I know that Python does not support method overloading, but I've run into a problem that I can't seem to solve in a nice Pythonic way.
I am making a game where a character needs to shoot a variety of bullets, but how do I write different functions for creating these bullets? For example suppose I have a function that creates a bullet travelling from point A to B with a given speed. I would write a function like this:
def add_bullet(sprite, start, headto, speed):
# Code ...
But I want to write other functions for creating bullets like:
def add_bullet(sprite, start, direction, speed):
def add_bullet(sprite, start, headto, spead, acceleration):
def add_bullet(sprite, script): # For bullets that are controlled by a script
def add_bullet(sprite, curve, speed): # for bullets with curved paths
# And so on ...
And so on with many variations. Is there a better way to do it without using so many keyword arguments cause its getting kinda ugly fast. Renaming each function is pretty bad too because you get either add_bullet1, add_bullet2, or add_bullet_with_really_long_name.
To address some answers:
No I can't create a Bullet class hierarchy because thats too slow. The actual code for managing bullets is in C and my functions are wrappers around C API.
I know about the keyword arguments but checking for all sorts of combinations of parameters is getting annoying, but default arguments help allot like acceleration=0
What you are asking for is called multiple dispatch. See Julia language examples which demonstrates different types of dispatches.
However, before looking at that, we'll first tackle why overloading is not really what you want in Python.
Why Not Overloading?
First, one needs to understand the concept of overloading and why it's not applicable to Python.
When working with languages that can discriminate data types at
compile-time, selecting among the alternatives can occur at
compile-time. The act of creating such alternative functions for
compile-time selection is usually referred to as overloading a
function. (Wikipedia)
Python is a dynamically typed language, so the concept of overloading simply does not apply to it. However, all is not lost, since we can create such alternative functions at run-time:
In programming languages that defer data type identification until
run-time the selection among alternative
functions must occur at run-time, based on the dynamically determined
types of function arguments. Functions whose alternative
implementations are selected in this manner are referred to most
generally as multimethods. (Wikipedia)
So we should be able to do multimethods in Python—or, as it is alternatively called: multiple dispatch.
Multiple dispatch
The multimethods are also called multiple dispatch:
Multiple dispatch or multimethods is the feature of some
object-oriented programming languages in which a function or method
can be dynamically dispatched based on the run time (dynamic) type of
more than one of its arguments. (Wikipedia)
Python does not support this out of the box1, but, as it happens, there is an excellent Python package called multipledispatch that does exactly that.
Solution
Here is how we might use multipledispatch2 package to implement your methods:
>>> from multipledispatch import dispatch
>>> from collections import namedtuple
>>> from types import * # we can test for lambda type, e.g.:
>>> type(lambda a: 1) == LambdaType
True
>>> Sprite = namedtuple('Sprite', ['name'])
>>> Point = namedtuple('Point', ['x', 'y'])
>>> Curve = namedtuple('Curve', ['x', 'y', 'z'])
>>> Vector = namedtuple('Vector', ['x','y','z'])
>>> #dispatch(Sprite, Point, Vector, int)
... def add_bullet(sprite, start, direction, speed):
... print("Called Version 1")
...
>>> #dispatch(Sprite, Point, Point, int, float)
... def add_bullet(sprite, start, headto, speed, acceleration):
... print("Called version 2")
...
>>> #dispatch(Sprite, LambdaType)
... def add_bullet(sprite, script):
... print("Called version 3")
...
>>> #dispatch(Sprite, Curve, int)
... def add_bullet(sprite, curve, speed):
... print("Called version 4")
...
>>> sprite = Sprite('Turtle')
>>> start = Point(1,2)
>>> direction = Vector(1,1,1)
>>> speed = 100 #km/h
>>> acceleration = 5.0 #m/s**2
>>> script = lambda sprite: sprite.x * 2
>>> curve = Curve(3, 1, 4)
>>> headto = Point(100, 100) # somewhere far away
>>> add_bullet(sprite, start, direction, speed)
Called Version 1
>>> add_bullet(sprite, start, headto, speed, acceleration)
Called version 2
>>> add_bullet(sprite, script)
Called version 3
>>> add_bullet(sprite, curve, speed)
Called version 4
1. Python 3 currently supports single dispatch
2. Take care not to use multipledispatch in a multi-threaded environment or you will get weird behavior.
Python does support "method overloading" as you present it. In fact, what you just describe is trivial to implement in Python, in so many different ways, but I would go with:
class Character(object):
# your character __init__ and other methods go here
def add_bullet(self, sprite=default, start=default,
direction=default, speed=default, accel=default,
curve=default):
# do stuff with your arguments
In the above code, default is a plausible default value for those arguments, or None. You can then call the method with only the arguments you are interested in, and Python will use the default values.
You could also do something like this:
class Character(object):
# your character __init__ and other methods go here
def add_bullet(self, **kwargs):
# here you can unpack kwargs as (key, values) and
# do stuff with them, and use some global dictionary
# to provide default values and ensure that ``key``
# is a valid argument...
# do stuff with your arguments
Another alternative is to directly hook the desired function directly to the class or instance:
def some_implementation(self, arg1, arg2, arg3):
# implementation
my_class.add_bullet = some_implementation_of_add_bullet
Yet another way is to use an abstract factory pattern:
class Character(object):
def __init__(self, bfactory, *args, **kwargs):
self.bfactory = bfactory
def add_bullet(self):
sprite = self.bfactory.sprite()
speed = self.bfactory.speed()
# do stuff with your sprite and speed
class pretty_and_fast_factory(object):
def sprite(self):
return pretty_sprite
def speed(self):
return 10000000000.0
my_character = Character(pretty_and_fast_factory(), a1, a2, kw1=v1, kw2=v2)
my_character.add_bullet() # uses pretty_and_fast_factory
# now, if you have another factory called "ugly_and_slow_factory"
# you can change it at runtime in python by issuing
my_character.bfactory = ugly_and_slow_factory()
# In the last example you can see abstract factory and "method
# overloading" (as you call it) in action
You can use "roll-your-own" solution for function overloading. This one is copied from Guido van Rossum's article about multimethods (because there is little difference between multimethods and overloading in Python):
registry = {}
class MultiMethod(object):
def __init__(self, name):
self.name = name
self.typemap = {}
def __call__(self, *args):
types = tuple(arg.__class__ for arg in args) # a generator expression!
