Update deps & bump to 0.16.1

[pxl-th]

PR Checklist

• Tests are added
• Entry in NEWS.md
• Documentation, if applicable

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 32.50%. Comparing base (101106e) to head (fc5dac1).
Report is 5 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #2574   +/-   ##
=======================================
  Coverage   32.50%   32.50%           
=======================================
  Files          34       34           
  Lines        2003     2003           
=======================================
  Hits          651      651           
  Misses       1352     1352           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[mcabbott]

CI seems to get ⌃ [e88e6eb3] Zygote v0.6.75 not 0.7... maybe Optimisers.jl is to blame.

NNlib has errors FluxML/NNlib.jl#619

[pxl-th]

It's because of ComponentArrays used in tests. It needs a bump for Zygote@0.7.

UPD: Ah and optimisers as well.

[CarloLucibello]

Ok, we are now picking Zygote 0.7.1 but there are some real failures

[pxl-th]

I think conv failures should be fixed by FluxML/NNlib.jl#620

[CarloLucibello]

Now tests are not passing due to g containing thunks in the following example

julia> m = Chain([Dense(3 => 4, tanh; bias=false), Dense(4 => 2)]);

julia> x = randn(Float32,3,5);

julia> y = rand(Bool,2,5);

julia> g = gradient(m -> Flux.logitcrossentropy(m(x), y), m)[1]
(layers = NamedTuple{(:weight, :bias, :σ)}[(weight = InplaceableThunk(ChainRules.var"#..., Thunk(ChainRules.var"#...)), bias = nothing, σ = nothing), (weight = Thunk(Zygote.var"#...), bias = Float32[-0.088241935, 0.088241935], σ = nothing)],)

[mcabbott]

Maybe Flux.gradient should recursively un-thunk?

[CarloLucibello]

I wouldn't expect any thunk to appear in the output of Zygote.gradient as well, right?

[CarloLucibello]

Reduced to

julia> using Flux

julia> layers = [Dense(3, 4)]
1-element Vector{Dense{typeof(identity), Matrix{Float32}, Vector{Float32}}}:
 Dense(3 => 4)       # 16 parameters

julia> x = ones(Float32, 3);

julia> g = gradient(layers -> sum(layers[1](x)), layers)[1]
1-element Vector{@NamedTuple{weight::ChainRulesCore.Thunk{Zygote.var"#163#164"{Nothing, ChainRulesCore.InplaceableThunk{ChainRulesCore.Thunk{ChainRules.var"#551#556"{Vector{Float32}, Vector{Float32}}}, ChainRules.var"#550#555"{Vector{Float32}, Vector{Float32}}}}}, bias::Vector{Float32}, σ::Nothing}}:
 (weight = Thunk(Zygote.var"#...), bias = [1.0, 1.0, 1.0, 1.0], σ = nothing)

@mcabbott @pxl-th any idea?

[pxl-th]

I have a fix, let me see if it breaks anything else.

[pxl-th]

Should be fixed by FluxML/Zygote.jl#1551

[pxl-th]

Should be good. Other errors are unrelated