Penn Jenks

I realize that a specific rule for SVector probably shouldn't be added to chain rules core -- but a more general solution for any reinterpreted composite types is the goal.

Does it make sense to add your suggestion to ChainRules?

Sorry that's a bad example. More generally i'm interested in reductions and accumulation over broadcasted functions i.e. `reduce(*, a .+ b)` or `accumulate(*, a .+ b)`. I see this as...

I think the CSR format is intended for sparse x dense matmul not sparse x sparse matmul. The following works as expected: ```julia using CUDA using SparseArrays CUDA.allowscalar(false) A =...

I missed this [_nearest_neighbors](https://github.com/FluxML/Flux3D.jl/blob/master/src/metrics/pcloud.jl#L72-L86) (gpu). I believe the advantages of the pytorch implementation are: - ~~Can use any distance metric~~ (the pytorch version uses euclidean distance) - Only the k...

FYI: [https://discourse.julialang.org/t/using-functions-in-gpu-kernel-via-kernelabstractions-jl-k-nearest-neighbor-kernel/53850](https://discourse.julialang.org/t/using-functions-in-gpu-kernel-via-kernelabstractions-jl-k-nearest-neighbor-kernel/53850)

Oops. Fixed it.

Is there a good way to change **all** bias initializers to zero?

I'm not pushing to add init args. I agree that it adds unnecessary complexity (and work :sweat_smile:)

Currently the `layers` attribute of `Sequential` can be a list -- which of course supports insert, append, ect ... However it may make sense to force this to be a...