How can I set an attribute in a frozen dataclass custom __init__ method?

Question

I'm trying to build a @dataclass that defines a schema but is not actually instantiated with the given members. (Basically, I'm hijacking the convenient @dataclass syntax for other purposes). This almost does what I want:

@dataclass(frozen=True, init=False)
class Tricky:
    thing1: int
    thing2: str

    def __init__(self, thing3):
        self.thing3 = thing3

But I get a FrozenInstanceError in the __init__ method:

dataclasses.FrozenInstanceError: cannot assign to field 'thing3'

I need the frozen=True (for hashability). Is there some way I can set a custom attribute in __init__ on a frozen @dataclass?

Answer 1

The problem is that the default __init__ implementation uses object.__setattr__() with frozen classes and by providing your own implementation, you have to use it too which would make your code pretty hacky:

@dataclass(frozen=True, init=False)
class Tricky:
    thing1: int
    thing2: str

    def __init__(self, thing3):
        object.__setattr__(self, "thing3", thing3)

Unfortunately, python does not provide a way to use the default implementation so we can't simply do something like:

@dataclass(frozen=True, init=False)
class Tricky:
    thing1: int
    thing2: str

    def __init__(self, thing3, **kwargs):
        self.__default_init__(DoSomething(thing3), **kwargs)

However, with we can implement that behavior quite easily:

def dataclass_with_default_init(_cls=None, *args, **kwargs):
    def wrap(cls):
        # Save the current __init__ and remove it so dataclass will
        # create the default __init__.
        user_init = getattr(cls, "__init__")
        delattr(cls, "__init__")

        # let dataclass process our class.
        result = dataclass(cls, *args, **kwargs)

        # Restore the user's __init__ save the default init to __default_init__.
        setattr(result, "__default_init__", result.__init__)
        setattr(result, "__init__", user_init)

        # Just in case that dataclass will return a new instance,
        # (currently, does not happen), restore cls's __init__.
        if result is not cls:
            setattr(cls, "__init__", user_init)

        return result

    # Support both dataclass_with_default_init() and dataclass_with_default_init
    if _cls is None:
        return wrap
    else:
        return wrap(_cls)

and then

@dataclass_with_default_init(frozen=True)
class DataClass:
    value: int

    def __init__(self, value: str):
        # error:
        # self.value = int(value)

        self.__default_init__(value=int(value))

Update: I opened this bug and I hope to implement that by 3.9.

Answer 2

I need the frozen=True (for hashability).

There is no strict need to freeze a class just to be hashable. You can opt to just not mutate the attributes from anywhere in your code, and set unsafe_hash=True instead.

However, you should really declare thing3 as a field, and not use a custom __init__:

from dataclasses import dataclass, field
from typing import Any

@dataclass(unsafe_hash=True)
class Tricky:
    thing1: int = field(init=False)
    thing2: str = field(init=False)
    thing3: Any

    def __post_init__(self):
        self.thing1 = 42
        self.thing2 = 'foo'

Here thing1 and thing2 have init=False set, so they are not passed to the __init__ method. You then set them in a __post_init__() method.

Note that this now requires that you don't freeze the class, otherwise you can't set thing1 and thing2 either, not in a custom __init__ and not in __post_init__.

Demo:

>>> Tricky('bar')
Tricky(thing1=42, thing2='foo', thing3='bar')
>>> hash(Tricky('bar'))
-3702476386127038381

If all you want is a schema definition, you don’t need dataclasses at all. You can get the class annotations from any class; either as raw annotations or with typing.get_type_hints().

Answer 3

Here's a simpler option - just add a static make function:

@dataclass(frozen=True)
class Tricky:
    thing1: str
    thing2: int
    thing3: bool

    @classmethod
    def make(cls, whatever: str, you: bool, want: float):
        return cls(whatever + "..", you * 4, want > 5)

x = Tricky.make("foo", false, 3)

Depending on what your make method does it may be a good idea to follow Rust's naming convention - from_foo(). E.g.

@dataclass(frozen=True)
class Coord:
    lat: float
    lon: float

    @classmethod
    def from_os_grid_reference(cls, x: int, y: int):
        return cls(...)

    @classmethod
    def from_gps_nema_string(cls, nema_string: str):
        return cls(...)