function = self.typemap.get(types)
if function is None:
raise TypeError("no match")
return function(*args)
def register(self, types, function):
if types in self.typemap:
raise TypeError("duplicate registration")
self.typemap[types] = function
def multimethod(*types):
def register(function):
name = function.__name__
mm = registry.get(name)
if mm is None:
mm = registry[name] = MultiMethod(name)
mm.register(types, function)
return mm
return register
The usage would be
from multimethods import multimethod
import unittest
# 'overload' makes more sense in this case
overload = multimethod
class Sprite(object):
pass
class Point(object):
pass
class Curve(object):
pass
#overload(Sprite, Point, Direction, int)
def add_bullet(sprite, start, direction, speed):
# ...
#overload(Sprite, Point, Point, int, int)
def add_bullet(sprite, start, headto, speed, acceleration):
# ...
#overload(Sprite, str)
def add_bullet(sprite, script):
# ...
#overload(Sprite, Curve, speed)
def add_bullet(sprite, curve, speed):
# ...
Most restrictive limitations at the moment are:
methods are not supported, only functions that are not class members;
inheritance is not handled;
kwargs are not supported;
registering new functions should be done at import time thing is not thread-safe
A possible option is to use the multipledispatch module as detailed here:
http://matthewrocklin.com/blog/work/2014/02/25/Multiple-Dispatch
Instead of doing this:
def add(self, other):
if isinstance(other, Foo):
...
elif isinstance(other, Bar):
...
else:
raise NotImplementedError()
You can do this:
from multipledispatch import dispatch
#dispatch(int, int)
def add(x, y):
return x + y
#dispatch(object, object)
def add(x, y):
return "%s + %s" % (x, y)
With the resulting usage:
>>> add(1, 2)
3
>>> add(1, 'hello')
'1 + hello'
In Python 3.4 PEP-0443. Single-dispatch generic functions was added.
Here is a short API description from PEP.
To define a generic function, decorate it with the #singledispatch decorator. Note that the dispatch happens on the type of the first argument. Create your function accordingly:
from functools import singledispatch
#singledispatch
def fun(arg, verbose=False):
if verbose:
print("Let me just say,", end=" ")
print(arg)
To add overloaded implementations to the function, use the register() attribute of the generic function. This is a decorator, taking a type parameter and decorating a function implementing the operation for that type:
#fun.register(int)
def _(arg, verbose=False):
if verbose:
print("Strength in numbers, eh?", end=" ")
print(arg)
#fun.register(list)
def _(arg, verbose=False):
if verbose:
print("Enumerate this:")
for i, elem in enumerate(arg):
print(i, elem)
The #overload decorator was added with type hints (PEP 484).
While this doesn't change the behaviour of Python, it does make it easier to understand what is going on, and for mypy to detect errors.
See: Type hints and PEP 484
This type of behaviour is typically solved (in OOP languages) using polymorphism. Each type of bullet would be responsible for knowing how it travels. For instance:
class Bullet(object):
def __init__(self):
self.curve = None
self.speed = None
self.acceleration = None
self.sprite_image = None
class RegularBullet(Bullet):
def __init__(self):
super(RegularBullet, self).__init__()
self.speed = 10
class Grenade(Bullet):
def __init__(self):
super(Grenade, self).__init__()
self.speed = 4
self.curve = 3.5
add_bullet(Grendade())
def add_bullet(bullet):
c_function(bullet.speed, bullet.curve, bullet.acceleration, bullet.sprite, bullet.x, bullet.y)
void c_function(double speed, double curve, double accel, char[] sprite, ...) {
if (speed != null && ...) regular_bullet(...)
else if (...) curved_bullet(...)
//..etc..
}
Pass as many arguments to the c_function that exist, and then do the job of determining which c function to call based on the values in the initial c function. So, Python should only ever be calling the one c function. That one c function looks at the arguments, and then can delegate to other c functions appropriately.
You're essentially just using each subclass as a different data container, but by defining all the potential arguments on the base class, the subclasses are free to ignore the ones they do nothing with.
When a new type of bullet comes along, you can simply define one more property on the base, change the one python function so that it passes the extra property, and the one c_function that examines the arguments and delegates appropriately. It doesn't sound too bad I guess.
It is impossible by definition to overload a function in python (read on for details), but you can achieve something similar with a simple decorator
class overload:
def __init__(self, f):
self.cases = {}
def args(self, *args):
def store_function(f):
self.cases[tuple(args)] = f
return self
return store_function
def __call__(self, *args):
function = self.cases[tuple(type(arg) for arg in args)]
return function(*args)
You can use it like this
#overload
def f():
pass
#f.args(int, int)
def f(x, y):
print('two integers')
#f.args(float)
def f(x):
print('one float')
f(5.5)
f(1, 2)
Modify it to adapt it to your use case.
A clarification of concepts
function dispatch: there are multiple functions with the same name. Which one should be called? two strategies
static/compile-time dispatch (aka. "overloading"). decide which function to call based on the compile-time type of the arguments. In all dynamic languages, there is no compile-time type, so overloading is impossible by definition
dynamic/run-time dispatch: decide which function to call based on the runtime type of the arguments. This is what all OOP languages do: multiple classes have the same methods, and the language decides which one to call based on the type of self/this argument. However, most languages only do it for the this argument only. The above decorator extends the idea to multiple parameters.
To clear up, assume that we define, in a hypothetical static language, the functions
void f(Integer x):
print('integer called')
void f(Float x):
print('float called')
void f(Number x):
print('number called')
Number x = new Integer('5')
f(x)
x = new Number('3.14')
f(x)
With static dispatch (overloading) you will see "number called" twice, because x has been declared as Number, and that's all overloading cares about. With dynamic dispatch you will see "integer called, float called", because those are the actual types of x at the time the function is called.
By passing keyword args.
def add_bullet(**kwargs):
#check for the arguments listed above and do the proper things
Python 3.8 added functools.singledispatchmethod
Transform a method into a single-dispatch generic function.
