python3 dataclass with **kwargs(asterisk)

Dataclass only relies on the __init__ method so you’re free to alter your class in the __new__ method.

from dataclasses import dataclass


@dataclass
class Container:
    user_id: int
    body: str

    def __new__(cls, *args, **kwargs):
        try:
            initializer = cls.__initializer
        except AttributeError:
            # Store the original init on the class in a different place
            cls.__initializer = initializer = cls.__init__
            # replace init with something harmless
            cls.__init__ = lambda *a, **k: None

        # code from adapted from Arne
        added_args = {}
        for name in list(kwargs.keys()):
            if name not in cls.__annotations__:
                added_args[name] = kwargs.pop(name)

        ret = object.__new__(cls)
        initializer(ret, **kwargs)
        # ... and add the new ones by hand
        for new_name, new_val in added_args.items():
            setattr(ret, new_name, new_val)

        return ret


if __name__ == "__main__":
    params = {'user_id': 1, 'body': 'foo', 'bar': 'baz', 'amount': 10}
    c = Container(**params)
    print(c.bar)  # prints: 'baz'
    print(c.body)  # prints: 'baz'`

Here’s a neat variation on this I used.

from dataclasses import dataclass, field
from typing import Optional, Dict


@dataclass
class MyDataclass:
    data1: Optional[str] = None
    data2: Optional[Dict] = None
    data3: Optional[Dict] = None

    kwargs: field(default_factory=dict) = None

    def __post_init__(self):
        [setattr(self, k, v) for k, v in self.kwargs.items()]

This works as below:

>>> data = MyDataclass(data1="data1", kwargs={"test": 1, "test2": 2})
>>> data.test
1
>>> data.test2
2

However note that the dataclass does not seem to know that is has these new attributes:

>>> from dataclasses import asdict
>>> asdict(data)
{'data1': 'data1', 'data2': None, 'data3': None, 'kwargs': {'test': 1, 'test2': 2}}

This means that the keys have to be known. This worked for my use case and possibly others.

from dataclasses import make_dataclass
Clas = make_dataclass('A', 
                      ['d'], 
                      namespace={
                                 '__post_init__': lambda self: self.__dict__.update(self.d)
                      })
d = {'a':1, 'b': 2}
instance = Clas(d)
instance.a

Variation of answer from Trian Svinit:

You could use the following approach:

1. Extra attributes are added via a kwargs argument as such: MyDataclass(xx, yy, kwargs={...})
2. kwargs is a dataclasses.InitVar that is then processed in the __post_init__ of your dataclass
3. You can access all the values with instance.__dict__ (because asdict would not detect the attributes added via kwargs=...

This would only use native features from dataclasses and inheriting this class would still work.

from dataclasses import InitVar, asdict, dataclass
from typing import Dict, Optional


@dataclass
class MyDataclass:
    data1: Optional[str] = None
    data2: Optional[Dict] = None
    data3: Optional[Dict] = None

    kwargs: InitVar[Optional[Dict[str, Any]]] = None

    def __post_init__(self, kwargs: Optional[Dict[str, Any]]) -> None:
        if kwargs:
            for k, v in kwargs.items():
                setattr(self, k, v)


data = MyDataclass(data1="data_nb_1", kwargs={"test1": 1, "test2": 2})
print(data, "-", data.data1, "-", data.test1)
# MyDataclass(data1='data_nb_1', data2=None, data3=None) - data1 - 1
print(asdict(data))
# {'data1': 'data_nb_1', 'data2': None, 'data3': None}
print(data.__dict__)
# {'data1': 'data_nb_1', 'data2': None, 'data3': None, 'test1': 1, 'test2': 2}

If you really need to use asdict to get the attributes passed as kwargs, you could start to use private attributes in dataclasses to hack asdict:

from dataclasses import _FIELD, _FIELDSInitVar, asdict, dataclass, field
from typing import Dict, Optional


@dataclass
class MyDataclass:
    data1: Optional[str] = None
    data2: Optional[Dict] = None
    data3: Optional[Dict] = None

    kwargs: InitVar[Optional[Dict[str, Any]]] = None

    def __post_init__(self, kwargs: Optional[Dict[str, Any]]) -> None:
        if kwargs:
            for k, v in kwargs.items():
                setattr(self, k, v)
                self._add_to_asdict(k)

    def _add_to_asdict(self, attr:str) -> None:
        """Add an attribute to the list of keys returned by asdict"""
        f = field(repr=True)
        f.name = attr
        f._field_type = _FIELD
        getattr(self, _FIELDS)[attr] = f

data = MyDataclass(data1="data_nb_1", kwargs={"test1": 1, "test2": 2})
print(asdict(data))
# {'data1': 'data_nb_1', 'data2': None, 'data3': None, 'test1': 1, 'test2': 2}

Based on Arnes Answer, I create a class decorator which extends the dataclass decorator with the from_kwargs method.

from dataclasses import dataclass
from inspect import signature


def dataclass_init_kwargs(cls, *args, **kwargs):
    cls = dataclass(cls, *args, **kwargs)

    def from_kwargs(**kwargs):
        cls_fields = {field for field in signature(cls).parameters}
        native_arg_keys = cls_fields & set(kwargs.keys())
        native_args = {k: kwargs[k] for k in native_arg_keys}
        ret = cls(**native_args)
        return ret

    setattr(cls, 'from_kwargs', from_kwargs)
    return cls

All of these changes are well-meaning but pretty clearly against the spirit of dataclasses, which is to avoid writing a bunch of boilerplate to set up a class.

Python 3.10 introduces the match statement and with it dataclasses get a match_args=True default argument in the constructor (i.e. the decorator).

This means that you get a dunder attribute __match_args__ which stores a tuple of the init (kw)args, importantly without runtime inspection.

So you can just create a classmethod

from dataclasses import dataclass

@dataclass
class A:
    a: int
    b: int = 0
    
    def from_kwargs(cls, **kwargs: dict) -> A:
        return cls(**{k: kwargs[k] for k in kwargs if k in cls.__match_args__})

It works:

>>> A.from_kwargs(a=1, b=2, c=3)
A(a=1, b=2)
>>> A.from_kwargs(a=1)
A(a=1, b=0)

However we also have access to these same keys in Python 3.9 thanks to __dataclass_fields__, which is the next best option if you can’t rely on the 3.10 runtime.

    def from_kwargs(cls, **kwargs: dict) -> A:
        return cls(**{k: kwargs[k] for k in kwargs if k in cls.__dataclass_fields__})

This gives the same result.

For the (unusual but reasonable!) use case in the question, you can just change the class method to pop rather than access the kwargs dict when building the init_kw dict, so that the remaining keys will be left over in kwargs and can be passed as their own kwarg, rest.

from dataclasses import dataclass

@dataclass
class A:
    a: int
    b: int = 0
    rest: dict = {}
    
    def from_kwargs(cls, **kwargs: dict) -> A:
        init_kw = {k: kwargs.pop(k) for k in dict(kwargs) if k in cls.__match_args__}
        return cls(**init_kw, rest=kwargs)

Note that you have to wrap the kwargs in a call to dict (make a copy) to avoid the error of "dict size changed during iteration"

For this problem, you should use default_factory, as it states in the dataclass documentation.

@dataclass
class Foo:
    a: Dict = field(default_factory=dict)