Models

The primary means of defining objects in pydantic is via models (models are simply classes which inherit from BaseModel).

You can think of models as similar to types in strictly typed languages, or as the requirements of a single endpoint in an API.

Untrusted data can be passed to a model, and after parsing and validation pydantic guarantees that the fields of the resultant model instance will conform to the field types defined on the model.

Note

pydantic is primarily a parsing library, not a validation library. Validation is a means to an end: building a model which conforms to the types and constraints provided.

In other words, pydantic guarantees the types and constraints of the output model, not the input data.

This might sound like an esoteric distinction, but it is not. If you're unsure what this means or how it might effect your usage you should read the section about Data Conversion below.

Basic model usage🔗

from pydantic import BaseModel

class User(BaseModel):
    id: int
    name = 'Jane Doe'

User here is a model with two fields id which is an integer and is required, and name which is a string and is not required (it has a default value). The type of name is inferred from the default value, and so a type annotation is not required (however note this warning about field order when some fields do not have type annotations).

user = User(id='123')

user here is an instance of User. Initialisation of the object will perform all parsing and validation, if no ValidationError is raised, you know the resulting model instance is valid.

assert user.id == 123

fields of a model can be accessed as normal attributes of the user object the string '123' has been cast to an int as per the field type

assert user.name == 'Jane Doe'

name wasn't set when user was initialised, so it has the default value

assert user.__fields_set__ == {'id'}

the fields which were supplied when user was initialised:

assert user.dict() == dict(user) == {'id': 123, 'name': 'Jane Doe'}

either .dict() or dict(user) will provide a dict of fields, but .dict() can take numerous other arguments.

user.id = 321
assert user.id == 321

This model is mutable so field values can be changed.

Model properties🔗

The example above only shows the tip of the iceberg of what models can do. Models possess the following methods and attributes:

dict()
returns a dictionary of the model's fields and values; cf. exporting models
json()
returns a JSON string representation dict(); cf. exporting models
copy()
returns a deep copy of the model; cf. exporting models
parse_obj()
a utility for loading any object into a model with error handling if the object is not a dictionary; cf. helper functions
parse_raw()
a utility for loading strings of numerous formats; cf. helper functions
parse_file()
like parse_raw() but for files; cf. helper function
from_orm()
loads data into a model from an arbitrary class; cf. ORM mode
schema()
returns a dictionary representing the model as JSON Schema; cf. Schema
schema_json()
returns a JSON string representation of schema(); cf. Schema
construct()
a class method for creating models without running validation; cf. Creating models without validation
__fields_set__
Set of names of fields which were set when the model instance was initialised
__fields__
a dictionary of the model's fields
__config__
the configuration class for the model, cf. model config

Recursive Models🔗

More complex hierarchical data structures can be defined using models themselves as types in annotations.

from typing import List
from pydantic import BaseModel

class Foo(BaseModel):
    count: int
    size: float = None

class Bar(BaseModel):
    apple = 'x'
    banana = 'y'

class Spam(BaseModel):
    foo: Foo
    bars: List[Bar]

m = Spam(foo={'count': 4}, bars=[{'apple': 'x1'}, {'apple': 'x2'}])
print(m)
#> foo=Foo(count=4, size=None) bars=[Bar(apple='x1', banana='y'),
#> Bar(apple='x2', banana='y')]
print(m.dict())
"""
{
    'foo': {'count': 4, 'size': None},
    'bars': [
        {'apple': 'x1', 'banana': 'y'},
        {'apple': 'x2', 'banana': 'y'},
    ],
}
"""

(This script is complete, it should run "as is")

For self-referencing models, see postponed annotations.

ORM Mode (aka Arbitrary Class Instances)🔗

Pydantic models can be created from arbitrary class instances to support models that map to ORM objects.

To do this:

  1. The Config property orm_mode must be set to True.
  2. The special constructor from_orm must be used to create the model instance.

The example here uses SQLAlchemy, but the same approach should work for any ORM.

from typing import List
from sqlalchemy import Column, Integer, String
from sqlalchemy.dialects.postgresql import ARRAY
from sqlalchemy.ext.declarative import declarative_base
from pydantic import BaseModel, constr

Base = declarative_base()

class CompanyOrm(Base):
    __tablename__ = 'companies'
    id = Column(Integer, primary_key=True, nullable=False)
    public_key = Column(String(20), index=True, nullable=False, unique=True)
    name = Column(String(63), unique=True)
    domains = Column(ARRAY(String(255)))

class CompanyModel(BaseModel):
    id: int
    public_key: constr(max_length=20)
    name: constr(max_length=63)
    domains: List[constr(max_length=255)]

    class Config:
        orm_mode = True

co_orm = CompanyOrm(
    id=123,
    public_key='foobar',
    name='Testing',
    domains=['example.com', 'foobar.com']
)
print(co_orm)
#> <models_orm_mode.CompanyOrm object at 0x7f3843236ad0>
co_model = CompanyModel.from_orm(co_orm)
print(co_model)
#> id=123 public_key='foobar' name='Testing' domains=['example.com',
#> 'foobar.com']

(This script is complete, it should run "as is")

ORM instances will be parsed with from_orm recursively as well as at the top level.

