Support generic origins in transformation rules

[bharel]

For example np.ndarray[...], with a rule np.ndarray -> callable.

Right now we require every possible combination np.ndarray[...] to be in the rules which is not practical.

The fix is quite simple - just add another check after get_origin to see if it's in the rules.

[Udzu]

I'm not 100% sure what you mean, sorry! Can you not just use np.ndarray as the type if you don't care about the details? I.e. something like transform = { target_type: (np.ndarray, transformation_fn) }. Or am I misunderstanding?

[bharel]

I'm unable to create a transformation for the following:

class Foo:
 bar: np.ndarray[tuple[int,int], bool]

Deserializing using transform={np.ndarray: func} doesn't work:

  File "/Users/bar.harel/PycharmProjects/SimuRF/.venv/lib/python3.12/site-packages/dataglasses/core.py", line 181, in _from_dict
    elif not isinstance(value, cls):
             ^^^^^^^^^^^^^^^^^^^^^^
TypeError: isinstance() argument 2 cannot be a parameterized generic

That is because the hook checks for exact, and not for generic origin.

I can create transform={np.ndarray[tuple[int,int], bool]: func}, which will work.

However, that requires creating a transformation for every possible combination of ndarray[...] arguments if I wish to create a hook that simply converts all possible ndarrays.

[Udzu]

Makes sense! I've merged and released a likely fix — I've tested it with other generic types but not directly with np.ndarray. Does this fix your problem?

[bharel]

Works great 🎉
Next step is to resolve type aliases:

type MyGeneric = set[str]

@dataclass
class A:
 a: MyGeneric

In the case of the TypeAlias type, we should recurse from_dict on the .__value__.

Will open a new issue for this.