To define a generic method, decorate it with the #singledispatchmethod
decorator. Note that the dispatch happens on the type of the first
non-self or non-cls argument, create your function accordingly:
from functools import singledispatchmethod
class Negator:
#singledispatchmethod
def neg(self, arg):
raise NotImplementedError("Cannot negate a")
#neg.register
def _(self, arg: int):
return -arg
#neg.register
def _(self, arg: bool):
return not arg
negator = Negator()
for v in [42, True, "Overloading"]:
neg = negator.neg(v)
print(f"{v=}, {neg=}")
Output
v=42, neg=-42
v=True, neg=False
NotImplementedError: Cannot negate a
#singledispatchmethod supports nesting with other decorators such as
#classmethod. Note that to allow for dispatcher.register,
singledispatchmethod must be the outer most decorator. Here is the
Negator class with the neg methods being class bound:
from functools import singledispatchmethod
class Negator:
#singledispatchmethod
#staticmethod
def neg(arg):
raise NotImplementedError("Cannot negate a")
#neg.register
def _(arg: int) -> int:
return -arg
#neg.register
def _(arg: bool) -> bool:
return not arg
for v in [42, True, "Overloading"]:
neg = Negator.neg(v)
print(f"{v=}, {neg=}")
Output:
v=42, neg=-42
v=True, neg=False
NotImplementedError: Cannot negate a
The same pattern can be used for other similar decorators:
staticmethod, abstractmethod, and others.
I think your basic requirement is to have a C/C++-like syntax in Python with the least headache possible. Although I liked Alexander Poluektov's answer it doesn't work for classes.
The following should work for classes. It works by distinguishing by the number of non-keyword arguments (but it doesn't support distinguishing by type):
class TestOverloading(object):
def overloaded_function(self, *args, **kwargs):
# Call the function that has the same number of non-keyword arguments.
getattr(self, "_overloaded_function_impl_" + str(len(args)))(*args, **kwargs)
def _overloaded_function_impl_3(self, sprite, start, direction, **kwargs):
print "This is overload 3"
print "Sprite: %s" % str(sprite)
print "Start: %s" % str(start)
print "Direction: %s" % str(direction)
def _overloaded_function_impl_2(self, sprite, script):
print "This is overload 2"
print "Sprite: %s" % str(sprite)
print "Script: "
print script
And it can be used simply like this:
test = TestOverloading()
test.overloaded_function("I'm a Sprite", 0, "Right")
print
test.overloaded_function("I'm another Sprite", "while x == True: print 'hi'")
Output:
This is overload 3
Sprite: I'm a Sprite
Start: 0
Direction: Right
This is overload 2
Sprite: I'm another Sprite
Script:
while x == True: print 'hi'
You can achieve this with the following Python code:
#overload
def test(message: str):
return message
#overload
def test(number: int):
return number + 1
Either use multiple keyword arguments in the definition, or create a Bullet hierarchy whose instances are passed to the function.
I think a Bullet class hierarchy with the associated polymorphism is the way to go. You can effectively overload the base class constructor by using a metaclass so that calling the base class results in the creation of the appropriate subclass object. Below is some sample code to illustrate the essence of what I mean.
Updated
The code has been modified to run under both Python 2 and 3 to keep it relevant. This was done in a way that avoids the use Python's explicit metaclass syntax, which varies between the two versions.
To accomplish that objective, a BulletMetaBase instance of the BulletMeta class is created by explicitly calling the metaclass when creating the Bullet baseclass (rather than using the __metaclass__= class attribute or via a metaclass keyword argument depending on the Python version).
class BulletMeta(type):
def __new__(cls, classname, bases, classdict):
""" Create Bullet class or a subclass of it. """
classobj = type.__new__(cls, classname, bases, classdict)
if classname != 'BulletMetaBase':
if classname == 'Bullet': # Base class definition?
classobj.registry = {} # Initialize subclass registry.
else:
try:
alias = classdict['alias']
except KeyError:
raise TypeError("Bullet subclass %s has no 'alias'" %
classname)
if alias in Bullet.registry: # unique?
raise TypeError("Bullet subclass %s's alias attribute "
"%r already in use" % (classname, alias))
# Register subclass under the specified alias.
classobj.registry[alias] = classobj
return classobj
def __call__(cls, alias, *args, **kwargs):
""" Bullet subclasses instance factory.
Subclasses should only be instantiated by calls to the base
class with their subclass' alias as the first arg.
"""
if cls != Bullet:
raise TypeError("Bullet subclass %r objects should not to "
"be explicitly constructed." % cls.__name__)
elif alias not in cls.registry: # Bullet subclass?
raise NotImplementedError("Unknown Bullet subclass %r" %
str(alias))
# Create designated subclass object (call its __init__ method).
subclass = cls.registry[alias]
return type.__call__(subclass, *args, **kwargs)
class Bullet(BulletMeta('BulletMetaBase', (object,), {})):
# Presumably you'd define some abstract methods that all here
# that would be supported by all subclasses.
# These definitions could just raise NotImplementedError() or
# implement the functionality is some sub-optimal generic way.
# For example:
def fire(self, *args, **kwargs):
raise NotImplementedError(self.__class__.__name__ + ".fire() method")
# Abstract base class's __init__ should never be called.
# If subclasses need to call super class's __init__() for some
# reason then it would need to be implemented.
def __init__(self, *args, **kwargs):
raise NotImplementedError("Bullet is an abstract base class")
