We construct a computational graph and need to validate (or search for possible new) edges at runtime. The data is specific to computer vision. One example is clipping geometry within 2D boundaries. A minimal implementation of this data would be:

@dataclass
class FramedGeometry:
    width: PositiveInt
    height: PositiveInt
    geometry: Geometry

However, different properties are manifest in different encodings. Height, width can be meta data from an image database or inherent to a decoded image (as np.ndarray.shape). The transform will then `dataclasses.replace(geometry=...)` its attribute(s).. If width and height are not implemented, another transition is needed to produce them whilst only looking into the image header, not fully decoding potentially large and many images.

The read-only interface ensures that transitions are generic wrt some forms of inputs. The replace interface preserves runtime types.

Inputs and outputs are annotated with @dataclass or tuples of them. Those dataclasses are a mixin of base dataclasses that declare concrete properties like a URI of an image and ABCs that declare accessors like get_width(self) -> PositiveInt. We use @pydantic.dataclass to parse, validate and deserialize concrete classes to great extent [0]. 

In order to not implement accessors on top of dataclasses, we'd want that abstract properties are compatible with dataclasses and issubclass works with data protocols (given the necessary constraints).

PS:
Polymorphism for computer-vision data is an interesting problem. Other approaches exist, each with their own struggle to model "traits" the right way [1]. E.g., scaling would be valid for `FramedGeometry` since no image data is included but invalid when images are referenced but cannot be resized, like:

class EncodedSizedAnnotatedFrame:
    width: PositiveInt
    height: PositiveInt
    image_bin: bytes
    geometry: Geometry

Thanks!

[0] https://pydantic-docs.helpmanual.io/usage/dataclasses/
[1] https://github.com/pytorch/vision/issues/1406