Conversation

A neat proposal by

, and

on naming the axes of matrices in order to aid understanding: namedtensor.github.io Reminds of how I often prefer named parameters and records in programming over their positional alternatives. 🤔

A snippet showing Section 2 of the linked article on "Named Tensor Notation".

read image description

ALT

Brendan Zabarauskas

@brendanzab

It's interesting to note that many numerical libraries and languages seem to allow naming the axes of arrays these days. I'm guessing might be where the inspiration came from?

@contextfree

compared and contrasted some of these this video:

youtube.com

ND Arrays (Tensors) in different languages

Code: https://github.com/contextfreecode/tensorThanks much to Casper Sørensen for hi and bye (and Bjarne Stroustrup) in Danish! - https://theparallelthread.c...

3:44 AM · Feb 11, 2021Twitter Web App

Replying to

I also think the 'index type' extension of the proposal pretty neat - reminds me of how Pascal lets you define the type of array indices: freepascal.org/docs-html/ref/ - I often wish Rust supported this!

A snippet of Section 7 of the "Named Tensor Notation" article linked in the first tweet of the thread.

read image description

ALT

Replying to

and

Yeah, worked on this with the torch folks a couple years ago. But we had trouble convincing them to remove orders as well (since it matters for speed). Until there are better JITs, we're re-focusing on notation.

Brendan Zabarauskas

@brendanzab

Feb 11, 2021

Yeah, ultimately removing the restriction on ordering seems super handy. Seems like the matrices would then become something like homogeneous nested records?