I will fully acknowledge my ignorance on swift, but how robust is the matrix operations/scientific packages? As someone who does DL research, I simply cannot imagine a programming environment where I don't have access to a lot of scientific matrix operations.

Some of my more clever DL approaches have used lots of complex windowing of data, and would be pretty stumped on how to implement these same preprocessing steps without numpy/scipy/skimage.

Since it was all written by Google it probably has only the packages they need...though in the interest of adoption perhaps not. You can just check in the repo on github.

OTOH you may not need some of those tools as the point of this project was actually as an environment supporting differentiable programming, so some of that slicing and dicing is approached automatically, at a higher level.