Add support for generic types

[alexmirrington]

Overview:

Serialiasation support for generic types would allow for serialisation of more complex nested classes/dataclasses

Example:

Consider the following:

SplitType = TypeVar("SplitType")
 
@dataclass(frozen=True)
class DatasetConfig(Generic[SplitType]):
    """A class specifying the common fields used across all datasets."""
 
    split: SplitType
 
@dataclass(frozen=True)
class FirstDatasetConfig(DatasetConfig[FirstDatasetSplit]):
    """A class specifying fields for the first dataset."""
 
    version: str

@dataclass(frozen=True)
class SecondDatasetConfig(DatasetConfig[SecondDatasetSplit]):
    """A class specifying fields for the second dataset."""

    features: List[str]

class FirstDatasetSplit(Enum):
    TRAIN: "train"
    TEST: "test"
    CHALLENGE: "challenge"

class FirstDatasetSplit(Enum):
    TRAIN: "train"
    VAL: "val"
    TEST: "test"
    DEV: "testdev"

The general problem we aim to solve in this case is that we want to define shared fields in a common superclass, but the type of that field is often dependent on the subclass implementation. In the example above, we know that all datasets have a split attribute, however the values that these can take are different depending on the type of dataset.

A workaround for this is to perform checks in a __post_init__ hook, to ensure the value of split is valid, however this often means working with less rigid types like strings instead of enums.

[ramonhagenaars]

Good idea @alexmirrington.

[JobaDiniz]

I assumed Generics were supported since this lib supports generic Lists for example.