enable JSON serialization of dataclasses produced by hydra-zen builds

[cameronraysmith]

flytekit's dataclass transformer requires JSON-serializable dataclasses (docs). Currently, we construct JSON-serializable dataclasses for arguments or whole function interfaces along the lines of the pseudocode

from datclasses import make_dataclass
from dataclasses_json import DataClassJsonMixin
# OR
# from mashumaro.mixins.json import DataClassJSONMixin
from hydra_zen import builds
from sklearn.linear_model import LogisticRegression

# ...
# passing DataClassJsonMixin to bases via make_dataclass
logistic_regression_fields = create_dataclass_from_callable(
    LogisticRegression, custom_types_defaults
)

LRI_DataClass = make_dataclass(
    "LRI_DataClass",
    logistic_regression_fields,
    bases=(DataClassJsonMixin,),
)
LRI_DataClass.__module__ = __name__

approximating the usage in the logistic regression example, which, accounting for the dependency on

flytezen/src/flytezen/configuration.py

Lines 36 to 117 in 526203d

	def create_dataclass_from_callable(
	callable_obj: Callable,
	overrides: Optional[Dict[str, Tuple[Type, Any]]] = None,
	) -> List[Tuple[str, Type, Any]]:
	"""
	Creates the fields of a dataclass from a `Callable` that includes all
	parameters of the callable as typed fields with default values inferred or
	taken from type hints. The function also accepts a dictionary containing
	parameter names together with a tuple of a type and default to allow
	specification of or override (un)typed defaults from the target callable.
	Args:
	callable_obj (Callable): The callable object to create a dataclass from.
	overrides (Optional[Dict[str, Tuple[Type, Any]]]): Dictionary to
	override inferred types and default values. Each dict value is a tuple
	(Type, default_value).
	Returns:
	Fields that can be used to construct a new dataclass type that
	represents the interface of the callable.
	Examples:
	>>> from pprint import pprint
	>>> custom_types_defaults: Dict[str, Tuple[Type, Any]] = {
	... "penalty": (str, "l2"),
	... "class_weight": (Optional[dict], None),
	... "random_state": (Optional[int], None),
	... "max_iter": (int, 2000),
	... "n_jobs": (Optional[int], None),
	... "l1_ratio": (Optional[float], None),
	... }
	>>> fields = create_dataclass_from_callable(LogisticRegression, custom_types_defaults)
	>>> LogisticRegressionInterface = dataclasses.make_dataclass(
	... "LogisticRegressionInterface", fields, bases=(DataClassJSONMixin,)
	... )
	>>> lr_instance = LogisticRegressionInterface()
	>>> isinstance(lr_instance, DataClassJSONMixin)
	True
	>>> pprint(lr_instance)
	LogisticRegressionInterface(penalty='l2',
	dual=False,
	tol=0.0001,
	C=1.0,
	fit_intercept=True,
	intercept_scaling=1,
	class_weight=None,
	random_state=None,
	solver='lbfgs',
	max_iter=2000,
	multi_class='auto',
	verbose=0,
	warm_start=False,
	n_jobs=None,
	l1_ratio=None)
	"""
	if inspect.isclass(callable_obj):
	func = callable_obj.__init__
	else:
	func = callable_obj
	signature = inspect.signature(func)
	type_hints = get_type_hints(func)
	fields = []
	for name, param in signature.parameters.items():
	if name == "self":
	continue
	if overrides and name in overrides:
	field_type, default_value = overrides[name]
	else:
	inferred_type = infer_type_from_default(param.default)
	field_type = type_hints.get(name, inferred_type)
	default_value = (
	param.default
	if param.default is not inspect.Parameter.empty
	else dataclasses.field(default_factory=lambda: None)
	)
	fields.append((name, field_type, default_value))
	return fields

create_dataclass_from_callable, is verbose, even though it behaves as expected.

Based on the documentation for hydra_zen.builds zen_dataclass argument it seems like it should be possible to use the dataclasses constructed by hydra-zen instead

# passing DataClassJsonMixin to bases via zen_dataclass
Builds_LRI = builds(
    LogisticRegression,
    populate_full_signature=True,
    dataclass_name="Builds_LRI",
    zen_dataclass={"bases": (DataClassJsonMixin,), "module": __name__},
)

and eliminate create_dataclass_from_callable from flytezen altogether, but this produces

TypeError: dataclass option `bases` must be a tuple of dataclass types

[rsokl]

Sorry for lurking but I just came across this!

I think you can make inheritable via builds by making a dataclass-typed subclass of it

from dataclasses import dataclass

from dataclasses_json import DataClassJsonMixin as _DataClassJsonMixin

@dataclass
class DataClassJsonMixin(_DataClassJsonMixin): pass

then you can pass DataClassJsonMixin into bases

[cameronraysmith]

Sorry for lurking but I just came across this!

Not at all! Thank you for chiming in. I would have posted this in a discussion on the hydra-zen repository, but had not yet invested enough time to understand what I was missing, so I logged it here for future work.

I think you can make inheritable via builds by making a dataclass-typed subclass of it

from dataclasses import dataclass

from dataclasses_json import DataClassJsonMixin as _DataClassJsonMixin

@dataclass
class DataClassJsonMixin(_DataClassJsonMixin): pass

then you can pass DataClassJsonMixin into bases

That sounds like it should essentially lead to the resolution of this issue. Thanks again!