ConvTranspose errors with symmetric non-constant pad

[piever]

Description

The ConvTranspose layer errors if given a symmetric non-constant pad (that is, length(pad) == ndims(w) - 2), whereas pad::Int (constant pad) and length(pad) == 2 * (ndims(w) - 2) (non-symmetric non-constant pad) work fine.

MWE:

using Flux

c = ConvTranspose((3, 3), 1 => 1, pad = (1, 0))

x = randn(Float32, 5, 5, 1, 16)

c(x)

errors with

ERROR: MethodError: no method matching DenseConvDims(::Tuple{Int64, Int64, Int64}, ::NTuple{4, Int64}; stride::Tuple{Int64, Int64}, padding::Tuple{Int64, Int64}, dilation::Tuple{Int64, Int64}, groups::Int64)

Closest candidates are:
  DenseConvDims(::Tuple{Vararg{Int64, N}}, ::Tuple{Vararg{Int64, K}}, ::Int64, ::Int64, ::Int64, ::Tuple{Vararg{Int64, S}}, ::Tuple{Vararg{Int64, P}}, ::Tuple{Vararg{Int64, D}}, ::Bool) where {N, K, S, P, D} got unsupported keyword arguments "stride", "padding", "dilation", "groups"
   @ NNlib C:\Users\pietro\.julia\packages\NNlib\O0zGY\src\dim_helpers\DenseConvDims.jl:7
  DenseConvDims(::Tuple{Vararg{T, M}} where T, ::Tuple{Vararg{T, M}} where T; stride, padding, dilation, groups, flipkernel) where M
   @ NNlib C:\Users\pietro\.julia\packages\NNlib\O0zGY\src\dim_helpers\DenseConvDims.jl:20

Stacktrace:
 [1] conv_transpose_dims(c::ConvTranspose{2, 2, typeof(identity), Array{Float32, 4}, Vector{Float32}}, x::Array{Float32, 4})
   @ Flux C:\Users\pietro\.julia\packages\Flux\Wz6D4\src\layers\conv.jl:323
 [2] (::ConvTranspose{2, 2, typeof(identity), Array{Float32, 4}, Vector{Float32}})(x::Array{Float32, 4})
   @ Flux C:\Users\pietro\.julia\packages\Flux\Wz6D4\src\layers\conv.jl:336
 [3] top-level scope
   @ REPL[6]:1

Environment

The above reproducer works in Julia 1.10.2 on a fresh environment

(jl_tMb72B) pkg> st
Status `C:\Users\pietro\AppData\Local\Temp\jl_tMb72B\Project.toml`
  [587475ba] Flux v0.14.15

Suggestion

I suspect the problem is in combined_pad (here and here). It assumes that c.pad has length 2 * (ndims(c.weight) - 2), which is not the case.