Validating detailed types in python dataclasses

Question:

Python 3.7 was released a while ago, and I wanted to test some of the fancy new dataclass+typing features. Getting hints to work right is easy enough, with both native types and those from the typing module:

>>> import dataclasses
>>> import typing as ty
>>> 
... @dataclasses.dataclass
... class Structure:
...     a_str: str
...     a_str_list: ty.List[str]
...
>>> my_struct = Structure(a_str='test', a_str_list=['t', 'e', 's', 't'])
>>> my_struct.a_str_list[0].  # IDE suggests all the string methods :)

But one other thing that I wanted to try was forcing the type hints as conditions during runtime, i.e. it should not be possible for a dataclass with incorrect types to exist. It can be implemented nicely with __post_init__:

>>> @dataclasses.dataclass
... class Structure:
...     a_str: str
...     a_str_list: ty.List[str]
...     
...     def validate(self):
...         ret = True
...         for field_name, field_def in self.__dataclass_fields__.items():
...             actual_type = type(getattr(self, field_name))
...             if actual_type != field_def.type:
...                 print(f"t{field_name}: '{actual_type}' instead of '{field_def.type}'")
...                 ret = False
...         return ret
...     
...     def __post_init__(self):
...         if not self.validate():
...             raise ValueError('Wrong types')

This kind of validate function works for native types and custom classes, but not those specified by the typing module:

>>> my_struct = Structure(a_str='test', a_str_list=['t', 'e', 's', 't'])
Traceback (most recent call last):
  a_str_list: '<class 'list'>' instead of 'typing.List[str]'
  ValueError: Wrong types

Is there a better approach to validate an untyped list with a typing-typed one? Preferably one that doesn’t include checking the types of all elements in any list, dict, tuple, or set that is a dataclass‘ attribute.


Revisiting this question after a couple of years, I’ve now moved to use pydantic in cases where I want to validate classes that I’d normally just define a dataclass for. I’ll leave my mark with the currently accepted answer though, since it correctly answers the original question and has outstanding educational value.

Asked By: Arne

||

Answers:

Instead of checking for type equality, you should use isinstance. But you cannot use a parametrized generic type (typing.List[int]) to do so, you must use the "generic" version (typing.List). So you will be able to check for the container type but not the contained types. Parametrized generic types define an __origin__ attribute that you can use for that.

Contrary to Python 3.6, in Python 3.7 most type hints have a useful __origin__ attribute. Compare:

# Python 3.6
>>> import typing
>>> typing.List.__origin__
>>> typing.List[int].__origin__
typing.List

and

# Python 3.7
>>> import typing
>>> typing.List.__origin__
<class 'list'>
>>> typing.List[int].__origin__
<class 'list'>

Python 3.8 introduce even better support with the typing.get_origin() introspection function:

# Python 3.8
>>> import typing
>>> typing.get_origin(typing.List)
<class 'list'>
>>> typing.get_origin(typing.List[int])
<class 'list'>

Notable exceptions being typing.Any, typing.Union and typing.ClassVar… Well, anything that is a typing._SpecialForm does not define __origin__. Fortunately:

>>> isinstance(typing.Union, typing._SpecialForm)
True
>>> isinstance(typing.Union[int, str], typing._SpecialForm)
False
>>> typing.get_origin(typing.Union[int, str])
typing.Union

But parametrized types define an __args__ attribute that store their parameters as a tuple; Python 3.8 introduce the typing.get_args() function to retrieve them:

# Python 3.7
>>> typing.Union[int, str].__args__
(<class 'int'>, <class 'str'>)

# Python 3.8
>>> typing.get_args(typing.Union[int, str])
(<class 'int'>, <class 'str'>)

So we can improve type checking a bit:

for field_name, field_def in self.__dataclass_fields__.items():
    if isinstance(field_def.type, typing._SpecialForm):
        # No check for typing.Any, typing.Union, typing.ClassVar (without parameters)
        continue
    try:
        actual_type = field_def.type.__origin__
    except AttributeError:
        # In case of non-typing types (such as <class 'int'>, for instance)
        actual_type = field_def.type
    # In Python 3.8 one would replace the try/except with
    # actual_type = typing.get_origin(field_def.type) or field_def.type
    if isinstance(actual_type, typing._SpecialForm):
        # case of typing.Union[…] or typing.ClassVar[…]
        actual_type = field_def.type.__args__

    actual_value = getattr(self, field_name)
    if not isinstance(actual_value, actual_type):
        print(f"t{field_name}: '{type(actual_value)}' instead of '{field_def.type}'")
        ret = False

This is not perfect as it won’t account for typing.ClassVar[typing.Union[int, str]] or typing.Optional[typing.List[int]] for instance, but it should get things started.


Next is the way to apply this check.

