Comment by kccqzy

I have an unpopular opinion: I don't like numpy.einsum because it is too different from the rest of numpy. You label your axes with letters but none of the other regular numpy functions do that. I usually avoid using numpy.einsum in favor of using a combination of indexing notation with numpy.newaxis (None), broadcasting, and numpy.swapaxes.

And I learned from someone more senior than me that you should instead label your variables with single-letter axes names. This way, the reader reads regular non-einsum operations and they still have the axes information in their mental model. And when you use numpy.einsum these axes labeling information become duplicated.