Home

Awesome

Dataclasses JSON

This library provides a simple API for encoding and decoding dataclasses to and from JSON.

It's very easy to get started.

README / Documentation website. Features a navigation bar and search functionality, and should mirror this README exactly -- take a look!

Quickstart

pip install dataclasses-json

from dataclasses import dataclass
from dataclasses_json import dataclass_json


@dataclass_json
@dataclass
class Person:
    name: str


person = Person(name='lidatong')
person.to_json()  # '{"name": "lidatong"}' <- this is a string
person.to_dict()  # {'name': 'lidatong'} <- this is a dict
Person.from_json('{"name": "lidatong"}')  # Person(1)
Person.from_dict({'name': 'lidatong'})  # Person(1)

# You can also apply _schema validation_ using an alternative API
# This can be useful for "typed" Python code

Person.from_json('{"name": 42}')  # This is ok. 42 is not a `str`, but
                                  # dataclass creation does not validate types
Person.schema().loads('{"name": 42}')  # Error! Raises `ValidationError`

What if you want to work with camelCase JSON?

# same imports as above, with the additional `LetterCase` import
from dataclasses import dataclass
from dataclasses_json import dataclass_json, LetterCase

@dataclass_json(letter_case=LetterCase.CAMEL)  # now all fields are encoded/decoded from camelCase
@dataclass
class ConfiguredSimpleExample:
    int_field: int

ConfiguredSimpleExample(1).to_json()  # {"intField": 1}
ConfiguredSimpleExample.from_json('{"intField": 1}')  # ConfiguredSimpleExample(1)

Supported types

It's recursive (see caveats below), so you can easily work with nested dataclasses. In addition to the supported types in the py to JSON table, this library supports the following:

The latest release is compatible with both Python 3.7 and Python 3.6 (with the dataclasses backport).

Usage

Approach 1: Class decorator

from dataclasses import dataclass
from dataclasses_json import dataclass_json

@dataclass_json
@dataclass
class Person:
    name: str

lidatong = Person('lidatong')

# Encoding to JSON
lidatong.to_json()  # '{"name": "lidatong"}'

# Decoding from JSON
Person.from_json('{"name": "lidatong"}')  # Person(name='lidatong')

Note that the @dataclass_json decorator must be stacked above the @dataclass decorator (order matters!)

Approach 2: Inherit from a mixin

from dataclasses import dataclass
from dataclasses_json import DataClassJsonMixin

@dataclass
class Person(DataClassJsonMixin):
    name: str

lidatong = Person('lidatong')

# A different example from Approach 1 above, but usage is the exact same
assert Person.from_json(lidatong.to_json()) == lidatong

Pick whichever approach suits your taste. Note that there is better support for the mixin approach when using static analysis tools (e.g. linting, typing), but the differences in implementation will be invisible in runtime usage.

How do I...

Use my dataclass with JSON arrays or objects?

from dataclasses import dataclass
from dataclasses_json import dataclass_json

@dataclass_json
@dataclass
class Person:
    name: str

Encode into a JSON array containing instances of my Data Class

people_json = [Person('lidatong')]
Person.schema().dumps(people_json, many=True)  # '[{"name": "lidatong"}]'

Decode a JSON array containing instances of my Data Class

people_json = '[{"name": "lidatong"}]'
Person.schema().loads(people_json, many=True)  # [Person(name='lidatong')]

Encode as part of a larger JSON object containing my Data Class (e.g. an HTTP request/response)

import json

response_dict = {
    'response': {
        'person': Person('lidatong').to_dict()
    }
}

response_json = json.dumps(response_dict)

In this case, we do two steps. First, we encode the dataclass into a python dictionary rather than a JSON string, using .to_dict.

Second, we leverage the built-in json.dumps to serialize our dataclass into a JSON string.

Decode as part of a larger JSON object containing my Data Class (e.g. an HTTP response)

import json

response_dict = json.loads('{"response": {"person": {"name": "lidatong"}}}')

person_dict = response_dict['response']

person = Person.from_dict(person_dict)

In a similar vein to encoding above, we leverage the built-in json module.

First, call json.loads to read the entire JSON object into a dictionary. We then access the key of the value containing the encoded dict of our Person that we want to decode (response_dict['response']).

Second, we load in the dictionary using Person.from_dict.

Encode or decode into Python lists/dictionaries rather than JSON?

This can be by calling .schema() and then using the corresponding encoder/decoder methods, ie. .load(...)/.dump(...).

