Add dataclass_factory argument to dataclasses.make_dataclass for custom dataclass transformation support

[XuehaiPan]

Feature or enhancement

Proposal:

typing.dataclass_transform (PEP 681 – Data Class Transforms) allows users define their own dataclass decorator that can be recognized by the type checker.

Here is a real-world example use case:

• flax.struct.dataclass

Also, dataclasses.asdict and dataclasses.astuple allow users pass an extra argument for the factory of the returned instance.

cpython/Lib/dataclasses.py

Lines 1299 to 1317 in 0fb18b0

	def asdict(obj, *, dict_factory=dict):
	"""Return the fields of a dataclass instance as a new dictionary mapping
	field names to field values.
	Example usage::
	@dataclass
	class C:
	x: int
	y: int
	c = C(1, 2)
	assert asdict(c) == {'x': 1, 'y': 2}
	If given, 'dict_factory' will be used instead of built-in dict.
	The function applies recursively to field values that are
	dataclass instances. This will also look into built-in containers:
	tuples, lists, and dicts. Other objects are copied with 'copy.deepcopy()'.
	"""

cpython/Lib/dataclasses.py

Lines 1380 to 1397 in 0fb18b0

	def astuple(obj, *, tuple_factory=tuple):
	"""Return the fields of a dataclass instance as a new tuple of field values.
	Example usage::
	@dataclass
	class C:
	x: int
	y: int
	c = C(1, 2)
	assert astuple(c) == (1, 2)
	If given, 'tuple_factory' will be used instead of built-in tuple.
	The function applies recursively to field values that are
	dataclass instances. This will also look into built-in containers:
	tuples, lists, and dicts. Other objects are copied with 'copy.deepcopy()'.
	"""

However, the make_dataclass function does not support third-party dataclass factory (e.g., flax.struct.dataclass):

cpython/Lib/dataclasses.py

Lines 1441 to 1528 in 0fb18b0

	def make_dataclass(cls_name, fields, *, bases=(), namespace=None, init=True,
	repr=True, eq=True, order=False, unsafe_hash=False,
	frozen=False, match_args=True, kw_only=False, slots=False,
	weakref_slot=False, module=None):
	"""Return a new dynamically created dataclass.
	The dataclass name will be 'cls_name'. 'fields' is an iterable
	of either (name), (name, type) or (name, type, Field) objects. If type is
	omitted, use the string 'typing.Any'. Field objects are created by
	the equivalent of calling 'field(name, type [, Field-info])'.::
	C = make_dataclass('C', ['x', ('y', int), ('z', int, field(init=False))], bases=(Base,))
	is equivalent to::
	@dataclass
	class C(Base):
	x: 'typing.Any'
	y: int
	z: int = field(init=False)
	For the bases and namespace parameters, see the builtin type() function.
	The parameters init, repr, eq, order, unsafe_hash, frozen, match_args, kw_only,
	slots, and weakref_slot are passed to dataclass().
	If module parameter is defined, the '__module__' attribute of the dataclass is
	set to that value.
	"""
	if namespace is None:
	namespace = {}
	# While we're looking through the field names, validate that they
	# are identifiers, are not keywords, and not duplicates.
	seen = set()
	annotations = {}
	defaults = {}
	for item in fields:
	if isinstance(item, str):
	name = item
	tp = 'typing.Any'
	elif len(item) == 2:
	name, tp, = item
	elif len(item) == 3:
	name, tp, spec = item
	defaults[name] = spec
	else:
	raise TypeError(f'Invalid field: {item!r}')
	if not isinstance(name, str) or not name.isidentifier():
	raise TypeError(f'Field names must be valid identifiers: {name!r}')
	if keyword.iskeyword(name):
	raise TypeError(f'Field names must not be keywords: {name!r}')
	if name in seen:
	raise TypeError(f'Field name duplicated: {name!r}')
	seen.add(name)
	annotations[name] = tp
	# Update 'ns' with the user-supplied namespace plus our calculated values.
	def exec_body_callback(ns):
	ns.update(namespace)
	ns.update(defaults)
	ns['__annotations__'] = annotations
	# We use `types.new_class()` instead of simply `type()` to allow dynamic creation
	# of generic dataclasses.
	cls = types.new_class(cls_name, bases, {}, exec_body_callback)
	# For pickling to work, the __module__ variable needs to be set to the frame
	# where the dataclass is created.
	if module is None:
	try:
	module = sys._getframemodulename(1) or '__main__'
	except AttributeError:
	try:
	module = sys._getframe(1).f_globals.get('__name__', '__main__')
	except (AttributeError, ValueError):
	pass
	if module is not None:
	cls.__module__ = module
	# Apply the normal decorator.
	return dataclass(cls, init=init, repr=repr, eq=eq, order=order,
	unsafe_hash=unsafe_hash, frozen=frozen,
	match_args=match_args, kw_only=kw_only, slots=slots,
	weakref_slot=weakref_slot)

It can only apply dataclasses.dataclass (see the return statement above).

This feature request issue will discuss the possibility of adding a new dataclass_factory argument to the dataclasses.make_dataclass to support third-party dataclasss transformation, similar to dict_factory for dataclasses.asdict.

# dataclasses.py

def make_dataclass(cls_name, fields, *, bases=(), namespace=None, init=True,
                   repr=True, eq=True, order=False, unsafe_hash=False,
                   frozen=False, match_args=True, kw_only=False, slots=False,
                   weakref_slot=False, module=None,
                   dataclass_factory=dataclass):
    ...

    # Apply the normal decorator.
    return dataclass_factory(cls, init=init, repr=repr, eq=eq, order=order,
                             unsafe_hash=unsafe_hash, frozen=frozen,
                             match_args=match_args, kw_only=kw_only, slots=slots,
                             weakref_slot=weakref_slot)

Has this already been discussed elsewhere?

https://discuss.python.org/t/add-dataclass-factory-argument-to-dataclasses-make-dataclass-for-custom-dataclass-transformation-support/53188

Links to previous discussion of this feature:

No response

[ericvsmith]

That doesn't seem unreasonable to me. But as the issue template says, please discuss this on Discourse first: https://discuss.python.org/c/ideas/6

[XuehaiPan]

Thanks for the hint. I opened a new thread and let's discuss this there first.