Python dataclass from a nested dict

Question

The standard library in 3.7 can recursively convert a dataclass into a dict (example from the docs):

from dataclasses import dataclass, asdict
from typing import List

@dataclass
class Point:
     x: int
     y: int

@dataclass
class C:
     mylist: List[Point]

p = Point(10, 20)
assert asdict(p) == {'x': 10, 'y': 20}

c = C([Point(0, 0), Point(10, 4)])
tmp = {'mylist': [{'x': 0, 'y': 0}, {'x': 10, 'y': 4}]}
assert asdict(c) == tmp

I am looking for a way to turn a dict back into a dataclass when there is nesting. Something like C(**tmp) only works if the fields of the data class are simple types and not themselves dataclasses. I am familiar with [jsonpickle][1], which however comes with a prominent security warning.

---

EDIT:

Answers have suggested the following libraries:

• dacite
• mashumaro (I used for a while, works well but I quickly ran into tricky corner cases)
• pydantic (works very well, excellent documentation and fewer corner cases) [1]: https://jsonpickle.github.io/

Answer 1

I'm the author of dacite - the tool that simplifies creation of data classes from dictionaries.

This library has only one function from_dict - this is a quick example of usage:

from dataclasses import dataclass
from dacite import from_dict

@dataclass
class User:
    name: str
    age: int
    is_active: bool

data = {
    'name': 'john',
    'age': 30,
    'is_active': True,
}

user = from_dict(data_class=User, data=data)

assert user == User(name='john', age=30, is_active=True)

Moreover dacite supports following features:

• nested structures
• (basic) types checking
• optional fields (i.e. typing.Optional)
• unions
• collections
• values casting and transformation
• remapping of fields names

... and it's well tested - 100% code coverage!

To install dacite, simply use pip (or pipenv):

$ pip install dacite

Answer 2

Below is the CPython implementation of asdict – or specifically, the internal recursive helper function _asdict_inner that it uses:

# Source: https://github.com/python/cpython/blob/master/Lib/dataclasses.py

def _asdict_inner(obj, dict_factory):
    if _is_dataclass_instance(obj):
        result = []
        for f in fields(obj):
            value = _asdict_inner(getattr(obj, f.name), dict_factory)
            result.append((f.name, value))
        return dict_factory(result)
    elif isinstance(obj, tuple) and hasattr(obj, '_fields'):
        # [large block of author comments]
        return type(obj)(*[_asdict_inner(v, dict_factory) for v in obj])
    elif isinstance(obj, (list, tuple)):
        # [ditto]
        return type(obj)(_asdict_inner(v, dict_factory) for v in obj)
    elif isinstance(obj, dict):
        return type(obj)((_asdict_inner(k, dict_factory),
                          _asdict_inner(v, dict_factory))
                         for k, v in obj.items())
    else:
        return copy.deepcopy(obj)

asdict simply calls the above with some assertions, and dict_factory=dict by default.

How can this be adapted to create an output dictionary with the required type-tagging, as mentioned in the comments?

---

1. Adding type information

My attempt involved creating a custom return wrapper inheriting from dict:

class TypeDict(dict):
    def __init__(self, t, *args, **kwargs):
        super(TypeDict, self).__init__(*args, **kwargs)

        if not isinstance(t, type):
            raise TypeError("t must be a type")

        self._type = t

    @property
    def type(self):
        return self._type

Looking at the original code, only the first clause needs to be modified to use this wrapper, as the other clauses only handle containers of dataclass-es:

# only use dict for now; easy to add back later
def _todict_inner(obj):
    if is_dataclass_instance(obj):
        result = []
        for f in fields(obj):
            value = _todict_inner(getattr(obj, f.name))
            result.append((f.name, value))
        return TypeDict(type(obj), result)

    elif isinstance(obj, tuple) and hasattr(obj, '_fields'):
        return type(obj)(*[_todict_inner(v) for v in obj])
    elif isinstance(obj, (list, tuple)):
        return type(obj)(_todict_inner(v) for v in obj)
    elif isinstance(obj, dict):
        return type(obj)((_todict_inner(k), _todict_inner(v))
                         for k, v in obj.items())
    else:
        return copy.deepcopy(obj)

Imports:

from dataclasses import dataclass, fields, is_dataclass

# thanks to Patrick Haugh
from typing import *

# deepcopy 
import copy

Functions used:

# copy of the internal function _is_dataclass_instance
def is_dataclass_instance(obj):
    return is_dataclass(obj) and not is_dataclass(obj.type)

# the adapted version of asdict
def todict(obj):
    if not is_dataclass_instance(obj):
         raise TypeError("todict() should be called on dataclass instances")
    return _todict_inner(obj)

Tests with the example dataclasses:

c = C([Point(0, 0), Point(10, 4)])

print(c)
cd = todict(c)

print(cd)
# {'mylist': [{'x': 0, 'y': 0}, {'x': 10, 'y': 4}]}

print(cd.type)
# <class '__main__.C'>

Results are as expected.

