I've done some filtering to leave out the most repetitive ones, things which clearly don't apply (like Params), and things testing rules for external...

mcabbott

Failure with ComposedFunction `∘`

1

I'm pretty confident this worked in September, ~~but have no idea whether changes here or in ChainRules broke it:~~ Edit -- the change is https://github.com/JuliaDiff/ChainRulesCore.jl/pull/495, discussed there. ```julia julia> Diffractor.gradient(cbrt,...

mcabbott

`x.re` gives a Tangent not a Complex

1

```julia julia> gradient(x -> x.re, 2+3im) (Tangent{Complex{Int64}}(re = 1,),) julia> gradient(x -> abs2(x * x.re), 4+5im) ERROR: MethodError: no method matching +(::ComplexF64, ::Tangent{Complex{Int64}, NamedTuple{(:re,), Tuple{Float64}}}) ```

mcabbott

These should agree, right? ```julia julia> Diffractor.PrimeDerivativeBack(x -> real(sin(x)))(1+im) 0.8337300251311491 + 0.9888977057628651im julia> Diffractor.PrimeDerivativeFwd(x -> real(sin(x)))(1+im) ERROR: Tangent space not defined for `Complex{Int64}. ```

mcabbott

Cutting a release

7

Even though Diffractor isn't yet feature complete, it can already be a [major improvement](https://github.com/FluxML/Flux.jl/pull/1809#discussion_r777770025) over Zygote for real-world use cases.

ToucheSir