# Subclass definitions.
class Bullet1(Bullet):
alias = 'B1'
def __init__(self, sprite, start, direction, speed):
print('creating %s object' % self.__class__.__name__)
def fire(self, trajectory):
print('Bullet1 object fired with %s trajectory' % trajectory)
class Bullet2(Bullet):
alias = 'B2'
def __init__(self, sprite, start, headto, spead, acceleration):
print('creating %s object' % self.__class__.__name__)
class Bullet3(Bullet):
alias = 'B3'
def __init__(self, sprite, script): # script controlled bullets
print('creating %s object' % self.__class__.__name__)
class Bullet4(Bullet):
alias = 'B4'
def __init__(self, sprite, curve, speed): # for bullets with curved paths
print('creating %s object' % self.__class__.__name__)
class Sprite: pass
class Curve: pass
b1 = Bullet('B1', Sprite(), (10,20,30), 90, 600)
b2 = Bullet('B2', Sprite(), (-30,17,94), (1,-1,-1), 600, 10)
b3 = Bullet('B3', Sprite(), 'bullet42.script')
b4 = Bullet('B4', Sprite(), Curve(), 720)
b1.fire('uniform gravity')
b2.fire('uniform gravity')
Output:
creating Bullet1 object
creating Bullet2 object
creating Bullet3 object
creating Bullet4 object
Bullet1 object fired with uniform gravity trajectory
Traceback (most recent call last):
File "python-function-overloading.py", line 93, in <module>
b2.fire('uniform gravity') # NotImplementedError: Bullet2.fire() method
File "python-function-overloading.py", line 49, in fire
raise NotImplementedError(self.__class__.__name__ + ".fire() method")
NotImplementedError: Bullet2.fire() method
You can easily implement function overloading in Python. Here is an example using floats and integers:
class OverloadedFunction:
def __init__(self):
self.router = {int : self.f_int ,
float: self.f_float}
def __call__(self, x):
return self.router[type(x)](x)
def f_int(self, x):
print('Integer Function')
return x**2
def f_float(self, x):
print('Float Function (Overloaded)')
return x**3
# f is our overloaded function
f = OverloadedFunction()
print(f(3 ))
print(f(3.))
# Output:
# Integer Function
# 9
# Float Function (Overloaded)
# 27.0
The main idea behind the code is that a class holds the different (overloaded) functions that you would like to implement, and a Dictionary works as a router, directing your code towards the right function depending on the input type(x).
PS1. In case of custom classes, like Bullet1, you can initialize the internal dictionary following a similar pattern, such as self.D = {Bullet1: self.f_Bullet1, ...}. The rest of the code is the same.
PS2. The time/space complexity of the proposed solution is fairly good as well, with an average cost of O(1) per operation.
Use keyword arguments with defaults. E.g.
def add_bullet(sprite, start=default, direction=default, script=default, speed=default):
In the case of a straight bullet versus a curved bullet, I'd add two functions: add_bullet_straight and add_bullet_curved.
Overloading methods is tricky in Python. However, there could be usage of passing the dict, list or primitive variables.
I have tried something for my use cases, and this could help here to understand people to overload the methods.
Let's take your example:
A class overload method with call the methods from different class.
def add_bullet(sprite=None, start=None, headto=None, spead=None, acceleration=None):
Pass the arguments from the remote class:
add_bullet(sprite = 'test', start=Yes,headto={'lat':10.6666,'long':10.6666},accelaration=10.6}
Or
add_bullet(sprite = 'test', start=Yes, headto={'lat':10.6666,'long':10.6666},speed=['10','20,'30']}
So, handling is being achieved for list, Dictionary or primitive variables from method overloading.
Try it out for your code.
Plum supports it in a straightforward pythonic way. Copying an example from the README below.
from plum import dispatch
#dispatch
def f(x: str):
return "This is a string!"
#dispatch
def f(x: int):
return "This is an integer!"
>>> f("1")
'This is a string!'
>>> f(1)
'This is an integer!'
You can also try this code. We can try any number of arguments
# Finding the average of given number of arguments
def avg(*args): # args is the argument name we give
sum = 0
for i in args:
sum += i
average = sum/len(args) # Will find length of arguments we given
print("Avg: ", average)
# call function with different number of arguments
avg(1,2)
avg(5,6,4,7)
avg(11,23,54,111,76)
Although I have never needed this, it just struck me that making an immutable object in Python could be slightly tricky. You can't just override __setattr__, because then you can't even set attributes in the __init__. Subclassing a tuple is a trick that works:
class Immutable(tuple):
def __new__(cls, a, b):
return tuple.__new__(cls, (a, b))
#property
def a(self):
return self[0]
#property
def b(self):
return self[1]
def __str__(self):
return "<Immutable {0}, {1}>".format(self.a, self.b)
def __setattr__(self, *ignored):
raise NotImplementedError
def __delattr__(self, *ignored):
raise NotImplementedError
But then you have access to the a and b variables through self[0] and self[1], which is annoying.
Is this possible in Pure Python? If not, how would I do it with a C extension?
(Answers that work only in Python 3 are acceptable).
Update:
As of Python 3.7, the way to go is to use the #dataclass decorator, see the newly accepted answer.
Yet another solution I just thought of: The simplest way to get the same behaviour as your original code is
Immutable = collections.namedtuple("Immutable", ["a", "b"])
It does not solve the problem that attributes can be accessed via [0] etc., but at least it's considerably shorter and provides the additional advantage of being compatible with pickle and copy.
namedtuple creates a type similar to what I described in this answer, i.e. derived from tuple and using __slots__. It is available in Python 2.6 or above.
The easiest way to do this is using __slots__:
class A(object):
__slots__ = []
Instances of A are immutable now, since you can't set any attributes on them.
If you want the class instances to contain data, you can combine this with deriving from tuple:
from operator import itemgetter
class Point(tuple):
__slots__ = []
def __new__(cls, x, y):
return tuple.__new__(cls, (x, y))
x = property(itemgetter(0))
y = property(itemgetter(1))
p = Point(2, 3)
p.x
# 2
p.y
# 3
Edit: If you want to get rid of indexing either, you can override __getitem__():
class Point(tuple):
__slots__ = []
def __new__(cls, x, y):
return tuple.__new__(cls, (x, y))
#property
def x(self):
return tuple.__getitem__(self, 0)
#property
def y(self):
return tuple.__getitem__(self, 1)
def __getitem__(self, item):
raise TypeError
Note that you can't use operator.itemgetter for the properties in thise case, since this would rely on Point.__getitem__() instead of tuple.__getitem__(). Fuerthermore this won't prevent the use of tuple.__getitem__(p, 0), but I can hardly imagine how this should constitute a problem.
I don't think the "right" way of creating an immutable object is writing a C extension. Python usually relies on library implementers and library users being consenting adults, and instead of really enforcing an interface, the interface should be clearly stated in the documentation. This is why I don't consider the possibility of circumventing an overridden __setattr__() by calling object.__setattr__() a problem. If someone does this, it's on her own risk.