---

2. Converting back to a dataclass

The recursive routine used by asdict can be re-used for the reverse process, with some relatively minor changes:

def _fromdict_inner(obj):
    # reconstruct the dataclass using the type tag
    if is_dataclass_dict(obj):
        result = {}
        for name, data in obj.items():
            result[name] = _fromdict_inner(data)
        return obj.type(**result)

    # exactly the same as before (without the tuple clause)
    elif isinstance(obj, (list, tuple)):
        return type(obj)(_fromdict_inner(v) for v in obj)
    elif isinstance(obj, dict):
        return type(obj)((_fromdict_inner(k), _fromdict_inner(v))
                         for k, v in obj.items())
    else:
        return copy.deepcopy(obj)

Functions used:

def is_dataclass_dict(obj):
    return isinstance(obj, TypeDict)

def fromdict(obj):
    if not is_dataclass_dict(obj):
        raise TypeError("fromdict() should be called on TypeDict instances")
    return _fromdict_inner(obj)

Test:

c = C([Point(0, 0), Point(10, 4)])
cd = todict(c)
cf = fromdict(cd)

print(c)
# C(mylist=[Point(x=0, y=0), Point(x=10, y=4)])

print(cf)
# C(mylist=[Point(x=0, y=0), Point(x=10, y=4)])

Again as expected.

Answer 3

All it takes is a five-liner:

def dataclass_from_dict(klass, d):
    try:
        fieldtypes = {f.name:f.type for f in dataclasses.fields(klass)}
        return klass(**{f:dataclass_from_dict(fieldtypes[f],d[f]) for f in d})
    except:
        return d # Not a dataclass field

Sample usage:

from dataclasses import dataclass, asdict

@dataclass
class Point:
    x: float
    y: float

@dataclass
class Line:
    a: Point
    b: Point

line = Line(Point(1,2), Point(3,4))
assert line == dataclass_from_dict(Line, asdict(line))

Full code, including to/from json, here at gist: https://gist.github.com/gatopeich/1efd3e1e4269e1e98fae9983bb914f22

Answer 4

You can use mashumaro for creating dataclass object from a dict according to the scheme. Mixin from this library adds convenient from_dict and to_dict methods to dataclasses:

from dataclasses import dataclass
from typing import List
from mashumaro import DataClassDictMixin

@dataclass
class Point(DataClassDictMixin):
     x: int
     y: int

@dataclass
class C(DataClassDictMixin):
     mylist: List[Point]

p = Point(10, 20)
tmp = {'x': 10, 'y': 20}
assert p.to_dict() == tmp
assert Point.from_dict(tmp) == p

c = C([Point(0, 0), Point(10, 4)])
tmp = {'mylist': [{'x': 0, 'y': 0}, {'x': 10, 'y': 4}]}
assert c.to_dict() == tmp
assert C.from_dict(tmp) == c

Answer 5

Using no additional modules, you can make use of the __post_init__ function to automatically convert the dict values to the correct type. This function is called after __init__.

from dataclasses import dataclass, asdict


@dataclass
class Bar:
    fee: str
    far: str

@dataclass
class Foo:
    bar: Bar

    def __post_init__(self):
        if isinstance(self.bar, dict):
            self.bar = Bar(**self.bar)

foo = Foo(bar=Bar(fee="La", far="So"))

d= asdict(foo)
print(d)  # {'bar': {'fee': 'La', 'far': 'So'}}
o = Foo(**d)
print(o)  # Foo(bar=Bar(fee='La', far='So'))

This solution has the added benefit of being able to use non-dataclass objects. As long as its str function can be converted back, it's fair game. For example, it can be used to keep str fields as IP4Address internally.

Answer 6

If your goal is to produce JSON from and to existing, predefined dataclasses, then just write custom encoder and decoder hooks. Do not use dataclasses.asdict() here, instead record in JSON a (safe) reference to the original dataclass.

jsonpickle is not safe because it stores references to arbitrary Python objects and passes in data to their constructors. With such references I can get jsonpickle to reference internal Python data structures and create and execute functions, classes and modules at will. But that doesn't mean you can't handle such references unsafely. Just verify that you only import (not call) and then verify that the object is an actual dataclass type, before you use it.

