Julia is still such an absolute joy to use. I have a dream that somehow, Torch.jl, which attempts to provide a julia interface to pytorch, will magically take off. Python is ok, but julia is just so fast, fun, and easy to scale.
Have you tried using PythonCall.jl (which allows you to basically use any random Python library, in principle) with Pytorch?
I haven't, but I have used PythonCall, and it's really rough to use and constantly go back and forth between types. The main reason I love julia is because all julia packages build upon julia arrays, and don't require conversion to special arrays for that specific package.