This behaves the same as in numpy:
- For dense tensors it is the product of the shape elements.
- For sparse tensors it is the number of elements actually stored
(including zeros that are explicitly stored).