The framework can be made generic enough but still limited only to JSON-serialisable types plus dataclass-based instances:

import dataclasses
import importlib
import sys

def dataclass_object_dump(ob):
    datacls = type(ob)
    if not dataclasses.is_dataclass(datacls):
        raise TypeError(f"Expected dataclass instance, got '{datacls!r}' object")
    mod = sys.modules.get(datacls.__module__)
    if mod is None or not hasattr(mod, datacls.__qualname__):
        raise ValueError(f"Can't resolve '{datacls!r}' reference")
    ref = f"{datacls.__module__}.{datacls.__qualname__}"
    fields = (f.name for f in dataclasses.fields(ob))
    return {**{f: getattr(ob, f) for f in fields}, '__dataclass__': ref}

def dataclass_object_load(d):
    ref = d.pop('__dataclass__', None)
    if ref is None:
        return d
    try:
        modname, hasdot, qualname = ref.rpartition('.')
        module = importlib.import_module(modname)
        datacls = getattr(module, qualname)
        if not dataclasses.is_dataclass(datacls) or not isinstance(datacls, type):
            raise ValueError
        return datacls(**d)
    except (ModuleNotFoundError, ValueError, AttributeError, TypeError):
        raise ValueError(f"Invalid dataclass reference {ref!r}") from None

This uses JSON-RPC-style class hints to name the dataclass, and on loading this is verified to still be a data class with the same fields. No type checking is done on the values of the fields (as that's a whole different kettle of fish).

Use these as the default and object_hook arguments to json.dump[s]() and json.dump[s]():

>>> print(json.dumps(c, default=dataclass_object_dump, indent=4))
{
    "mylist": [
        {
            "x": 0,
            "y": 0,
            "__dataclass__": "__main__.Point"
        },
        {
            "x": 10,
            "y": 4,
            "__dataclass__": "__main__.Point"
        }
    ],
    "__dataclass__": "__main__.C"
}
>>> json.loads(json.dumps(c, default=dataclass_object_dump), object_hook=dataclass_object_load)
C(mylist=[Point(x=0, y=0), Point(x=10, y=4)])
>>> json.loads(json.dumps(c, default=dataclass_object_dump), object_hook=dataclass_object_load) == c
True

or create instances of the JSONEncoder and JSONDecoder classes with those same hooks.

Instead of using fully qualifying module and class names, you could also use a separate registry to map permissible type names; check against the registry on encoding, and again on decoding to ensure you don't forget to register dataclasses as you develop.

Answer 7

A possible solution that I haven't seen mentioned yet is to use dataclasses-json. This library provides conversions of dataclass instances to/from JSON, but also to/from dict (like dacite and mashumaro, which were suggested in earlier answers).

dataclasses-json requires decorating the classes with @dataclass_json in addition to @dataclass. The decorated classes then get a couple of member functions for conversions to/from JSON and to/from dict:

• from_dict(...)
• from_json(...)
• to_dict(...)
• to_json(...)

Here is a slightly modified version of the original code in the question. I've added the required @dataclass_json decorators and asserts for the conversion from dicts to instances of Point and C:

from dataclasses import dataclass, asdict
from dataclasses_json import dataclass_json
from typing import List

@dataclass_json
@dataclass
class Point:
     x: int
     y: int

@dataclass_json
@dataclass
class C:
     mylist: List[Point]

p = Point(10, 20)

assert asdict(p) == {'x': 10, 'y': 20}
assert p == Point.from_dict({'x': 10, 'y': 20})

c = C([Point(0, 0), Point(10, 4)])

tmp = {'mylist': [{'x': 0, 'y': 0}, {'x': 10, 'y': 4}]}

assert asdict(c) == tmp
assert c == C.from_dict(tmp)

Answer 8

Simple solution that supports lists as well (and can be extended for other generic uses)

from dataclasses import dataclass, asdict, fields, is_dataclass
from typing import List
from types import GenericAlias

def asdataclass(klass, d):
    if not is_dataclass(klass):
        return d
    values = {}
    for f in fields(klass):
        if isinstance(f.type, GenericAlias) and f.type.__origin__ == list:
            values[f.name] = [asdataclass(f.type.__args__[0], d2) for d2 in d[f.name]]
        else:
            values[f.name] = asdataclass(f.type,d[f.name])
    return klass(**values)

@dataclass
class Point:
    x: int
    y: int

@dataclass
class C:
    mylist: list[Point]
    title: str = ""

c = C([Point(0, 0), Point(10, 4)])

assert c == asdataclass(C, asdict(c))