Using a Frozen Dataclass
For Python 3.7+ you can use a Data Class with a frozen=True option, which is a very pythonic and maintainable way to do what you want.
It would look something like that:
from dataclasses import dataclass
#dataclass(frozen=True)
class Immutable:
a: Any
b: Any
As type hinting is required for dataclasses' fields, I have used Any from the typing module.
Reasons NOT to use a Namedtuple
Before Python 3.7 it was frequent to see namedtuples being used as immutable objects. It can be tricky in many ways, one of them is that the __eq__ method between namedtuples does not consider the objects' classes. For example:
from collections import namedtuple
ImmutableTuple = namedtuple("ImmutableTuple", ["a", "b"])
ImmutableTuple2 = namedtuple("ImmutableTuple2", ["a", "c"])
obj1 = ImmutableTuple(a=1, b=2)
obj2 = ImmutableTuple2(a=1, c=2)
obj1 == obj2 # will be True
As you see, even if the types of obj1 and obj2 are different, even if their fields' names are different, obj1 == obj2 still gives True. That's because the __eq__ method used is the tuple's one, which compares only the values of the fields given their positions. That can be a huge source of errors, specially if you are subclassing these classes.
..howto do it "properly" in C..
You could use Cython to create an extension type for Python:
cdef class Immutable:
cdef readonly object a, b
cdef object __weakref__ # enable weak referencing support
def __init__(self, a, b):
self.a, self.b = a, b
It works both Python 2.x and 3.
Tests
# compile on-the-fly
import pyximport; pyximport.install() # $ pip install cython
from immutable import Immutable
o = Immutable(1, 2)
assert o.a == 1, str(o.a)
assert o.b == 2
try: o.a = 3
except AttributeError:
pass
else:
assert 0, 'attribute must be readonly'
try: o[1]
except TypeError:
pass
else:
assert 0, 'indexing must not be supported'
try: o.c = 1
except AttributeError:
pass
else:
assert 0, 'no new attributes are allowed'
o = Immutable('a', [])
assert o.a == 'a'
assert o.b == []
o.b.append(3) # attribute may contain mutable object
assert o.b == [3]
try: o.c
except AttributeError:
pass
else:
assert 0, 'no c attribute'
o = Immutable(b=3,a=1)
assert o.a == 1 and o.b == 3
try: del o.b
except AttributeError:
pass
else:
assert 0, "can't delete attribute"
d = dict(b=3, a=1)
o = Immutable(**d)
assert o.a == d['a'] and o.b == d['b']
o = Immutable(1,b=3)
assert o.a == 1 and o.b == 3
try: object.__setattr__(o, 'a', 1)
except AttributeError:
pass
else:
assert 0, 'attributes are readonly'
try: object.__setattr__(o, 'c', 1)
except AttributeError:
pass
else:
assert 0, 'no new attributes'
try: Immutable(1,c=3)
except TypeError:
pass
else:
assert 0, 'accept only a,b keywords'
for kwd in [dict(a=1), dict(b=2)]:
try: Immutable(**kwd)
except TypeError:
pass
else:
assert 0, 'Immutable requires exactly 2 arguments'
If you don't mind indexing support then collections.namedtuple suggested by #Sven Marnach is preferrable:
Immutable = collections.namedtuple("Immutable", "a b")
Another idea would be to completely disallow __setattr__ and use object.__setattr__ in the constructor:
class Point(object):
def __init__(self, x, y):
object.__setattr__(self, "x", x)
object.__setattr__(self, "y", y)
def __setattr__(self, *args):
raise TypeError
def __delattr__(self, *args):
raise TypeError
Of course you could use object.__setattr__(p, "x", 3) to modify a Point instance p, but your original implementation suffers from the same problem (try tuple.__setattr__(i, "x", 42) on an Immutable instance).
You can apply the same trick in your original implementation: get rid of __getitem__(), and use tuple.__getitem__() in your property functions.
You could create a #immutable decorator that either overrides the __setattr__ and change the __slots__ to an empty list, then decorate the __init__ method with it.
Edit: As the OP noted, changing the __slots__ attribute only prevents the creation of new attributes, not the modification.
Edit2: Here's an implementation:
Edit3: Using __slots__ breaks this code, because if stops the creation of the object's __dict__. I'm looking for an alternative.
Edit4: Well, that's it. It's a but hackish, but works as an exercise :-)
class immutable(object):
def __init__(self, immutable_params):
self.immutable_params = immutable_params
def __call__(self, new):
params = self.immutable_params
def __set_if_unset__(self, name, value):
if name in self.__dict__:
raise Exception("Attribute %s has already been set" % name)
if not name in params:
raise Exception("Cannot create atribute %s" % name)
self.__dict__[name] = value;
def __new__(cls, *args, **kws):
cls.__setattr__ = __set_if_unset__
return super(cls.__class__, cls).__new__(cls, *args, **kws)
return __new__
class Point(object):
#immutable(['x', 'y'])
def __new__(): pass
def __init__(self, x, y):
self.x = x
self.y = y
p = Point(1, 2)
p.x = 3 # Exception: Attribute x has already been set
p.z = 4 # Exception: Cannot create atribute z
I don't think it is entirely possible except by using either a tuple or a namedtuple. No matter what, if you override __setattr__() the user can always bypass it by calling object.__setattr__() directly. Any solution that depends on __setattr__ is guaranteed not to work.
The following is about the nearest you can get without using some sort of tuple:
class Immutable:
__slots__ = ['a', 'b']
def __init__(self, a, b):
object.__setattr__(self, 'a', a)
object.__setattr__(self, 'b', b)
def __setattr__(self, *ignored):
raise NotImplementedError
__delattr__ = __setattr__
but it breaks if you try hard enough:
>>> t = Immutable(1, 2)
>>> t.a
1
>>> object.__setattr__(t, 'a', 2)
>>> t.a
2
but Sven's use of namedtuple is genuinely immutable.
