search

JuliaArrays / ArrayInterface.jl

Designs for new Base array interface primitives, used widely through scientific machine learning (SciML) and other organizations
https://docs.sciml.ai/ArrayInterface/stable/
MIT License
134 stars 37 forks source link

`axis_keys` #250

Open Tokazama opened 2 years ago

Tokazama commented 2 years ago

@rafaqz and @mcabbott, what do you think about having this method defined here to to interfaces together?

rafaqz commented 2 years ago

I like the idea, but I'm not sure about overloading the term "keys" as that already means something else for an array.

Tokazama commented 2 years ago

Do you have anything in mind that would more clearly convey what the method should return?

rafaqz commented 2 years ago

Sorry forgot to answer, but no. I've struggled with this too.

Tokazama commented 2 years ago

I'm pretty sure what we are all talking about is "keys" of some sort, but the concept of keys hasn't been formalized into a dedicated abstract interface in Julia yet. I think the definition of keys(::AbstractArray) in base is mostly so that functions can work generically with dictionaries and arrays, not necessarily a formalization of that interface. So my opinion is that anything that doesn't break those generic functions is fine. But I'm always open to alternative suggestions.

mcabbott commented 2 years ago

FWIW, I think calling these "keys" in my package was a mistake, as Base.keys already has a meaning. (And because these are auxiliary info, if you are using them to look up entries often, then you probably want a dictionary not an array.) "Labels" would have been better.

Tokazama commented 2 years ago

IRC, @ChrisRackauckas didn't like "labels" when this was discussed (well over a year ago).

Tokazama commented 2 years ago

I'm assuming this won't get more attention until after people have more time freed up after JuliaCon. I'd be fine if we changed from axes_keys to labels, but I think it is at least as problematic as using the term "keys". For example, if I have a time series recording of subject data I want the time to be the keys and events to be labels/annotations. There can also be multiple layers of conceptually independent labels. That has a place but I'm specifically thinking of what we call the key-value relationship.