Understanding Python Dataclasses — Part 1

Shikhar Chauhan
MindOrks
Published in
8 min readJul 3, 2018

If you’re reading this, then you are already aware of Python 3.7 and the new features that come packed with it. Personally, I am most excited about Dataclasses. I have been waiting for them to arrive for a while.

This is a two part post:
1. Dataclass features overview in this post
2. Dataclass fields overview in the next post

Introduction

Dataclasses are python classes but are suited for storing data objects. What are data objects, you ask? Here is a non-exhaustive list of features that define data objects:

  • They store data and represent a certain data type. Ex: A number. For people familiar with ORMs, a model instance is a data object. It represents a specific kind of entity. It holds attributes that define or represent the entity.
  • They can be compared to other objects of the same type. Ex: A number can be greater than, less than, or equal to another number

There are certainly more features, but this list is sufficient to help you understand the crux.

To understand Dataclasses, we shall be implementing a simple class that holds a number, and allows us to perform the above mentioned operations.
First, we shall be using normal classes, and then we shall use Dataclasses to achieve the same result.

But before we begin, a word on the usage of Dataclasses

Python 3.7 provides a decorator dataclass that is used to convert a class into a dataclass.

All you have to do is wrap the class in the decorator:

from dataclasses import dataclass@dataclass
class A:

Now, lets dive into the usage of how and what dataclass changes for us.

Initialization

Usual

class Number:  __init__(self, val):
self.val = val

>>> one = Number(1)
>>> one.val
>>> 1

With dataclass

@dataclass
class Number:
val:int

>>> one = Number(1)
>>> one.val
>>> 1

Here’s what’s changed with the dataclass decorator:

1. No need of defining __init__and then assigning values to self, d takes care of it
2. We defined the member attributes in advance in a much more readable fashion, along with type hinting. We now know instantly that val is of type int. This is definitely more readable than the usual way of defining class members.

Zen of Python: Readability counts

It is also possible to define default values:

@dataclass
class Number:
val:int = 0

Representation

Object representation is a meaningful string representation of the object that is very useful in debugging.

Default python objects representation is not very meaningful:

class Number:
def __init__(self, val = 0):
self.val = val

>>> a = Number(1)
>>> a
>>> <__main__.Number object at 0x7ff395b2ccc0>

This gives us no insight as to the utility of the object, and will result in horrible a debugging experience.

A meaningful representation could be implemented by defining a __repr__ method in the class definition.

def __repr__(self):
return self.val

Now we get a meaningful representation of the object:

>>> a = Number(1)
>>> a
>>> 1

dataclass automatically add a __repr__ function, so that we don’t have to manually implement it.

@dataclass
class Number:
val: int = 0
>>> a = Number(1)
>>> a
>>> Number(val = 1)

Data Comparison

Generally, data objects come with a need to be compared with each other.

Comparison between two objects `a` and `b` generally consists of the following operations:

  • a < b
  • a > b
  • a == b
  • a >= b
  • a <= b

In python, it is possible to define methods in classes that can do the above operations. For the sake of simplicity and to not let this post run amuck, I shall be only demonstrating implementation of == and <.

Usual

class Number:
def __init__( self, val = 0):
self.val = val

def __eq__(self, other):
return self.val == other.val

def __lt__(self, other):
return self.val < other.val

With dataclass

@dataclass(order = True)
class Number:
val: int = 0

Yup, that’s it.

We dont need to define the __eq__and __lt__ methods, because dataclass decorator automatically adds them to the class definition for us when called with order = True

Well, how does it do that?

When you use dataclass, it adds a functions __eq__ and __lt__ to the class definition. We already know that. So, how do these functions know how to check equality and do comparison?

A dataclass generated __eq__ function will compare a tuple of its attributes with a tuple of attributes of the other instance of the same class. In our case here’s what the automatically generated __eq__ function would be equivalent to:

def __eq__(self, other):
return (self.val,) == (other.val,)

Let’s look at a more elaborate example:

We shall write a dataclass Person to hold their name and age.

@dataclass(order = True)
class Person:
name: str
age:int = 0

The automatically generated __eq__ method will be equivalent of:

def __eq__(self, other):
return (self.name, self.age) == ( other.name, other.age)

Pay attention to the order of the attributes. They will always be generated in the order you defined them in the dataclass definition.

Similarly, the equivalent __le__ function would be akin to:

def __le__(self, other):
return (self.name, self.age) <= (other.name, other.age)

A need for defining a function like __le__ generally arises, when you have to sort a list of your data objects. Python’s built-in sorted function relies on comparing two objects.


>>> import random
>>> a = [Number(random.randint(1,10)) for _ in range(10)] #generate list of random numbers>>> a>>> [Number(val=2), Number(val=7), Number(val=6), Number(val=5), Number(val=10), Number(val=9), Number(val=1), Number(val=10), Number(val=1), Number(val=7)]>>> sorted_a = sorted(a) #Sort Numbers in ascending order>>> [Number(val=1), Number(val=1), Number(val=2), Number(val=5), Number(val=6), Number(val=7), Number(val=7), Number(val=9), Number(val=10), Number(val=10)]>>> reverse_sorted_a = sorted(a, reverse = True) #Sort Numbers in descending order >>> reverse_sorted_a>>> [Number(val=10), Number(val=10), Number(val=9), Number(val=7), Number(val=7), Number(val=6), Number(val=5), Number(val=2), Number(val=1), Number(val=1)]

dataclass as a callable decorator

It is not always desirable to have all the `dunder` methods defined. Your use case might only consist of storing the values and checking equality. Thus, you only need the __init__ and __eq__ methods defined. If we could tell the decorator to not generate the other methods, it would reduce some overhead and we shall have correct operations available on the data object.