Update
Since the question has been updated to ask how to do it properly in C, here's my answer on how to do it properly in Cython:
First immutable.pyx:
cdef class Immutable:
cdef object _a, _b
def __init__(self, a, b):
self._a = a
self._b = b
property a:
def __get__(self):
return self._a
property b:
def __get__(self):
return self._b
def __repr__(self):
return "<Immutable {0}, {1}>".format(self.a, self.b)
and a setup.py to compile it (using the command setup.py build_ext --inplace:
from distutils.core import setup
from distutils.extension import Extension
from Cython.Distutils import build_ext
ext_modules = [Extension("immutable", ["immutable.pyx"])]
setup(
name = 'Immutable object',
cmdclass = {'build_ext': build_ext},
ext_modules = ext_modules
)
Then to try it out:
>>> from immutable import Immutable
>>> p = Immutable(2, 3)
>>> p
<Immutable 2, 3>
>>> p.a = 1
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
AttributeError: attribute 'a' of 'immutable.Immutable' objects is not writable
>>> object.__setattr__(p, 'a', 1)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
AttributeError: attribute 'a' of 'immutable.Immutable' objects is not writable
>>> p.a, p.b
(2, 3)
>>>
I've made immutable classes by overriding __setattr__, and allowing the set if the caller is __init__:
import inspect
class Immutable(object):
def __setattr__(self, name, value):
if inspect.stack()[2][3] != "__init__":
raise Exception("Can't mutate an Immutable: self.%s = %r" % (name, value))
object.__setattr__(self, name, value)
This isn't quite enough yet, since it allows anyone's ___init__ to change the object, but you get the idea.
Here's an elegant solution:
class Immutable(object):
def __setattr__(self, key, value):
if not hasattr(self, key):
super().__setattr__(key, value)
else:
raise RuntimeError("Can't modify immutable object's attribute: {}".format(key))
Inherit from this class, initialize your fields in the constructor, and you'e all set.
In addition to the excellent other answers I like to add a method for python 3.4 (or maybe 3.3). This answer builds upon several previouse answers to this question.
In python 3.4, you can use properties without setters to create class members that cannot be modified. (In earlier versions assigning to properties without a setter was possible.)
class A:
__slots__=['_A__a']
def __init__(self, aValue):
self.__a=aValue
#property
def a(self):
return self.__a
You can use it like this:
instance=A("constant")
print (instance.a)
which will print "constant"
But calling instance.a=10 will cause:
AttributeError: can't set attribute
Explaination: properties without setters are a very recent feature of python 3.4 (and I think 3.3). If you try to assign to such a property, an Error will be raised.
Using slots I restrict the membervariables to __A_a (which is __a).
Problem: Assigning to _A__a is still possible (instance._A__a=2). But if you assign to a private variable, it is your own fault...
This answer among others, however, discourages the use of __slots__. Using other ways to prevent attribute creation might be preferrable.
So, I am writing respective of python 3:
I) with the help of data class decorator and set frozen=True.
we can create immutable objects in python.
for this need to import data class from data classes lib and needs to set frozen=True
ex.
from dataclasses import dataclass
#dataclass(frozen=True)
class Location:
name: str
longitude: float = 0.0
latitude: float = 0.0
o/p:
>>> l = Location("Delhi", 112.345, 234.788)
>>> l.name
'Delhi'
>>> l.longitude
112.345
>>> l.latitude
234.788
>>> l.name = "Kolkata"
dataclasses.FrozenInstanceError: cannot assign to field 'name'
>>>
Source: https://realpython.com/python-data-classes/
If you are interested in objects with behavior, then namedtuple is almost your solution.
As described at the bottom of the namedtuple documentation, you can derive your own class from namedtuple; and then, you can add the behavior you want.
For example (code taken directly from the documentation):
class Point(namedtuple('Point', 'x y')):
__slots__ = ()
#property
def hypot(self):
return (self.x ** 2 + self.y ** 2) ** 0.5
def __str__(self):
return 'Point: x=%6.3f y=%6.3f hypot=%6.3f' % (self.x, self.y, self.hypot)
for p in Point(3, 4), Point(14, 5/7):
print(p)
This will result in:
Point: x= 3.000 y= 4.000 hypot= 5.000
Point: x=14.000 y= 0.714 hypot=14.018
This approach works for both Python 3 and Python 2.7 (tested on IronPython as well).
The only downside is that the inheritance tree is a bit weird; but this is not something you usually play with.
As of Python 3.7, you can use the #dataclass decorator in your class and it will be immutable like a struct! Though, it may or may not add a __hash__() method to your class. Quote:
hash() is used by built-in hash(), and when objects are added to hashed collections such as dictionaries and sets. Having a hash() implies that instances of the class are immutable. Mutability is a complicated property that depends on the programmer’s intent, the existence and behavior of eq(), and the values of the eq and frozen flags in the dataclass() decorator.
By default, dataclass() will not implicitly add a hash() method unless it is safe to do so. Neither will it add or change an existing explicitly defined hash() method. Setting the class attribute hash = None has a specific meaning to Python, as described in the hash() documentation.
If hash() is not explicit defined, or if it is set to None, then dataclass() may add an implicit hash() method. Although not recommended, you can force dataclass() to create a hash() method with unsafe_hash=True. This might be the case if your class is logically immutable but can nonetheless be mutated. This is a specialized use case and should be considered carefully.
Here the example from the docs linked above:
#dataclass
class InventoryItem:
'''Class for keeping track of an item in inventory.'''
name: str
unit_price: float
quantity_on_hand: int = 0
def total_cost(self) -> float:
return self.unit_price * self.quantity_on_hand
Just Like a dict
I have an open source library where I'm doing things in a functional way so moving data around in an immutable object is helpful. However, I don't want to have to transform my data object for the client to interact with them. So, I came up with this - it gives you a dict like object thats immutable + some helper methods.