Here a vanilla class is used to demonstrate the principle, but any ORM class could be used instead.

from typing import List
from pydantic import BaseModel

class PetCls:
    def __init__(self, *, name: str, species: str):
        self.name = name
        self.species = species

class PersonCls:
    def __init__(self, *, name: str, age: float = None, pets: List[PetCls]):
        self.name = name
        self.age = age
        self.pets = pets

class Pet(BaseModel):
    name: str
    species: str

    class Config:
        orm_mode = True

class Person(BaseModel):
    name: str
    age: float = None
    pets: List[Pet]

    class Config:
        orm_mode = True

bones = PetCls(name='Bones', species='dog')
orion = PetCls(name='Orion', species='cat')
anna = PersonCls(name='Anna', age=20, pets=[bones, orion])
anna_model = Person.from_orm(anna)
print(anna_model)
#> name='Anna' age=20.0 pets=[Pet(name='Bones', species='dog'),
#> Pet(name='Orion', species='cat')]

(This script is complete, it should run "as is")

Arbitrary classes are processed by pydantic using the GetterDict class (see utils.py), which attempts to provide a dictionary-like interface to any class. You can customise how this works by setting your own sub-class of GetterDict as the value of Config.getter_dict (see config).

You can also customise class validation using root_validators with pre=True. In this case your validator function will be passed a GetterDict instance which you may copy and modify.

Error Handling🔗

pydantic will raise ValidationError whenever it finds an error in the data it's validating.

Note

Validation code should not raise ValidationError itself, but rather raise ValueError, TypeError or AssertionError (or subclasses of ValueError or TypeError) which will be caught and used to populate ValidationError.

One exception will be raised regardless of the number of errors found, that ValidationError will contain information about all the errors and how they happened.

You can access these errors in a several ways:

e.errors()
method will return list of errors found in the input data.
e.json()
method will return a JSON representation of errors.
str(e)
method will return a human readable representation of the errors.

Each error object contains:

loc
the error's location as a list. The first item in the list will be the field where the error occurred, and if the field is a sub-model, subsequent items will be present to indicate the nested location of the error.
type
a computer-readable identifier of the error type.
msg
a human readable explanation of the error.
ctx
an optional object which contains values required to render the error message.

As a demonstration:

from typing import List
from pydantic import BaseModel, ValidationError, conint

class Location(BaseModel):
    lat = 0.1
    lng = 10.1

class Model(BaseModel):
    is_required: float
    gt_int: conint(gt=42)
    list_of_ints: List[int] = None
    a_float: float = None
    recursive_model: Location = None

data = dict(
    list_of_ints=['1', 2, 'bad'],
    a_float='not a float',
    recursive_model={'lat': 4.2, 'lng': 'New York'},
    gt_int=21,
)

try:
    Model(**data)
except ValidationError as e:
    print(e)
"""
5 validation errors for Model
is_required
  field required (type=value_error.missing)
gt_int
  ensure this value is greater than 42 (type=value_error.number.not_gt;
limit_value=42)
list_of_ints -> 2
  value is not a valid integer (type=type_error.integer)
a_float
  value is not a valid float (type=type_error.float)
recursive_model -> lng
  value is not a valid float (type=type_error.float)
"""

try:
    Model(**data)
except ValidationError as e:
    print(e.json())
"""
[
  {
    "loc": [
      "is_required"
    ],
    "msg": "field required",
    "type": "value_error.missing"
  },
  {
    "loc": [
      "gt_int"
    ],
    "msg": "ensure this value is greater than 42",
    "type": "value_error.number.not_gt",
    "ctx": {
      "limit_value": 42
    }
  },
  {
    "loc": [
      "list_of_ints",
      2
    ],
    "msg": "value is not a valid integer",
    "type": "type_error.integer"
  },
  {
    "loc": [
      "a_float"
    ],
    "msg": "value is not a valid float",
    "type": "type_error.float"
  },
  {
    "loc": [
      "recursive_model",
      "lng"
    ],
    "msg": "value is not a valid float",
    "type": "type_error.float"
  }
]
"""

(This script is complete, it should run "as is". json() has indent=2 set by default, but I've tweaked the JSON here and below to make it slightly more concise.)