Encode into a single Python dictionary

person = Person('lidatong')
person.to_dict()  # {'name': 'lidatong'}

Encode into a list of Python dictionaries

people = [Person('lidatong')]
Person.schema().dump(people, many=True)  # [{'name': 'lidatong'}]

Decode a dictionary into a single dataclass instance

person_dict = {'name': 'lidatong'}
Person.from_dict(person_dict)  # Person(name='lidatong')

Decode a list of dictionaries into a list of dataclass instances

people_dicts = [{"name": "lidatong"}]
Person.schema().load(people_dicts, many=True)  # [Person(name='lidatong')]

Encode or decode from camelCase (or kebab-case)?

JSON letter case by convention is camelCase, in Python members are by convention snake_case.

You can configure it to encode/decode from other casing schemes at both the class level and the field level.

from dataclasses import dataclass, field

from dataclasses_json import LetterCase, config, dataclass_json


# changing casing at the class level
@dataclass_json(letter_case=LetterCase.CAMEL)
@dataclass
class Person:
    given_name: str
    family_name: str
    
Person('Alice', 'Liddell').to_json()  # '{"givenName": "Alice"}'
Person.from_json('{"givenName": "Alice", "familyName": "Liddell"}')  # Person('Alice', 'Liddell')

# at the field level
@dataclass_json
@dataclass
class Person:
    given_name: str = field(metadata=config(letter_case=LetterCase.CAMEL))
    family_name: str
    
Person('Alice', 'Liddell').to_json()  # '{"givenName": "Alice"}'
# notice how the `family_name` field is still snake_case, because it wasn't configured above
Person.from_json('{"givenName": "Alice", "family_name": "Liddell"}')  # Person('Alice', 'Liddell')

This library assumes your field follows the Python convention of snake_case naming. If your field is not snake_case to begin with and you attempt to parameterize LetterCase, the behavior of encoding/decoding is undefined (most likely it will result in subtle bugs).

Encode or decode using a different name

from dataclasses import dataclass, field

from dataclasses_json import config, dataclass_json

@dataclass_json
@dataclass
class Person:
    given_name: str = field(metadata=config(field_name="overriddenGivenName"))

Person(given_name="Alice")  # Person('Alice')
Person.from_json('{"overriddenGivenName": "Alice"}')  # Person('Alice')
Person('Alice').to_json()  # {"overriddenGivenName": "Alice"}

Handle missing or optional field values when decoding?

By default, any fields in your dataclass that use default or default_factory will have the values filled with the provided default, if the corresponding field is missing from the JSON you're decoding.

Decode JSON with missing field

@dataclass_json
@dataclass
class Student:
    id: int
    name: str = 'student'

Student.from_json('{"id": 1}')  # Student(id=1, name='student')

Notice from_json filled the field name with the specified default 'student' when it was missing from the JSON.

Sometimes you have fields that are typed as Optional, but you don't necessarily want to assign a default. In that case, you can use the infer_missing kwarg to make from_json infer the missing field value as None.

Decode optional field without default

@dataclass_json
@dataclass
class Tutor:
    id: int
    student: Optional[Student] = None

Tutor.from_json('{"id": 1}')  # Tutor(id=1, student=None)

Personally I recommend you leverage dataclass defaults rather than using infer_missing, but if for some reason you need to decouple the behavior of JSON decoding from the field's default value, this will allow you to do so.

Handle unknown / extraneous fields in JSON?

By default, it is up to the implementation what happens when a json_dataclass receives input parameters that are not defined. (the from_dict method ignores them, when loading using schema() a ValidationError is raised.) There are three ways to customize this behavior.

Assume you want to instantiate a dataclass with the following dictionary:

dump_dict = {"endpoint": "some_api_endpoint", "data": {"foo": 1, "bar": "2"}, "undefined_field_name": [1, 2, 3]}
  1. You can enforce to always raise an error by setting the undefined keyword to Undefined.RAISE ('RAISE' as a case-insensitive string works as well). Of course it works normally if you don't pass any undefined parameters.
from dataclasses_json import Undefined

@dataclass_json(undefined=Undefined.RAISE)
@dataclass()
class ExactAPIDump:
    endpoint: str
    data: Dict[str, Any]

dump = ExactAPIDump.from_dict(dump_dict)  # raises UndefinedParameterError
  1. You can simply ignore any undefined parameters by setting the undefined keyword to Undefined.EXCLUDE ('EXCLUDE' as a case-insensitive string works as well). Note that you will not be able to retrieve them using to_dict:
from dataclasses_json import Undefined

