ArrayNode reflectinmodel sometimes results in weird error

[racinmat]

The gradient computation crashes on following code, using following data:
https://ufile.io/8wl0eit1

using JLD2, FileIO, Flux, Mill
@load "weird_node.jld2" x1 y
model = reflectinmodel(x1, d -> Chain(Dense(d, settings.k, relu),),
	d -> SegmentedMeanMax(d),
	b = Dict("" =>  d -> Chain(Dense(d, 2),)))
ps = Flux.params(model)
loss = (model, x, y) -> Flux.logitcrossentropy(model(x).data,y)
loss(model, x1,y)
Flux.logitcrossentropy(model(x1).data,y)
gradient(() -> loss(model, x1,y), ps)

with error
ERROR: MethodError: no method matching zero(::Type{Any}) Closest candidates are: zero(::Type{Union{Missing, T}}) where T at missing.jl:105 zero(::Type{Missing}) at missing.jl:103 zero(::Type{LibGit2.GitHash}) at D:\buildbot\worker\package_win64\build\usr\share\julia\stdlib\v1.3\LibGit2\src\oid.jl:220

[simonmandlik]

This is most probably because Flux fails when computing gradient of sparse input when the model contains standard dense matrices. See here FluxML/Flux.jl#965

[racinmat]

This seems to be related also to FluxML/Zygote.jl#810 which is not fixed.

[racinmat]

Seems to be fixed right now, because currently, Flux is not using explicit cast