Custom Errors🔗

In your custom data types or validators you should use ValueError, TypeError or AssertionError to raise errors.

See validators for more details on use of the @validator decorator.

from pydantic import BaseModel, ValidationError, validator

class Model(BaseModel):
    foo: str

    @validator('foo')
    def name_must_contain_space(cls, v):
        if v != 'bar':
            raise ValueError('value must be "bar"')

        return v

try:
    Model(foo='ber')
except ValidationError as e:
    print(e.errors())
"""
[
    {
        'loc': ('foo',),
        'msg': 'value must be "bar"',
        'type': 'value_error',
    },
]
"""

(This script is complete, it should run "as is")

You can also define your own error classes, which can specify a custom error code, message template, and context:

from pydantic import BaseModel, PydanticValueError, ValidationError, validator

class NotABarError(PydanticValueError):
    code = 'not_a_bar'
    msg_template = 'value is not "bar", got "{wrong_value}"'

class Model(BaseModel):
    foo: str

    @validator('foo')
    def name_must_contain_space(cls, v):
        if v != 'bar':
            raise NotABarError(wrong_value=v)
        return v

try:
    Model(foo='ber')
except ValidationError as e:
    print(e.json())
"""
[
  {
    "loc": [
      "foo"
    ],
    "msg": "value is not \"bar\", got \"ber\"",
    "type": "value_error.not_a_bar",
    "ctx": {
      "wrong_value": "ber"
    }
  }
]
"""

(This script is complete, it should run "as is")

Helper Functions🔗

Pydantic provides three classmethod helper functions on models for parsing data:

  • parse_obj: this is very similar to the __init__ method of the model, except it takes a dict rather than keyword arguments. If the object passed is not a dict a ValidationError will be raised.
  • parse_raw: this takes a str or bytes and parses it as json, then passes the result to parse_obj. Parsing pickle data is also supported by setting the content_type argument appropriately.
  • parse_file: this reads a file and passes the contents to parse_raw. If content_type is omitted, it is inferred from the file's extension.
import pickle
from datetime import datetime
from pydantic import BaseModel, ValidationError

class User(BaseModel):
    id: int
    name = 'John Doe'
    signup_ts: datetime = None

m = User.parse_obj({'id': 123, 'name': 'James'})
print(m)
#> id=123 signup_ts=None name='James'

try:
    User.parse_obj(['not', 'a', 'dict'])
except ValidationError as e:
    print(e)
"""
1 validation error for User
__root__
  User expected dict not list (type=type_error)
"""

# assumes json as no content type passed
m = User.parse_raw('{"id": 123, "name": "James"}')
print(m)
#> id=123 signup_ts=None name='James'

pickle_data = pickle.dumps({
    'id': 123,
    'name': 'James',
    'signup_ts': datetime(2017, 7, 14)
})
m = User.parse_raw(pickle_data, content_type='application/pickle',
                   allow_pickle=True)
print(m)
#> id=123 signup_ts=datetime.datetime(2017, 7, 14, 0, 0) name='James'

(This script is complete, it should run "as is")

Warning

To quote the official pickle docs, "The pickle module is not secure against erroneous or maliciously constructed data. Never unpickle data received from an untrusted or unauthenticated source."

Info

Because it can result in arbitrary code execution, as a security measure, you need to explicitly pass allow_pickle to the parsing function in order to load pickle data.

Creating models without validation🔗

pydantic also provides the construct() method which allows models to be created without validation this can be useful when data has already been validated or comes from a trusted source and you want to create a model as efficiently as possible (construct() is generally around 30x faster than creating a model with full validation).

Warning

construct() does not do any validation, meaning it can create models which are invalid. You should only ever use the construct() method with data which has already been validated, or you trust.

from pydantic import BaseModel

class User(BaseModel):
    id: int
    age: int
    name: str = 'John Doe'

original_user = User(id=123, age=32)

user_data = original_user.dict()
print(user_data)
#> {'id': 123, 'age': 32, 'name': 'John Doe'}
fields_set = original_user.__fields_set__
print(fields_set)
#> {'id', 'age'}

# ...
# pass user_data and fields_set to RPC or save to the database etc.
# ...

# you can then create a new instance of User without
# re-running validation which would be unnecessary at this point:
new_user = User.construct(_fields_set=fields_set, **user_data)
print(repr(new_user))
#> User(name='John Doe', id=123, age=32)
print(new_user.__fields_set__)
#> {'id', 'age'}

# construct can be dangerous, only use it with validated data!:
bad_user = User.construct(id='dog')
print(repr(bad_user))
#> User(name='John Doe', id='dog')

(This script is complete, it should run "as is")

The _fields_set keyword argument to construct() is optional, but allows you to be more precise about which fields were originally set and which weren't. If it's omitted __fields_set__ will just be the keys of the data provided.

