Alternatives to zero-padding conv layers?

Specific Domains Machine Learning
1

Is there a way to use same-padding or reflection-padding in Flux convolutional layers, such as NNlib.jl supplies in https://github.com/FluxML/NNlib.jl/blob/master/src/padding.jl, rather than only using zero-padding?

I’m seeing significant edge-effects in an image GAN I’m trying to compose and wondered if using one of these other padding regimens would help.

2

Can you file an issue on NNlib so we can track this? In the meantime, applying the pad_* functions before a conv layer is not too bad of a workaround.

3

I finally got back to looking at this project a couple of weeks ago. This is the best construction I could come up with which let me use it the same as a standard convolutional layer:

struct rConv
    a
    layer
end

@functor rConv

function rConv(k::NTuple{N,Integer}, ch::Pair{<:Integer,<:Integer}, σ = identity;
        init = Flux.glorot_uniform, stride = 1, pad = 0, dilation = 1, groups = 1,
        bias = true) where N
    
    a = Flux.calc_padding(Flux.Conv,pad,k,dilation,stride)
    weight = Flux.convfilter(k, ch; init, groups)

    layer = Flux.Conv(weight, bias, σ; stride=stride, pad=0, dilation=dilation, groups=groups)

    rConv(a,layer)
end

function (m::rConv)(x)
    m.layer(NNlib.pad_reflect(x,m.a))
end

Not sure if it’s the most elegant solution, but it works for me and definitely reduces the edge effects in my specific project.

Edit to say: I’ve not tried using this on a gpu, so not sure if it will work on that.

4

Glad you figured it out, I don’t see any reason it wouldn’t work on GPU :slight_smile:. We’ve talked about what the interface for fancier padding in Conv might look like, but never arrived at a design. If you have any ideas there, let us know on GitHub!