@dataclass_json(undefined=Undefined.EXCLUDE)
@dataclass()
class DontCareAPIDump:
    endpoint: str
    data: Dict[str, Any]

dump = DontCareAPIDump.from_dict(dump_dict)  # DontCareAPIDump(endpoint='some_api_endpoint', data={'foo': 1, 'bar': '2'})
dump.to_dict()  # {"endpoint": "some_api_endpoint", "data": {"foo": 1, "bar": "2"}}
  1. You can save them in a catch-all field and do whatever needs to be done later. Simply set the undefined keyword to Undefined.INCLUDE ('INCLUDE' as a case-insensitive string works as well) and define a field of type CatchAll where all unknown values will end up. This simply represents a dictionary that can hold anything. If there are no undefined parameters, this will be an empty dictionary.
from dataclasses_json import Undefined, CatchAll

@dataclass_json(undefined=Undefined.INCLUDE)
@dataclass()
class UnknownAPIDump:
    endpoint: str
    data: Dict[str, Any]
    unknown_things: CatchAll

dump = UnknownAPIDump.from_dict(dump_dict)  # UnknownAPIDump(endpoint='some_api_endpoint', data={'foo': 1, 'bar': '2'}, unknown_things={'undefined_field_name': [1, 2, 3]})
dump.to_dict()  # {'endpoint': 'some_api_endpoint', 'data': {'foo': 1, 'bar': '2'}, 'undefined_field_name': [1, 2, 3]}

Notes:

  1. All 3 options work as well using schema().loads and schema().dumps, as long as you don't overwrite it by specifying schema(unknown=<a marshmallow value>). marshmallow uses the same 3 keywords 'include', 'exclude', 'raise'.

  2. All 3 operations work as well using __init__, e.g. UnknownAPIDump(**dump_dict) will not raise a TypeError, but write all unknown values to the field tagged as CatchAll. Classes tagged with EXCLUDE will also simply ignore unknown parameters. Note that classes tagged as RAISE still raise a TypeError, and not a UndefinedParameterError if supplied with unknown keywords.

Override the default encode / decode / marshmallow field of a specific field?

See Overriding

Handle recursive dataclasses?

Object hierarchies where fields are of the type that they are declared within require a small type hinting trick to declare the forward reference.

from typing import Optional
from dataclasses import dataclass
from dataclasses_json import dataclass_json

@dataclass_json
@dataclass
class Tree():
    value: str
    left: Optional['Tree']
    right: Optional['Tree']

Avoid using

from __future__ import annotations

as it will cause problems with the way dataclasses_json accesses the type annotations.

Use numpy or pandas types?

Data types specific to libraries commonly used in data analysis and machine learning like numpy and pandas are not supported by default, but you can easily enable them by using custom decoders and encoders. Below are two examples for numpy and pandas types.

from dataclasses import field, dataclass
from dataclasses_json import config, dataclass_json
import numpy as np
import pandas as pd

@dataclass_json
@dataclass
class DataWithNumpy:
    my_int: np.int64 = field(metadata=config(decoder=np.int64))
    my_float: np.float64 = field(metadata=config(decoder=np.float64))
    my_array: np.ndarray = field(metadata=config(decoder=np.asarray))
DataWithNumpy.from_json("{\"my_int\": 42, \"my_float\": 13.37, \"my_array\": [1,2,3]}")

@dataclass_json
@dataclass
class DataWithPandas:
    my_df: pd.DataFrame = field(metadata=config(decoder=pd.DataFrame.from_records, encoder=lambda x: x.to_dict(orient="records")))
data = DataWithPandas.from_dict({"my_df": [{"col1": 1, "col2": 2}, {"col1": 3, "col2": 4}]})
# my_df results in:
# col1  col2
# 1    2    
# 3    4
data.to_dict()
# {"my_df": [{"col1": 1, "col2": 2}, {"col1": 3, "col2": 4}]}

Marshmallow interop

Using the dataclass_json decorator or mixing in DataClassJsonMixin will provide you with an additional method .schema().

.schema() generates a schema exactly equivalent to manually creating a marshmallow schema for your dataclass. You can reference the marshmallow API docs to learn other ways you can use the schema returned by .schema().