For example, in the example above, if _fields_set was not provided, new_user.__fields_set__ would be {'id', 'age', 'name'}.

Generic Models🔗

Note

New in version v0.29.

This feature requires Python 3.7+.

Pydantic supports the creation of generic models to make it easier to reuse a common model structure.

In order to declare a generic model, you perform the following steps:

  • Declare one or more typing.TypeVar instances to use to parameterize your model.
  • Declare a pydantic model that inherits from pydantic.generics.GenericModel and typing.Generic, where you pass the TypeVar instances as parameters to typing.Generic.
  • Use the TypeVar instances as annotations where you will want to replace them with other types or pydantic models.

Here is an example using GenericModel to create an easily-reused HTTP response payload wrapper:

from typing import Generic, TypeVar, Optional, List

from pydantic import BaseModel, validator, ValidationError
from pydantic.generics import GenericModel

DataT = TypeVar('DataT')

class Error(BaseModel):
    code: int
    message: str

class DataModel(BaseModel):
    numbers: List[int]
    people: List[str]

class Response(GenericModel, Generic[DataT]):
    data: Optional[DataT]
    error: Optional[Error]

    @validator('error', always=True)
    def check_consistency(cls, v, values):
        if v is not None and values['data'] is not None:
            raise ValueError('must not provide both data and error')
        if v is None and values.get('data') is None:
            raise ValueError('must provide data or error')
        return v

data = DataModel(numbers=[1, 2, 3], people=[])
error = Error(code=404, message='Not found')

print(Response[int](data=1))
#> data=1 error=None
print(Response[str](data='value'))
#> data='value' error=None
print(Response[str](data='value').dict())
#> {'data': 'value', 'error': None}
print(Response[DataModel](data=data).dict())
"""
{
    'data': {'numbers': [1, 2, 3], 'people': []},
    'error': None,
}
"""
print(Response[DataModel](error=error).dict())
"""
{
    'data': None,
    'error': {'code': 404, 'message': 'Not found'},
}
"""
try:
    Response[int](data='value')
except ValidationError as e:
    print(e)
"""
2 validation errors for Response[int]
data
  value is not a valid integer (type=type_error.integer)
error
  must provide data or error (type=value_error)
"""

(This script is complete, it should run "as is")

If you set Config or make use of validator in your generic model definition, it is applied to concrete subclasses in the same way as when inheriting from BaseModel. Any methods defined on your generic class will also be inherited.

Pydantic's generics also integrate properly with mypy, so you get all the type checking you would expect mypy to provide if you were to declare the type without using GenericModel.

Note

Internally, pydantic uses create_model to generate a (cached) concrete BaseModel at runtime, so there is essentially zero overhead introduced by making use of GenericModel.

If the name of the concrete subclasses is important, you can also override the default behavior:

from typing import Generic, TypeVar, Type, Any, Tuple

from pydantic.generics import GenericModel

DataT = TypeVar('DataT')

class Response(GenericModel, Generic[DataT]):
    data: DataT

    @classmethod
    def __concrete_name__(cls: Type[Any], params: Tuple[Type[Any], ...]) -> str:
        return f'{params[0].__name__.title()}Response'

print(Response[int](data=1))
#> data=1
print(Response[str](data='a'))
#> data='a'

(This script is complete, it should run "as is")

Dynamic model creation🔗

There are some occasions where the shape of a model is not known until runtime. For this pydantic provides the create_model method to allow models to be created on the fly.

from pydantic import BaseModel, create_model

DynamicFoobarModel = create_model('DynamicFoobarModel', foo=(str, ...), bar=123)

class StaticFoobarModel(BaseModel):
    foo: str
    bar: int = 123

Here StaticFoobarModel and DynamicFoobarModel are identical.

Fields are defined by either a tuple of the form (<type>, <default value>) or just a default value. The special key word arguments __config__ and __base__ can be used to customise the new model. This includes extending a base model with extra fields.

from pydantic import BaseModel, create_model

class FooModel(BaseModel):
    foo: str
    bar: int = 123

BarModel = create_model(
    'BarModel',
    apple='russet',
    banana='yellow',
    __base__=FooModel,
)
print(BarModel)
#> <class 'BarModel'>
print(BarModel.__fields__.keys())
#> dict_keys(['foo', 'bar', 'apple', 'banana'])

Custom Root Types🔗

Pydantic models which do not represent a dict ("object" in JSON parlance) can have a custom root type defined via the __root__ field. The root type can be of any type: list, float, int, etc.

