Carlo Lucibello

Those 2 usage examples seem pretty common. Moreover, I would find it very useful to have a `pyconvert(Any, x)` with a recursive behavior, something like what I poorly implemented [https://github.com/CarloLucibello/HuggingFaceDatasets.jl](https://github.com/CarloLucibello/HuggingFaceDatasets.jl/blob/main/src/transforms.jl)

If for many functions it is just a matter of broadening the allowed input types, would it make sense to just have SpecialFunctions depend on VectorizationBase?

That's a very nice project that we should definitely integrate (when it is AD and gpu ready)

> rename examples/gcn.jl --> examples/gcn_gpu.jl. (this actually does not run, but IDK if its my setup or a bug) > added examples/gcn.jl (original example but w/o the |> gpu parts...

It feels more natural to me to consider the graph as input data instead of part of the model. Also, layers shouldn't modify the internalized graph, it is unsafe since...

> For the same issue, there may be another approach to deal with this. Does parallelism being considered? in GNN the graph size is essentially equivalent to the batch size,...

that was very clear, thanks!

Can this be merged? Seems well tested and well documented

Agree, NNlib is used by different DL frameworks, it should have its own docs

@mcabbott thanks for all these suggestions, feel free to push on this branch. I like the `topk` - `topkperm` decoupling, so that it mimics base's sort functions. Representation of the...