Credit to Sven Marnach in his answer for the basic implementation of restricting property updating and deleting.
import json
# ^^ optional - If you don't care if it prints like a dict
# then rip this and __str__ and __repr__ out
class Immutable(object):
def __init__(self, **kwargs):
"""Sets all values once given
whatever is passed in kwargs
"""
for k,v in kwargs.items():
object.__setattr__(self, k, v)
def __setattr__(self, *args):
"""Disables setting attributes via
item.prop = val or item['prop'] = val
"""
raise TypeError('Immutable objects cannot have properties set after init')
def __delattr__(self, *args):
"""Disables deleting properties"""
raise TypeError('Immutable objects cannot have properties deleted')
def __getitem__(self, item):
"""Allows for dict like access of properties
val = item['prop']
"""
return self.__dict__[item]
def __repr__(self):
"""Print to repl in a dict like fashion"""
return self.pprint()
def __str__(self):
"""Convert to a str in a dict like fashion"""
return self.pprint()
def __eq__(self, other):
"""Supports equality operator
immutable({'a': 2}) == immutable({'a': 2})"""
if other is None:
return False
return self.dict() == other.dict()
def keys(self):
"""Paired with __getitem__ supports **unpacking
new = { **item, **other }
"""
return self.__dict__.keys()
def get(self, *args, **kwargs):
"""Allows for dict like property access
item.get('prop')
"""
return self.__dict__.get(*args, **kwargs)
def pprint(self):
"""Helper method used for printing that
formats in a dict like way
"""
return json.dumps(self,
default=lambda o: o.__dict__,
sort_keys=True,
indent=4)
def dict(self):
"""Helper method for getting the raw dict value
of the immutable object"""
return self.__dict__
Helper methods
def update(obj, **kwargs):
"""Returns a new instance of the given object with
all key/val in kwargs set on it
"""
return immutable({
**obj,
**kwargs
})
def immutable(obj):
return Immutable(**obj)
Examples
obj = immutable({
'alpha': 1,
'beta': 2,
'dalet': 4
})
obj.alpha # 1
obj['alpha'] # 1
obj.get('beta') # 2
del obj['alpha'] # TypeError
obj.alpha = 2 # TypeError
new_obj = update(obj, alpha=10)
new_obj is not obj # True
new_obj.get('alpha') == 10 # True
This way doesn't stop object.__setattr__ from working, but I've still found it useful:
class A(object):
def __new__(cls, children, *args, **kwargs):
self = super(A, cls).__new__(cls)
self._frozen = False # allow mutation from here to end of __init__
# other stuff you need to do in __new__ goes here
return self
def __init__(self, *args, **kwargs):
super(A, self).__init__()
self._frozen = True # prevent future mutation
def __setattr__(self, name, value):
# need to special case setting _frozen.
if name != '_frozen' and self._frozen:
raise TypeError('Instances are immutable.')
else:
super(A, self).__setattr__(name, value)
def __delattr__(self, name):
if self._frozen:
raise TypeError('Instances are immutable.')
else:
super(A, self).__delattr__(name)
you may need to override more stuff (like __setitem__) depending on the use case.
Classes which inherit from the following Immutable class are immutable, as are their instances, after their __init__ method finishes executing. Since it's pure python, as others have pointed out, there's nothing stopping someone from using the mutating special methods from the base object and type, but this is enough to stop anyone from mutating a class/instance by accident.
It works by hijacking the class-creation process with a metaclass.
"""Subclasses of class Immutable are immutable after their __init__ has run, in
the sense that all special methods with mutation semantics (in-place operators,
setattr, etc.) are forbidden.
"""
# Enumerate the mutating special methods
mutation_methods = set()
# Arithmetic methods with in-place operations
iarithmetic = '''add sub mul div mod divmod pow neg pos abs bool invert lshift
rshift and xor or floordiv truediv matmul'''.split()
for op in iarithmetic:
mutation_methods.add('__i%s__' % op)
# Operations on instance components (attributes, items, slices)
for verb in ['set', 'del']:
for component in '''attr item slice'''.split():
mutation_methods.add('__%s%s__' % (verb, component))
# Operations on properties
mutation_methods.update(['__set__', '__delete__'])
def checked_call(_self, name, method, *args, **kwargs):
"""Calls special method method(*args, **kw) on self if mutable."""
self = args[0] if isinstance(_self, object) else _self
if not getattr(self, '__mutable__', True):
# self told us it's immutable, so raise an error
cname= (self if isinstance(self, type) else self.__class__).__name__
raise TypeError('%s is immutable, %s disallowed' % (cname, name))
return method(*args, **kwargs)
def method_wrapper(_self, name):
"Wrap a special method to check for mutability."
method = getattr(_self, name)
def wrapper(*args, **kwargs):
return checked_call(_self, name, method, *args, **kwargs)
wrapper.__name__ = name
wrapper.__doc__ = method.__doc__
return wrapper
def wrap_mutating_methods(_self):
"Place the wrapper methods on mutative special methods of _self"
for name in mutation_methods:
if hasattr(_self, name):
method = method_wrapper(_self, name)
type.__setattr__(_self, name, method)
def set_mutability(self, ismutable):
"Set __mutable__ by using the unprotected __setattr__"
b = _MetaImmutable if isinstance(self, type) else Immutable
super(b, self).__setattr__('__mutable__', ismutable)
class _MetaImmutable(type):
'''The metaclass of Immutable. Wraps __init__ methods via __call__.'''
def __init__(cls, *args, **kwargs):
# Make class mutable for wrapping special methods
set_mutability(cls, True)
wrap_mutating_methods(cls)
# Disable mutability
set_mutability(cls, False)
def __call__(cls, *args, **kwargs):
'''Make an immutable instance of cls'''
self = cls.__new__(cls)
# Make the instance mutable for initialization
set_mutability(self, True)
# Execute cls's custom initialization on this instance
self.__init__(*args, **kwargs)
# Disable mutability
set_mutability(self, False)
return self
# Given a class T(metaclass=_MetaImmutable), mutative special methods which
# already exist on _MetaImmutable (a basic type) cannot be over-ridden
# programmatically during _MetaImmutable's instantiation of T, because the
# first place python looks for a method on an object is on the object's
# __class__, and T.__class__ is _MetaImmutable. The two extant special
# methods on a basic type are __setattr__ and __delattr__, so those have to
# be explicitly overridden here.
def __setattr__(cls, name, value):
checked_call(cls, '__setattr__', type.__setattr__, cls, name, value)
def __delattr__(cls, name, value):
checked_call(cls, '__delattr__', type.__delattr__, cls, name, value)
class Immutable(object):
"""Inherit from this class to make an immutable object.
__init__ methods of subclasses are executed by _MetaImmutable.__call__,
which enables mutability for the duration.
