N-dimensional CNN models

[ToucheSir]

fastai has this already for certain models to support audio and time series processing. We should be able to do that but even better, because Flux's Conv and Pooling layers can (theoretically) handle an arbitrary number of dimensions.

[lorenzoh]

Yeah, it's just a matter of adding models now which take in 3D images. Projections and augmentations work on N-D images, masks, keypoints already; so does one-hot encoding N-D masks.

Then you should be able to do

sz = (64, 64, 64)
Method(
    blocks=(Image{3}(), Mask{3}(classes)),
    encodings=[
        ProjectiveTransforms(sz),
        ImagePreprocessing(),
        OneHot()
    ]
)
learner = methodlearner(method, data, Models.conv3dbackbone)

[ToucheSir]

Totally forgot I'd opened this issue! What do you think about an ndim parameter or something similar for XResNets and U-Nets to pass through to convx? Then instead of (ks, ks) on https://github.com/FluxML/FastAI.jl/blob/master/src/models/xresnet.jl#L5, you'd have ntuple(_ -> ks, ndim) (or maybe ntuple(Returns(ks), ndim) on 1.7+).

[lorenzoh]

Sounds like a good idea! I guess pooling layers would have to take that in as well 🤔

[lorenzoh]

See #181