Instead of using __post_init__, I would go the decorator route: this could be used on anything with type hints, not only dataclasses:

import inspect
import typing
from contextlib import suppress
from functools import wraps


def enforce_types(callable):
    spec = inspect.getfullargspec(callable)

    def check_types(*args, **kwargs):
        parameters = dict(zip(spec.args, args))
        parameters.update(kwargs)
        for name, value in parameters.items():
            with suppress(KeyError):  # Assume un-annotated parameters can be any type
                type_hint = spec.annotations[name]
                if isinstance(type_hint, typing._SpecialForm):
                    # No check for typing.Any, typing.Union, typing.ClassVar (without parameters)
                    continue
                try:
                    actual_type = type_hint.__origin__
                except AttributeError:
                    # In case of non-typing types (such as <class 'int'>, for instance)
                    actual_type = type_hint
                # In Python 3.8 one would replace the try/except with
                # actual_type = typing.get_origin(type_hint) or type_hint
                if isinstance(actual_type, typing._SpecialForm):
                    # case of typing.Union[…] or typing.ClassVar[…]
                    actual_type = type_hint.__args__

                if not isinstance(value, actual_type):
                    raise TypeError('Unexpected type for '{}' (expected {} but found {})'.format(name, type_hint, type(value)))

    def decorate(func):
        @wraps(func)
        def wrapper(*args, **kwargs):
            check_types(*args, **kwargs)
            return func(*args, **kwargs)
        return wrapper

    if inspect.isclass(callable):
        callable.__init__ = decorate(callable.__init__)
        return callable

    return decorate(callable)

Usage being:

@enforce_types
@dataclasses.dataclass
class Point:
    x: float
    y: float

@enforce_types
def foo(bar: typing.Union[int, str]):
    pass

Appart from validating some type hints as suggested in the previous section, this approach still have some drawbacks:

  • type hints using strings (class Foo: def __init__(self: 'Foo'): pass) are not taken into account by inspect.getfullargspec: you may want to use typing.get_type_hints and inspect.signature instead;

  • a default value which is not the appropriate type is not validated:

     @enforce_type
     def foo(bar: int = None):
         pass
    
     foo()
    

    does not raise any TypeError. You may want to use inspect.Signature.bind in conjuction with inspect.BoundArguments.apply_defaults if you want to account for that (and thus forcing you to define def foo(bar: typing.Optional[int] = None));

  • variable number of arguments can’t be validated as you would have to define something like def foo(*args: typing.Sequence, **kwargs: typing.Mapping) and, as said at the beginning, we can only validate containers and not contained objects.


Update

After this answer got some popularity and a library heavily inspired by it got released, the need to lift the shortcomings mentioned above is becoming a reality. So I played a bit more with the typing module and will propose a few findings and a new approach here.

For starter, typing is doing a great job in finding when an argument is optional:

>>> def foo(a: int, b: str, c: typing.List[str] = None):
...   pass
... 
>>> typing.get_type_hints(foo)
{'a': <class 'int'>, 'b': <class 'str'>, 'c': typing.Union[typing.List[str], NoneType]}

This is pretty neat and definitely an improvement over inspect.getfullargspec, so better use that instead as it can also properly handle strings as type hints. But typing.get_type_hints will bail out for other kind of default values:

>>> def foo(a: int, b: str, c: typing.List[str] = 3):
...   pass
... 
>>> typing.get_type_hints(foo)
{'a': <class 'int'>, 'b': <class 'str'>, 'c': typing.List[str]}

So you may still need extra strict checking, even though such cases feels very fishy.

Next is the case of typing hints used as arguments for typing._SpecialForm, such as typing.Optional[typing.List[str]] or typing.Final[typing.Union[typing.Sequence, typing.Mapping]]. Since the __args__ of these typing._SpecialForms is always a tuple, it is possible to recursively find the __origin__ of the hints contained in that tuple. Combined with the above checks, we will then need to filter any typing._SpecialForm left.

Proposed improvements:

import inspect
import typing
from functools import wraps


def _find_type_origin(type_hint):
    if isinstance(type_hint, typing._SpecialForm):
        # case of typing.Any, typing.ClassVar, typing.Final, typing.Literal,
        # typing.NoReturn, typing.Optional, or typing.Union without parameters
        return

    actual_type = typing.get_origin(type_hint) or type_hint  # requires Python 3.8
    if isinstance(actual_type, typing._SpecialForm):
        # case of typing.Union[…] or typing.ClassVar[…] or …
        for origins in map(_find_type_origin, typing.get_args(type_hint)):
            yield from origins
    else:
        yield actual_type


def _check_types(parameters, hints):
    for name, value in parameters.items():
        type_hint = hints.get(name, typing.Any)
        actual_types = tuple(_find_type_origin(type_hint))
        if actual_types and not isinstance(value, actual_types):
            raise TypeError(
                    f"Expected type '{type_hint}' for argument '{name}'"
                    f" but received type '{type(value)}' instead"
            )


def enforce_types(callable):
    def decorate(func):
        hints = typing.get_type_hints(func)
        signature = inspect.signature(func)

