# Zygote unexpected broadcasting

**URL:** https://discourse.julialang.org/t/zygote-unexpected-broadcasting/46476
**Category:** Machine Learning
**Tags:** zygote
**Created:** [September 11, 2020, 8:28pm UTC](https://discourse.julialang.org/t/zygote-unexpected-broadcasting/46476 "2020-09-11T20:28:19Z")
**Posts on this page:** 2
**Page:** 1

<div class="post-metadata">

### Author: ![merckxiaan](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/merckxiaan/32/4697_2.png) [@merckxiaan](https://discourse.julialang.org/u/merckxiaan)
#### Post date: [September 11, 2020, 8:28pm UTC](https://discourse.julialang.org/t/zygote-unexpected-broadcasting/46476/1 "2020-09-11T20:28:19Z")

</div>

This is a MWE for a problem i’m having, I’ve adapted matrix multiplication to accept `trans` arguments telling whether the matrix is to be transposed.  
As for the adjoint definition, I looked at [`NNlib.batched_mul`](https://github.com/FluxML/Zygote.jl/blob/b269ed3d3328be201c7ff73947cdcfa69c6e3cd5/src/lib/nnlib.jl#L100)

When I try to get the gradient of x.x^T w.r.t. x using `my_mul`, get an error shown below.

```julia
x = rand(2, 3)

function my_mul(A, B, transA, transB) #transA/transB is either 'N' or 'T'
    (transA == 'N' ? A : A')*(transB == 'N' ? B : B')
end

my_mul(x, x, 'N', 'T')

using Zygote

Zygote.@adjoint function my_mul(A, B, transA, transB)
    d = Dict('N'=>'T', 'T'=>'N')

    C = my_mul(A, B, transA, transB)
    C, Δ -> (my_mul(Δ, B, 'N', d[transB]), my_mul(A, Δ, d[transA], 'N'))
end

Zygote.gradient(x) do x
    sum(x*x')
end # expected result

Zygote.gradient(x) do x
    sum(my_mul(x, x, 'N', 'T'))
end # unexpected!

```

```julia
ERROR: DimensionMismatch("arrays could not be broadcast to a common size; got a dimension with lengths 2 and 3")
Stacktrace:
 [1] _bcs1 at .\broadcast.jl:501 [inlined]
 [2] _bcs at .\broadcast.jl:495 [inlined]
 [3] broadcast_shape at .\broadcast.jl:489 [inlined]
 [4] combine_axes at .\broadcast.jl:484 [inlined]  
 [5] instantiate at .\broadcast.jl:266 [inlined]   
 [6] materialize at .\broadcast.jl:837 [inlined]   
 [7] accum(::Array{Float64,2}, ::Array{Float64,2}) 
at C:\Users\jules\.julia\packages\Zygote\Xgcgs\src\lib\lib.jl:16
 [8] (::Zygote.var"#41#42"{typeof(∂(#13))})(::Float64) at C:\Users\jules\.julia\packages\Zygote\Xgcgs\src\compiler\interface.jl:45
 [9] gradient(::Function, ::Array{Float64,2}) at C:\Users\jules\.julia\packages\Zygote\Xgcgs\src\compiler\interface.jl:54
 [10] top-level scope at REPL[1]:1
 [11] include_string(::Function, ::Module, ::String, ::String) at .\loading.jl:1088

```

I believe the failed broadcasting is coming from `Zygote.accum` which tries to add both the adjoint results together.

Note that this only happens if I’m multiplying an array with itself, that is, replacing one of the `x`’s with a copy fixes it.  
When I take the gradient of `sum(x*x')` I do get a sensible answer.

Does anybody know what I’m doing wrong here?

Thanks in advance,  
Jules

---

<div class="post-metadata">

### Author: ![merckxiaan](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/merckxiaan/32/4697_2.png) [@merckxiaan](https://discourse.julialang.org/u/merckxiaan)
#### Post date: [September 15, 2020, 9:47am UTC](https://discourse.julialang.org/t/zygote-unexpected-broadcasting/46476/2 "2020-09-15T09:47:48Z")

</div>

Sorry for the noise, this problem arises because I’m implicitly transposing a matrix so this needs to be reflected in the adjoint definition.