Some gradient errors

[mcabbott]

I had a go at using Zygote on something, and got some errors, which survive in the following simplified examples:

using Flux, Zygote

function ff(x, t = [1,2,3])  
    a = dropdims(sum( x .^ t' , dims=1), dims=1)
    sum(x)
end

Flux.gradient(ff, ones(3))
Zygote.gradient(ff, ones(3)) # ERROR: Compiling Tuple{typeof(in),Int64,Tuple{Int64}}: BoundsError: attempt to access 56-element Array{Any,1} at index [84]

and

function hh(x)  
    a = sum( x .^ [1 2 3] , dims=1)
    sum(x)
end

Flux.gradient(hh, ones(3))
Zygote.gradient(hh, ones(3)) # ERROR: UndefVarError: T not defined

I also got the same error as #47, but only when using a custom type; could try to boil that down if it would be helpful.

All on master of today, and Julia 1.1.

[MikeInnes]

On Julia master I get

julia> function ff(x, t = [1,2,3])
           a = dropdims(sum( x .^ t' , dims=1), dims=1)
           sum(x)
       end
ff (generic function with 2 methods)

julia> ff(ones(3))
3.0

julia> gradient(ff, ones(3))
(Int8[1, 1, 1],)

julia> function hh(x)
           a = sum( x .^ [1 2 3] , dims=1)
           sum(x)
       end
hh (generic function with 1 method)

julia> gradient(hh, ones(3))
ERROR: Mutating arrays is not supported

Closing this as fixed since the compiler errors are gone; of course the second example should work ideally, but that's just a case of #81.

[mcabbott]

Thanks, sounds good. I have one more case which boils down to the following:

julia> using Zygote

julia> y, back = Zygote._forward((m,v) -> (v .+ v' .+ m)[1,1], rand(3,3), rand(3))
(1.7873601043525784, ∂(getfield(Main, Symbol("##43#44"))()))

julia> back(0.666)
ERROR: BoundsError: attempt to access 0-element Array{Core.Compiler.BasicBlock,1} at index [1]
Stacktrace:
 [1] getindex at ./array.jl:729 [inlined]
 [2] iterate(::Core.Compiler.IncrementalCompact, ::Tuple{Int64,Int64}) at ./compiler/ssair/ir.jl:1077
 [3] iterate at ./compiler/ssair/ir.jl:1070 [inlined]

[MikeInnes]

Yup, you found another compiler bug there. Will reopen in that case. (You are also free to open new issues for new errors though).