Fortunately, this can be achieved by using dataclass decorator as a callable.

From the official docs, the decorator can be used as a callable with the following arguments:

@dataclass(init=True, repr=True, eq=True, order=False, unsafe_hash=False, frozen=False)
class C:
  1. init : By default an __init__ method will be generated. If passed as False, the class will not have an __init__ method.
  2. repr : __repr__ method is generated by default. If passed as False, the class will not have an __repr__ method.
  3. eq: By default the __eq__ method will be generated. If passed as False, the __eq__ method will not be added by dataclass, but will default to the object.__eq__.
  4. order : By default __gt__ , __ge__, __lt__, __le__ methods will be generated. If passed as False, they are omitted.

We shall discuss frozen in a while. The unsafe_hash argument deserves a separate post because of its complicated use cases.

Now, back to our use case, here’s what we need:

1. __init__
2. __eq__

These functions are generated by default, so what we need is to not have the other functions generated. How do we do that? Simply pass the relevant arguments as false to the generator.

@dataclass(repr = False) # order, unsafe_hash and frozen are False
class Number:
val: int = 0
>>> a = Number(1)>>> a>>> <__main__.Number object at 0x7ff395afe898>>>> b = Number(2)>>> c = Number(1)>>> a == b>>> False>>> a < b>>> Traceback (most recent call last):
File “<stdin>”, line 1, in <module>
TypeError: ‘<’ not supported between instances of ‘Number’ and ‘Number’

Frozen Instances

Frozen Instances are objects whose attributes cannot be modified after the object has been initialized.

It is not possible to create truly immutable Python objects

To create immutable attributes on an object in Python is an arduous task, and something that I won’t dive into in this post.

Here’s what we expect from an immutable object:

>>> a = Number(10) #Assuming Number class is immutable>>> a.val = 10 # Raises Error

With Dataclasses it is possible to define a frozen object by using dataclass decorator as a callable with argument frozen=True .

When a frozen dataclass object is instantiated, any attempt to modify the attributes of the object raises FrozenInstanceError.

@dataclass(frozen = True)
class Number:
val: int = 0
>>> a = Number(1)>>> a.val>>> 1>>> a.val = 2>>> Traceback (most recent call last):
File “<stdin>”, line 1, in <module>
File “<string>”, line 3, in __setattr__
dataclasses.FrozenInstanceError: cannot assign to field ‘val’

So a frozen instance is a great way of storing

  • constants
  • settings

These generally do not change over the lifetime of the application and any attempt to modify them should generally be warded off.

Post init processing

With Dataclasses the requirement of defining an __init__ method to assign variables to self has been taken care of. But now we lose the flexibility of making function-calls/processing that might be required immediately after the variables have been assigned.

Let us discuss a use case where we define a class `Float` to contain float numbers, and we calculate the integer and decimal parts immediately after initialization.

Usual

import mathclass Float:
def __init__(self, val = 0):
self.val = val
self.process()

def process(self):
self.decimal, self.integer = math.modf(self.val)

>>> a = Float( 2.2)
>>> a.decimal>>> 0.2000>>> a.integer>>> 2.0

Fortunately, post initialization processing is already taken care of with __post_init__ method.

The generated __init__ method calls the __post_init__ method before returning. So, any processing can be made in this functions.

import math@dataclass
class FloatNumber:
val: float = 0.0

def __post_init__(self):
self.decimal, self.integer = math.modf(self.val)

>>> a = Number(2.2)
>>> a.val>>> 2.2>>> a.integer>>> 2.0>>> a.decimal>>> 0.2

Neat!

Inheritance

Dataclasses support inheritance like normal python classes.

So, the attributes defined in the parent class will be available in the child class.

@dataclass
class Person:
age: int = 0
name: str
@dataclass
class Student(Person):
grade: int
>>> s = Student(20, "John Doe", 12)>>> s.age>>> 20>>> s.name>>> "John Doe">>> s.grade>>> 12

Pay attention to the fact that the arguments to Student are in the order of fields defined in the class definition.

What about the behavior of __post_init__ during inheritance?

Since __post_init__ is just another function, it has to be invoked in the conventional form:

@dataclass
class A:
a: int

def __post_init__(self):
print("A")
@dataclass
class B(A):
b: int

def __post_init__(self):
print("B")
>>> a = B(1,2)>>> B

In the above example, only B's __post_init__ is called. How do we invoke A's __post_init__ ?

Since it is a function of the parent class, it can be invoked using super.

@dataclass
class B(A):
b: int

def __post_init__(self):
super().__post_init__() #Call post init of A
print("B")
>>> a = B(1,2)>>> A
B

Conclusion

So, above are a few ways in which Dataclasses make life easier for Python developers.
I have tried to be thorough and cover most of the use cases, yet, no man is perfect. Reach out if you find mistakes, or want me to pay attention to relevant use cases.

I shall cover dataclasses.field and unsafe_hash in different posts.

Follow me on Github, Twitter.

Update: Post for dataclasses.field can be found here.

--

--

Shikhar Chauhan
MindOrks

Machine Learning Engineer. Open Source Contributor. Mentor. Trekker.