The root type is defined via the type hint on the __root__ field. The root value can be passed to model __init__ via the __root__ keyword argument or as the first and only argument to parse_obj.

from typing import List
import json
from pydantic import BaseModel
from pydantic.schema import schema

class Pets(BaseModel):
    __root__: List[str]

print(Pets(__root__=['dog', 'cat']))
#> __root__=['dog', 'cat']
print(Pets(__root__=['dog', 'cat']).json())
#> ["dog", "cat"]
print(Pets.parse_obj(['dog', 'cat']))
#> __root__=['dog', 'cat']
print(Pets.schema())
"""
{
    'title': 'Pets',
    'type': 'array',
    'items': {'type': 'string'},
}
"""
pets_schema = schema([Pets])
print(json.dumps(pets_schema, indent=2))
"""
{
  "definitions": {
    "Pets": {
      "title": "Pets",
      "type": "array",
      "items": {
        "type": "string"
      }
    }
  }
}
"""

Faux Immutability🔗

Models can be configured to be immutable via allow_mutation = False. When this is set, attempting to change the values of instance attributes will raise errors. See model config for more details on Config.

Warning

Immutability in python is never strict. If developers are determined/stupid they can always modify a so-called "immutable" object.

from pydantic import BaseModel

class FooBarModel(BaseModel):
    a: str
    b: dict

    class Config:
        allow_mutation = False

foobar = FooBarModel(a='hello', b={'apple': 'pear'})

try:
    foobar.a = 'different'
except TypeError as e:
    print(e)
#> "FooBarModel" is immutable and does not support item assignment

print(foobar.a)
#> hello
print(foobar.b)
#> {'apple': 'pear'}
foobar.b['apple'] = 'grape'
print(foobar.b)
#> {'apple': 'grape'}

Trying to change a caused an error, and a remains unchanged. However, the dict b is mutable, and the immutability of foobar doesn't stop b from being changed.

Abstract Base Classes🔗

Pydantic models can be used alongside Python's Abstract Base Classes (ABCs).

import abc
from pydantic import BaseModel

class FooBarModel(BaseModel, abc.ABC):
    a: str
    b: int

    @abc.abstractmethod
    def my_abstract_method(self):
        pass

(This script is complete, it should run "as is")

Field Ordering🔗

Field order is important in models for the following reasons:

As of v1.0 all fields with annotations (whether annotation-only or with a default value) will precede all fields without an annotation. Within their respective groups, fields remain in the order they were defined.

from pydantic import BaseModel, ValidationError

class Model(BaseModel):
    a: int
    b = 2
    c: int = 1
    d = 0
    e: float

print(Model.__fields__.keys())
#> dict_keys(['a', 'c', 'e', 'b', 'd'])
m = Model(e=2, a=1)
print(m.dict())
#> {'a': 1, 'c': 1, 'e': 2.0, 'b': 2, 'd': 0}
try:
    Model(a='x', b='x', c='x', d='x', e='x')
except ValidationError as e:
    error_locations = [e['loc'] for e in e.errors()]

print(error_locations)
#> [('a',), ('c',), ('e',), ('b',), ('d',)]

(This script is complete, it should run "as is")

Warning

As demonstrated by the example above, combining the use of annotated and non-annotated fields in the same model can result in surprising field orderings. (This is due to limitations of python)

Therefore, we recommend adding type annotations to all fields, even when a default value would determine the type by itself to guarentee field order is preserved.

Required fields🔗

To declare a field as required, you may declare it using just an annotation, or you may use an ellipsis (...) as the value:

from pydantic import BaseModel, Field

class Model(BaseModel):
    a: int
    b: int = ...
    c: int = Field(...)

(This script is complete, it should run "as is")

Where Field refers to the field function.

Here a, b and c are all required. However, use of the ellipses in b will not work well with mypy, and as of v1.0 should be avoided in most cases.

Data Conversion🔗

pydantic may cast input data to force it to conform to model field types, and in some cases this may result in a loss of information. For example:

from pydantic import BaseModel

class Model(BaseModel):
    a: int
    b: float
    c: str

print(Model(a=3.1415, b=' 2.72 ', c=123).dict())
#> {'a': 3, 'b': 2.72, 'c': '123'}

(This script is complete, it should run "as is")

This is a deliberate decision of pydantic, and in general it's the most useful approach. See here for a longer discussion on the subject.