        @wraps(func)
        def wrapper(*args, **kwargs):
            parameters = dict(zip(signature.parameters, args))
            parameters.update(kwargs)
            _check_types(parameters, hints)

            return func(*args, **kwargs)
        return wrapper

    if inspect.isclass(callable):
        callable.__init__ = decorate(callable.__init__)
        return callable

    return decorate(callable)


def enforce_strict_types(callable):
    def decorate(func):
        hints = typing.get_type_hints(func)
        signature = inspect.signature(func)

        @wraps(func)
        def wrapper(*args, **kwargs):
            bound = signature.bind(*args, **kwargs)
            bound.apply_defaults()
            parameters = dict(zip(signature.parameters, bound.args))
            parameters.update(bound.kwargs)
            _check_types(parameters, hints)

            return func(*args, **kwargs)
        return wrapper

    if inspect.isclass(callable):
        callable.__init__ = decorate(callable.__init__)
        return callable

    return decorate(callable)

Thanks to @Aran-Fey that helped me improve this answer.

Just found this question.

pydantic can do full type validation for dataclasses out of the box. (admission: I built pydantic)

Just use pydantic’s version of the decorator, the resulting dataclass is completely vanilla.

from datetime import datetime
from pydantic.dataclasses import dataclass

@dataclass
class User:
    id: int
    name: str = 'John Doe'
    signup_ts: datetime = None

print(User(id=42, signup_ts='2032-06-21T12:00'))
"""
User(id=42, name='John Doe', signup_ts=datetime.datetime(2032, 6, 21, 12, 0))
"""

User(id='not int', signup_ts='2032-06-21T12:00')

The last line will give:

    ...
pydantic.error_wrappers.ValidationError: 1 validation error
id
  value is not a valid integer (type=type_error.integer)
Answered By: SColvin

For typing aliases, you must separately check the annotation.
I did like this:
https://github.com/EvgeniyBurdin/validated_dc

Answered By: Evgeniy_Burdin

I created a tiny Python library for this purpose: https://github.com/tamuhey/dataclass_utils

This library can be applied for such dataclass that holds another dataclass (nested dataclass), and nested container type (like Tuple[List[Dict...)

Answered By: tamuhey

Adding an alternative option – convtools models (docs / github).

The vision of this library is:

  • validation first
  • no implicit type casting
  • no implicit data losses during type casting – e.g. casting 10.0 to int is fine, 10.1 is not
  • if there’s a model instance, it is valid.

And it also does it’s best to allow for automated error processing (link).

1. Validation only

from typing import List
from convtools.contrib.models import DictModel, build

class Structure(DictModel):
    a_str: str
    a_str_list: List[str]


structure, errors = build(
    Structure, {"a_str": "test", "a_str_list": ["t", "e", "s", "t"]}
)

"""
>>> In [12]: structure
>>> Out[12]: Structure(a_str='test', a_str_list=['t', 'e', 's', 't'])
>>> 
>>> In [13]: structure.to_dict()
>>> Out[13]: {'a_str': 'test', 'a_str_list': ['t', 'e', 's', 't']}

>>> In [14]: structure.a_str_list[0]. # IDE suggests all the string methods :)
>>>     capitalize()   encode()       format()       isalpha()      ...
>>>     casefold()     endswith()     format_map()   isascii()      ...
>>>     center()       expandtabs()   index()        isdecimal()    ...
>>>     count()        find()         isalnum()      isdigit()      ...
"""

# Let's fail validation test
structure, errors = build(
    Structure, {"a_str": "test", "a_str_list": ["t", "e", "s", "t", 1]}
)
"""
>>> In [16]: errors
>>> Out[16]: {'a_str_list': {4: {'__ERRORS': {'type': 'int instead of str'}}}}
"""

2. Type casting

from convtools.contrib.models import cast

# FIELD-LEVEL GRANULARITY
class Structure(DictModel):
    # no type casting here
    a_str: str

    # casting to list of strs
    #   - when run with no args it infers caster from output type
    #   - OR you can pass built-in/custom casters
    a_str_list: List[str] = cast()


# MODEL-LEVEL GRANULARITY
class Structure(DictModel):
    a_str: str
    a_str_list: List[str]

    class Meta:
        # forces all fields to be cast to expected types
        cast = True

        # # JIC: to override automatic caster inference:
        # cast_overrides = {
        #     date: casters.DateFromStr("%m/%d/%Y")
        # }


# now let's try again but with type casting
structure, errors = build(
    Structure, {"a_str": "test", "a_str_list": ["t", "e", "s", "t", 1]}
)
"""
>>> In [23]: structure
>>> Out[23]: Structure(a_str='test', a_str_list=['t', 'e', 's', 't', '1'])
"""

Answered By: westandskif