Open DrChainsaw opened 1 week ago
Could this be the same issue as #2424 which is fixed by PR-2463?
If I understand how padding affects the ConvTranspose output size, your example will result in a negative output size, but the following should work with the changes in PR-2463:
julia> ConvTranspose((1,1), 1=>1; pad=(1,1))(ones(Float32, 3, 3, 1,1))
1×1×1×1 Array{Float32, 4}:
[:, :, 1, 1] =
-1.3000358
Not something I need to do, but the NaiveNASflux tests hits this.
Haven't dug further into it, but I guess it just hits some obsolete codepath: