Add output padding for ConvTranspose

[guiyrt]

This PR provides the ability to increase ConvTranspose's output shape when stride > 1, implemented in the same way as in PyTorch output_padding (see ConvTranspose3d and NaiveConvolutionTranspose3d.cpp).

The purpose of this is to conserve Conv/ConvTranspose inversability. For stride = N, N>1, there are N possibilities per dimension that result in the same Conv output (see what-output-padding-does-in-nn-convtranspose2d). The added argument outpad controls how much the output shape is increased to match the strided Conv counterpart. Can be provided as an integer to equally increase all dimensions, or as a tuple for individual dimensional control).

outpad is invalid if any outpad .>= stride. Currently there is no verification of this when constructing ConvTranspose, it only breaks when called and throws DimensionMismatch. A similar thing happens in PyTorch, it allows constructing ConvTransposeNd with invalid output_paddding, but then throws RuntimeError before doing the deconvolution. As the verification can be done in the constructor, it makes sense to handle it there, but let me know what you think.

I tried to follow the variable naming style, but it can also be changed if you see a better fitting alternative. The same applies for other things such as argument order.

PR Checklist

• Tests are added
• Entry in NEWS.md
• Documentation, if applicable

[CarloLucibello]

thanks!

[paulnovo]

Hi @CarloLucibello and @guiyrt, I was wondering if the output padding should also be applied in the AMDGPU implementation of conv_transpose_dims?

I am new to Flux, but I noticed this when rebasing my changes to ConvTranspose, so I thought I should ask.

[pxl-th]

@paulnovo yes, currently ConvTranspose is broken for AMDGPU because of this:
https://buildkite.com/julialang/flux-dot-jl/builds/4674#01903fca-3e24-4906-81a2-728846992074/361-742