Can't use Zygote.gradient on Symbolics.jl Matrix Functions

General Usage
,
1

When using a matrix function generated using Symbolics.jl variables, I get the error Mutating arrays is not supported when trying to autodifferentiate through the matrix function with either Zygote.gradient or Zygote.jacobian. Minimum working example is below, and full error stacktrace is at the bottom of this post.

MWE 
using Symbolics, Zygote
@variables x, y

M = [x y]
M_expr = build_function(M, [x,y])
M_func = eval(M_expr[1])

Zygote.gradient(z->sum(M_func(z)), [1,2]) # Error; full stacktrace at bottom of post

Julia version is 1.6.0.
Symbolics.jl is v0.1.12
Zygote.jl is v0.6.7

Is it possible to create matrix functions using Symbolics.jl that do not use mutation? Otherwise I’d like to know how to properly use Zygote.jl with matrix functions generated by Symbolics.jl.

Additional Details:

If I perform the sum() function on the symbolic matrix before calling build_function, then Zygote can take the gradient:

M_sum = sum(M)
M_sum_expr = build_function(M_sum, [x,y])
M_sum_func = eval(M_sum_expr)

Zygote.gradient(M_sum_func, [1,2]) # This works

I don’t think I have this option however–I’m trying to autodifferentiate through a DifferentialEquations.jl ODE of the form \dot{x} = A(p)x + Bu, y = C(p)x, where p contains the parameters I’m trying to differentiate with respect to. As far as I’m aware, the functions A(), C() can’t be passed into an ODE solver in their symbolic form, but I could be wrong.

Full Stacktrace 
ERROR: Mutating arrays is not supported
Stacktrace:
  [1] error(s::String)
    @ Base ./error.jl:33
  [2] (::Zygote.var"#399#400")(#unused#::Nothing)
    @ Zygote ~/.julia/packages/Zygote/fjuG8/src/lib/array.jl:58
  [3] (::Zygote.var"#2253#back#401"{Zygote.var"#399#400"})(Δ::Nothing)
    @ Zygote ~/.julia/packages/ZygoteRules/OjfTt/src/adjoint.jl:59
  [4] Pullback
    @ ~/.julia/packages/SymbolicUtils/aNxjZ/src/code.jl:383 [inlined]
  [5] (::Zygote.var"#178#179"{Tuple{Tuple{Nothing, Nothing, Nothing}, Tuple{Nothing, Nothing}}, typeof(∂(_create_array))})(Δ::FillArrays.Fill{Int64, 2, Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}})
    @ Zygote ~/.julia/packages/Zygote/fjuG8/src/lib/lib.jl:194
  [6] #1686#back
    @ ~/.julia/packages/ZygoteRules/OjfTt/src/adjoint.jl:59 [inlined]
  [7] Pullback
    @ ~/.julia/packages/SymbolicUtils/aNxjZ/src/code.jl:394 [inlined]
  [8] Pullback
    @ ~/.julia/packages/SymbolicUtils/aNxjZ/src/code.jl:371 [inlined]
  [9] (::typeof(∂(#11)))(Δ::FillArrays.Fill{Int64, 2, Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}})
    @ Zygote ~/.julia/packages/Zygote/fjuG8/src/compiler/interface2.jl:0
 [10] Pullback
    @ ./REPL[38]:1 [inlined]
 [11] (::typeof(∂(#13)))(Δ::Int64)
    @ Zygote ~/.julia/packages/Zygote/fjuG8/src/compiler/interface2.jl:0
 [12] (::Zygote.var"#41#42"{typeof(∂(#13))})(Δ::Int64)
    @ Zygote ~/.julia/packages/Zygote/fjuG8/src/compiler/interface.jl:41
 [13] gradient(f::Function, args::Vector{Int64})
    @ Zygote ~/.julia/packages/Zygote/fjuG8/src/compiler/interface.jl:59
 [14] top-level scope
    @ REPL[38]:1
2

Open an issue for this.

1 Like
3

Done. Issue has been opened here.

As an update, ForwardDiff.gradient works with M_func as defined above. For example:

julia> ForwardDiff.gradient(z->sum(M_func(z)),[1,2])
2-element Vector{Int64}:
 1
 1

I’m under the impression that Zygote is often the preferred package though.

4

Depends on the details of the function. But I think @dhairyagandhi96 just fixed this.

2 Likes