Why Isn’t There a Standard for Typing Array Dimensions in Python?

PEP 646 introduced a feature called `TypeVarTuple`, aimed at addressing this issue. However, I'm not sure how much it's being utilized in libraries or supported by type checkers yet.

I totally relate to this frustration! When I started with machine learning, I would receive functions that accepted multiple arrays, but they rarely specified their expected shapes. The docstring might say ‘a vector of dimension d,’ but the reality is that sometimes, vectors are treated as 2D in ML. For me, clear type hints for array dimensions would be a game changer, especially since functions often deal with arrays that must meet strict shape dependencies.

NumPy does support specifying shape types, though it’s a bit cumbersome. For example, to define a 100x100x3 float64 array, you'd write `np.ndarray[tuple[Literal[100], Literal[100], Literal[3]], np.dtype[np.float64]]`. It hasn't really caught on because it's clunky, and the benefits aren’t substantial enough for widespread adoption.

There's actually jaxtyping, which works with both NumPy and Torch. I've found it really useful for my machine learning projects. It helps clarify dimensions, especially when you’re working with JAX and using functions like vmap.

The dimensionality really depends on the value; you can create your own annotations to enforce this. For example, check out how `astropy.units.quantity_input` handles it. Although, this may vary based on how expressive your type system is.