25

Currently I used DTO(Data Transfer Object) like this.

class Test1:
    def __init__(self, 
        user_id: int = None,
        body: str = None):
        self.user_id = user_id
        self.body = body

Example code is very small, But when object scale growing up, I have to define every variable.

While digging into it, found that python 3.7 supported dataclass

Below code is DTO used dataclass.

from dataclasses import dataclass


@dataclass
class Test2:
    user_id: int
    body: str

In this case, How can I allow pass more argument that does not define into class Test2?

If I used Test1, it is easy. Just add **kwargs(asterisk) into __init__

class Test1:
    def __init__(self, 
        user_id: int = None,
        body: str = None,
        **kwargs):
        self.user_id = user_id
        self.body = body

But using dataclass, Can't found any way to implement it.

Is there any solution here?

Thanks.

EDIT

class Test1:
    def __init__(self,
        user_id: str = None, 
        body: str = None):
        self.user_id = user_id
        self.body = body

if __name__ == '__main__':
    temp = {'user_id': 'hide', 'body': 'body test'}
    t1 = Test1(**temp)
    print(t1.__dict__)

Result : {'user_id': 'hide', 'body': 'body test'}

As you know, I want to insert data with dictionary type -> **temp

Reason to using asterisk in dataclass is the same.

I have to pass dictinary type to class init.

Any idea here?

Improve this question
6
  • 1
    Pass init=False to the data-class decorator and implement __init__ by hand. Sort of defeats the purpose, and without explicitly defining the attributes your data-class won't use those variables. Essentially, you don't want a dataclass. Commented Mar 11, 2019 at 10:13
  • 1
    @juanpa.arrivillaga If I implement __init__ by hand, there is no advantage for use dataclass. Is it a only way? Commented Mar 11, 2019 at 10:14
  • 4
    How do you expect the dataclass to know what variables to use without you explicitly enumerating them? If you want dynamic attributes (generally a poor design choice IMO), you don't really want a dataclass. Commented Mar 11, 2019 at 10:16
  • @juanpa.arrivillaga I added more information about my question. Please refer it. Commented Mar 11, 2019 at 13:53
  • 1
    Yes, that doesn't change anything, although note, you don't need **kwargs to do that, i.e. pass a dict Commented Mar 11, 2019 at 15:34

8 Answers 8

Reset to default
28

The basic use case for dataclasses is to provide a container that maps arguments to attributes. If you have unknown arguments, you can't know the respective attributes during class creation.

You can work around it if you know during initialization which arguments are unknown by sending them to a catch-all attribute by hand:

from dataclasses import dataclass, field


@dataclass
class Container:
    user_id: int
    body: str
    meta: field(default_factory=dict)


# usage:
obligatory_args = {'user_id': 1, 'body': 'foo'}
other_args = {'bar': 'baz', 'amount': 10}
c = Container(**obligatory_args, meta=other_args)
print(c.meta['bar'])  # prints: 'baz'

But in this case you'll still have a dictionary you need to look into and can't access the arguments by their name, i.e. c.bar doesn't work.

If you care about accessing attributes by name, or if you can't distinguish between known and unknown arguments during initialisation, then your last resort without rewriting __init__ (which pretty much defeats the purpose of using dataclasses in the first place) is writing a @classmethod:

from dataclasses import dataclass
from inspect import signature


@dataclass
class Container:
    user_id: int
    body: str

    @classmethod
    def from_kwargs(cls, **kwargs):
        # fetch the constructor's signature
        cls_fields = {field for field in signature(cls).parameters}

        # split the kwargs into native ones and new ones
        native_args, new_args = {}, {}
        for name, val in kwargs.items():
            if name in cls_fields:
                native_args[name] = val
            else:
                new_args[name] = val

        # use the native ones to create the class ...
        ret = cls(**native_args)

        # ... and add the new ones by hand
        for new_name, new_val in new_args.items():
            setattr(ret, new_name, new_val)
        return ret

Usage:

params = {'user_id': 1, 'body': 'foo', 'bar': 'baz', 'amount': 10}
Container(**params)  # still doesn't work, raises a TypeError 
c = Container.from_kwargs(**params)
print(c.bar)  # prints: 'baz'
Improve this answer
Sign up to request clarification or add additional context in comments.

8 Comments

I added more information about my question. Please refer it.
I'm afraid there is no way to pass unknown arguments to a dataclass constructor without rewriting it, in which case it's probably not worth it any more to use dataclasses at all. The only other option would be to provide a classmethod, I can add that to my answer.
It's also worth noting that we must exclude ClassVar-annotated attributes since those don't get included in the generated __init__. Something like if getattr(cls.__annotations__.get(name, None), "__origin__", None) is not ClassVar:.
@Anakhand good catch, thanks. I added a similar check.
@Felix It's fixed now, thanks for the heads up
|
12

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'`
Improve this answer

2 Comments

But initializer(ret, *args, **kwargs)
If you're doing inheritance, you will run into a problem with this solution. You also need to save the class name so that subclasses will know not to reuse cls.__initializer. cls.__class_name = cls.__name__
7

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.

Improve this answer

4 Comments

Nice idea. Note that 1. you can use data.__dict__ if you want to access your kwargs fields. 2. it's slightly strange to put field as typing hint: you would better replace by InitVar (part of dataclasses). I tried to submit an alternative solution based on your! ^_^
... and also, it is better to use list comprehensions to store a processed list, and to avoid it when you need to process elements (in which case, a standard loop is clearer).
*Don't use list comprehensions for side-effects: [setattr(self, k, v) for k, v in self.kwargs.items()] just use a for-loop
I think you meant kwargs: dict = field(default_factory=dict)
3

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"

Improve this answer

Comments

2 
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
Improve this answer

Comments

1

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}
Improve this answer

Comments

1

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
Improve this answer

Comments

-1

For this problem, you can use default_factory, which is described in the dataclass documentation.

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

Comments

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.