Unable to get gradients of "Dense" models when sparse arrays are involved

[simonmandlik]

using Flux, SparseArrays

md = Dense(2,2)
ms = Dense(sparse(randn(2,2)), sparse(randn(2)))
xd = randn(2, 2)
xs = sparse(randn(2, 2))

gradient(() -> sum(md(xd)), Flux.params(md))
gradient(() -> sum(ms(xs)), Flux.params(ms))
gradient(() -> sum(ms(xd)), Flux.params(ms))
gradient(() -> sum(md(xs)), Flux.params(md))

The last call fails (on 0.10.0):

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 /Users/sabae/buildbot/worker/package_macos64/build/usr/share/julia/stdlib/v1.3/LibGit2/src/oid.jl:220
  ...
Stacktrace:
 [1] zero(::Type{Any}) at ./missing.jl:105
 [2] _zeros_eltypes at /Users/sabae/buildbot/worker/package_macos64/build/usr/share/julia/stdlib/v1.3/SparseArrays/src/higherorderfns.jl:203 [inlined]
 [3] _noshapecheck_map(::Zygote.var"#1457#1464", ::SparseMatrixCSC{Any,Int64}) at /Users/sabae/buildbot/worker/package_macos64/build/usr/share/julia/stdlib/v1.3/SparseArrays/src/highero
rderfns.jl:159
 [4] map(::Zygote.var"#1457#1464", ::SparseMatrixCSC{Any,Int64}) at /Users/sabae/buildbot/worker/package_macos64/build/usr/share/julia/stdlib/v1.3/SparseArrays/src/higherorderfns.jl:143
 [5] adjoint at /Users/simon.mandlik/.julia/packages/Zygote/8dVxG/src/lib/broadcast.jl:126 [inlined]
 [6] _pullback at /Users/simon.mandlik/.julia/packages/ZygoteRules/6nssF/src/adjoint.jl:47 [inlined]
 [7] adjoint at /Users/simon.mandlik/.julia/packages/Zygote/8dVxG/src/lib/lib.jl:139 [inlined]
 [8] _pullback at /Users/simon.mandlik/.julia/packages/ZygoteRules/6nssF/src/adjoint.jl:47 [inlined]
 [9] broadcasted at ./broadcast.jl:1231 [inlined]
 [10] Dense at /Users/simon.mandlik/.julia/packages/Flux/oX9Pi/src/layers/basic.jl:116 [inlined]
 [11] _pullback(::Zygote.Context, ::Dense{typeof(identity),Array{Float32,2},Array{Float32,1}}, ::SparseMatrixCSC{Float64,Int64}) at /Users/simon.mandlik/.julia/packages/Zygote/8dVxG/src
/compiler/interface2.jl:0
 [12] #103 at ./REPL[107]:1 [inlined]
 [13] _pullback(::Zygote.Context, ::var"#103#104") at /Users/simon.mandlik/.julia/packages/Zygote/8dVxG/src/compiler/interface2.jl:0
 [14] pullback(::Function, ::Params) at /Users/simon.mandlik/.julia/packages/Zygote/8dVxG/src/compiler/interface.jl:96
 [15] gradient(::Function, ::Params) at /Users/simon.mandlik/.julia/packages/Zygote/8dVxG/src/compiler/interface.jl:46
 [16] top-level scope at REPL[107]:1

On the other hand, computing gradients of only the matrix multiplication works for all possible combinations:

gradient(x -> sum(x*xd), xd)
gradient(x -> sum(x*xs), xs)
gradient(x -> sum(x*xs), xd)
gradient(x -> sum(x*xd), xs)

[simonmandlik]

Seems to not be a problem anymore on the latest release