Based on https://stackoverflow.com/a/54769644/871166

Answer 9

undictify is a library which could be of help. Here is a minimal usage example:

import json
from dataclasses import dataclass
from typing import List, NamedTuple, Optional, Any

from undictify import type_checked_constructor


@type_checked_constructor(skip=True)
@dataclass
class Heart:
    weight_in_kg: float
    pulse_at_rest: int


@type_checked_constructor(skip=True)
@dataclass
class Human:
    id: int
    name: str
    nick: Optional[str]
    heart: Heart
    friend_ids: List[int]


tobias_dict = json.loads('''
    {
        "id": 1,
        "name": "Tobias",
        "heart": {
            "weight_in_kg": 0.31,
            "pulse_at_rest": 52
        },
        "friend_ids": [2, 3, 4, 5]
    }''')

tobias = Human(**tobias_dict)

Answer 10

Validobj does just that. Compared to other libraries, it provides a simpler interface (just one function at the moment) and emphasizes informative error messages. For example, given a schema like

import dataclasses
from typing import Optional, List


@dataclasses.dataclass
class User:
    name: str
    phone: Optional[str] = None
    tasks: List[str] = dataclasses.field(default_factory=list)

One gets an error like

>>> import validobj
>>> validobj.parse_input({
...      'phone': '555-1337-000', 'address': 'Somewhereville', 'nme': 'Zahari'}, User
... )
Traceback (most recent call last):
...
WrongKeysError: Cannot process value into 'User' because fields do not match.
The following required keys are missing: {'name'}. The following keys are unknown: {'nme', 'address'}.
Alternatives to invalid value 'nme' include:
  - name

All valid options are:
  - name
  - phone
  - tasks

for a typo on a given field.

Answer 11

I would like to suggest using the Composite Pattern to solve this, the main advantage is that you could continue adding classes to this pattern and have them behave the same way.

from dataclasses import dataclass
from typing import List


@dataclass
class CompositeDict:
    def as_dict(self):
        retval = dict()
        for key, value in self.__dict__.items():
            if key in self.__dataclass_fields__.keys():
                if type(value) is list:
                    retval[key] = [item.as_dict() for item in value]
                else:
                    retval[key] = value
        return retval

@dataclass
class Point(CompositeDict):
    x: int
    y: int


@dataclass
class C(CompositeDict):
    mylist: List[Point]


c = C([Point(0, 0), Point(10, 4)])
tmp = {'mylist': [{'x': 0, 'y': 0}, {'x': 10, 'y': 4}]}
assert c.as_dict() == tmp

as a side note, you could employ a factory pattern within the CompositeDict class that would handle other cases like nested dicts, tuples and such, which would save much boilerplate.

Answer 12

from validated_dc import ValidatedDC
from dataclasses import dataclass

from typing import List, Union


@dataclass
class Foo(ValidatedDC):
    foo: int


@dataclass
class Bar(ValidatedDC):
    bar: Union[Foo, List[Foo]]


foo = {'foo': 1}
instance = Bar(bar=foo)
print(instance.get_errors())  # None
print(instance)               # Bar(bar=Foo(foo=1))

list_foo = [{'foo': 1}, {'foo': 2}]
instance = Bar(bar=list_foo)
print(instance.get_errors())  # None
print(instance)               # Bar(bar=[Foo(foo=1), Foo(foo=2)])

validated_dc:
https://github.com/EvgeniyBurdin/validated_dc

And see a more detailed example:
https://github.com/EvgeniyBurdin/validated_dc/blob/master/examples/detailed.py

Answer 13

I really think that concept presented by gatopeich in this answer is the best approach for this question.

I've fixed and civilized his code. This is a correct function to load dataclass back from dictionary:

def dataclass_from_dict(cls: type, src: t.Mapping[str, t.Any]) -> t.Any:
    field_types_lookup = {
        field.name: field.type
        for field in dataclasses.fields(cls)
    }

    constructor_inputs = {}
    for field_name, value in src.items():
        try:
            constructor_inputs[field_name] = dataclass_from_dict(field_types_lookup[field_name], value)
        except TypeError as e:
            # type error from fields() call in recursive call
            # indicates that field is not a dataclass, this is how we are
            # breaking the recursion. If not a dataclass - no need for loading
            constructor_inputs[field_name] = value
        except KeyError:
            # similar, field not defined on dataclass, pass as plain field value
            constructor_inputs[field_name] = value

return cls(**constructor_inputs)

Then you can test with following:

@dataclass
class Point:
    x: float
    y: float


@dataclass
class Line:
    a: Point
    b: Point

p1, p2 = Point(1,1), Point(2,2)
line = Line(p1, p1)

assert line == dataclass_from_dict(Line, asdict(line))