You can pass in the exact same arguments to .schema() that you would when constructing a PersonSchema instance, e.g. .schema(many=True), and they will get passed through to the marshmallow schema.

from dataclasses import dataclass
from dataclasses_json import dataclass_json

@dataclass_json
@dataclass
class Person:
    name: str

# You don't need to do this - it's generated for you by `.schema()`!
from marshmallow import Schema, fields

class PersonSchema(Schema):
    name = fields.Str()

Briefly, on what's going on under the hood in the above examples: calling .schema() will have this library generate a marshmallow schema for you. It also fills in the corresponding object hook, so that marshmallow will create an instance of your Data Class on load (e.g. Person.schema().load returns a Person) rather than a dict, which it does by default in marshmallow.

Performance note

.schema() is not cached (it generates the schema on every call), so if you have a nested Data Class you may want to save the result to a variable to avoid re-generation of the schema on every usage.

person_schema = Person.schema()
person_schema.dump(people, many=True)

# later in the code...

person_schema.dump(person)

Overriding / Extending

Overriding

For example, you might want to encode/decode datetime objects using ISO format rather than the default timestamp.

from dataclasses import dataclass, field
from dataclasses_json import dataclass_json, config
from datetime import datetime
from marshmallow import fields

@dataclass_json
@dataclass
class DataClassWithIsoDatetime:
    created_at: datetime = field(
        metadata=config(
            encoder=datetime.isoformat,
            decoder=datetime.fromisoformat,
            mm_field=fields.DateTime(format='iso')
        )
    )

Extending

Similarly, you might want to extend dataclasses_json to encode date objects.

from dataclasses import dataclass, field
from dataclasses_json import dataclass_json, config
from datetime import date
from marshmallow import fields

dataclasses_json.cfg.global_config.encoders[date] = date.isoformat
dataclasses_json.cfg.global_config.decoders[date] = date.fromisoformat

@dataclass_json
@dataclass
class DataClassWithIsoDatetime:
    created_at: date
    modified_at: date
    accessed_at: date

As you can see, you can override or extend the default codecs by providing a "hook" via a callable:

Note that these hooks will be invoked regardless if you're using .to_json/dump/dumps and .from_json/load/loads. So apply overrides / extensions judiciously, making sure to carefully consider whether the interaction of the encode/decode/mm_field is consistent with what you expect!

What if I have other dataclass field extensions that rely on metadata

All the dataclasses_json.config does is return a mapping, namespaced under the key 'dataclasses_json'.

Say there's another module, other_dataclass_package that uses metadata. Here's how you solve your problem:

metadata = {'other_dataclass_package': 'some metadata...'}  # pre-existing metadata for another dataclass package
dataclass_json_config = config(
            encoder=datetime.isoformat,
            decoder=datetime.fromisoformat,
            mm_field=fields.DateTime(format='iso')
        )
metadata.update(dataclass_json_config)

@dataclass_json
@dataclass
class DataClassWithIsoDatetime:
    created_at: datetime = field(metadata=metadata)

You can also manually specify the dataclass_json configuration mapping.

@dataclass_json
@dataclass
class DataClassWithIsoDatetime:
    created_at: date = field(
        metadata={'dataclasses_json': {
            'encoder': date.isoformat,
            'decoder': date.fromisoformat,
            'mm_field': fields.DateTime(format='iso')
        }}
    )

A larger example

from dataclasses import dataclass
from dataclasses_json import dataclass_json

from typing import List

@dataclass_json
@dataclass(frozen=True)
class Minion:
    name: str


@dataclass_json
@dataclass(frozen=True)
class Boss:
    minions: List[Minion]

boss = Boss([Minion('evil minion'), Minion('very evil minion')])
boss_json = """
{
    "minions": [
        {
            "name": "evil minion"
        },
        {
            "name": "very evil minion"
        }
    ]
}
""".strip()

assert boss.to_json(indent=4) == boss_json
assert Boss.from_json(boss_json) == boss

Performance

Take a look at this issue

Versioning

Note this library is still pre-1.0.0 (SEMVER).

The current convention is:

Once this library is 1.0.0, it will follow standard SEMVER conventions.

Python compatibility

Any version that is not listed in the table below we do not test against, though you might still be able to install the library. For future Python versions, please open an issue and/or a pull request, adding them to the CI suite.

Python version rangeCompatible dataclasses-json version
3.7.x - 3.12.x0.5.x - 0.6.x
>= 3.13.xNo official support (yet)

Roadmap

Currently the focus is on investigating and fixing bugs in this library, working on performance, and finishing this issue.

That said, if you think there's a feature missing / something new needed in the library, please see the contributing section below.

Contributing

First of all, thank you for being interested in contributing to this library. I really appreciate you taking the time to work on this project.

Setting up your environment

This project uses Poetry for dependency and venv management. It is quite simple to get ready for your first commit: