

Typesafe Tensors

The distinctive feature of Nexus is typesafe tensors.

In Nexus, a tensor is typed using a tuple of axis labels.

val image: FloatTensor[(Width, Height, Channel)]

This is in contrast with common deep learning libraries, where multidimensional arrays (tensors) all belong to one type (given the element type). For example, a tensor with float is typed as numpy.ndarray in NumPy and torch.FloatTensor in PyTorch, no matter how many dimensions are there (rank) in the tensor, or what each axis means in the tensor.