[RFC] A Named-Dimensions TypeScript Library for TensorFlowJS

[Lucas Dixon]

Hi everyone,

I'm excited to share an RFC about A Named-Dimensions TypeScript Library for TensorFlowJS that is now available for comment. 

The feedback phase will be open for 4 weeks, until 2021-09-13. Comments are invited to the pull request linked. You can view formatted RFC, and also leave comments inline on the RFC document sources, including a basic proof of concept implementation, and a few examples.

(Cross posting discuss & developers, as broad API interest. My understanding is that usually these notices just go to developers@, where you should definitely join if you are interested in reviewing design proposals.)

Summary: Support the naming of dimensions in TensorFlow JS to provide an abstraction that makes it easier for developers and researchers to write and read ML algorithms (and support some automated optimisation):

• Make it harder to make subtle mistakes (have type-checking give intelligible type-error messages)
• Leveraging TypeScript IDE support to get better auto-completion
• Make it easier to support writing an algorithm and then generalise it to support batches.
• Have code more closely resemble the pseudocode and algorithms in papers.
• Allow machine optimization of the ordering of indexes, transpositions and permutations.

Thank you for your review and participation!

--

Lucas Dixon : PAIR / Responsible AI & Brain / Google Research