Getting started

We use a functional API and try to follow most of PyTorch's convention. Check out the API and start playing.

Notable differences include the way sampling coordinates are encoded ([0, N-1] instead of [-1, 1], no index flipping), and the fact that we often use a "channel last" dimension ordering.