Open alan-cooney opened 1 year ago
Interesting idea! I agree that'd make sense.
In practice though, TorchTyping is now semi-deprecated in favour of jaxtyping. (Which supports PyTorch too, despite the name.)
It also supports static type checkers etc., doesn't do crazy monkey-patching of typeguard, etc, which is the reason I'm trying to transition people over to it.
In terms of this suggestion: jaxtyping does go all-in on using a string-based DSL. I imagine you could accomplish something similar to this suggestion by simply passing around strings.
For larger projects,
TensorType[T.BATCH]
(whereT
is an enum) can be better thanTensorType["batch"]
, as it enables standard naming conventions & additional code hints.It would be great to support this out of the box.