"""
__metaclass__ = _MetaImmutable
class T(int, Immutable): # Checks it works with multiple inheritance, too.
"Class for testing immutability semantics"
def __init__(self, b):
self.b = b
#classmethod
def class_mutation(cls):
cls.a = 5
def instance_mutation(self):
self.c = 1
def __iadd__(self, o):
pass
def not_so_special_mutation(self):
self +=1
def immutabilityTest(f, name):
"Call f, which should try to mutate class T or T instance."
try:
f()
except TypeError, e:
assert 'T is immutable, %s disallowed' % name in e.args
else:
raise RuntimeError('Immutability failed!')
immutabilityTest(T.class_mutation, '__setattr__')
immutabilityTest(T(6).instance_mutation, '__setattr__')
immutabilityTest(T(6).not_so_special_mutation, '__iadd__')
The third party attr module provides this functionality.
Edit: python 3.7 has adopted this idea into the stdlib with #dataclass.
$ pip install attrs
$ python
>>> #attr.s(frozen=True)
... class C(object):
... x = attr.ib()
>>> i = C(1)
>>> i.x = 2
Traceback (most recent call last):
...
attr.exceptions.FrozenInstanceError: can't set attribute
attr implements frozen classes by overriding __setattr__ and has a minor performance impact at each instantiation time, according to the documentation.
If you're in the habit of using classes as datatypes, attr may be especially useful as it takes care of the boilerplate for you (but doesn't do any magic). In particular, it writes nine dunder (__X__) methods for you (unless you turn any of them off), including repr, init, hash and all the comparison functions.
attr also provides a helper for __slots__.
You can override setattr and still use init to set the variable. You would use super class setattr. here is the code.
class Immutable:
__slots__ = ('a','b')
def __init__(self, a , b):
super().__setattr__('a',a)
super().__setattr__('b',b)
def __str__(self):
return "".format(self.a, self.b)
def __setattr__(self, *ignored):
raise NotImplementedError
def __delattr__(self, *ignored):
raise NotImplementedError
I found a way to do it without subclassing tuple, namedtuple etc. All you need to do is to disable setattr and delattr (and also setitem and delitem if you want to make a collection immutable) after the initiation:
def __init__(self, *args, **kwargs):
# something here
self.lock()
where lock can look like this:
#classmethod
def lock(cls):
def raiser(*a):
raise TypeError('this instance is immutable')
cls.__setattr__ = raiser
cls.__delattr__ = raiser
if hasattr(cls, '__setitem__'):
cls.__setitem__ = raiser
cls.__delitem__ = raiser
So you can create class Immutable with this method and use it the way I showed.
If you don't want to write self.lock() in every single init you can make it automatically with metaclasses:
class ImmutableType(type):
#classmethod
def change_init(mcs, original_init_method):
def __new_init__(self, *args, **kwargs):
if callable(original_init_method):
original_init_method(self, *args, **kwargs)
cls = self.__class__
def raiser(*a):
raise TypeError('this instance is immutable')
cls.__setattr__ = raiser
cls.__delattr__ = raiser
if hasattr(cls, '__setitem__'):
cls.__setitem__ = raiser
cls.__delitem__ = raiser
return __new_init__
def __new__(mcs, name, parents, kwargs):
kwargs['__init__'] = mcs.change_init(kwargs.get('__init__'))
return type.__new__(mcs, name, parents, kwargs)
class Immutable(metaclass=ImmutableType):
pass
Test
class SomeImmutableClass(Immutable):
def __init__(self, some_value: int):
self.important_attr = some_value
def some_method(self):
return 2 * self.important_attr
ins = SomeImmutableClass(3)
print(ins.some_method()) # 6
ins.important_attr += 1 # TypeError
ins.another_attr = 2 # TypeError
The basic solution below addresses the following scenario:
__init__() can be written accessing the attributes as usual.
AFTER that the OBJECT is frozen for attributes changes only:
The idea is to override __setattr__ method and replace its implementation each time the object frozen status is changed.
So we need some method (_freeze) which stores those two implementations and switches between them when requested.
This mechanism may be implemented inside the user class or inherited from a special Freezer class as shown below:
class Freezer:
def _freeze(self, do_freeze=True):
def raise_sa(*args):
raise AttributeError("Attributes are frozen and can not be changed!")
super().__setattr__('_active_setattr', (super().__setattr__, raise_sa)[do_freeze])
def __setattr__(self, key, value):
return self._active_setattr(key, value)
class A(Freezer):
def __init__(self):
self._freeze(False)
self.x = 10
self._freeze()
I needed this a little while ago and decided to make a Python package for it. The initial version is on PyPI now:
$ pip install immutable
To use:
>>> from immutable import ImmutableFactory
>>> MyImmutable = ImmutableFactory.create(prop1=1, prop2=2, prop3=3)
>>> MyImmutable.prop1
1
Full docs here: https://github.com/theengineear/immutable
Hope it helps, it wraps a namedtuple as has been discussed, but makes instantiation much simpler.
An alternative approach is to create a wrapper which makes an instance immutable.
class Immutable(object):
def __init__(self, wrapped):
super(Immutable, self).__init__()
object.__setattr__(self, '_wrapped', wrapped)
def __getattribute__(self, item):
return object.__getattribute__(self, '_wrapped').__getattribute__(item)
def __setattr__(self, key, value):
raise ImmutableError('Object {0} is immutable.'.format(self._wrapped))
__delattr__ = __setattr__
def __iter__(self):
return object.__getattribute__(self, '_wrapped').__iter__()
def next(self):
return object.__getattribute__(self, '_wrapped').next()
def __getitem__(self, item):
return object.__getattribute__(self, '_wrapped').__getitem__(item)
immutable_instance = Immutable(my_instance)
This is useful in situations where only some instances have to be immutable (like default arguments of function calls).
Can also be used in immutable factories like:
#classmethod
def immutable_factory(cls, *args, **kwargs):
return Immutable(cls.__init__(*args, **kwargs))
Also protects from object.__setattr__, but fallable to other tricks due to Python's dynamic nature.
I used the same idea as Alex: a meta-class and an "init marker", but in combination with over-writing __setattr__:
>>> from abc import ABCMeta
>>> _INIT_MARKER = '_#_in_init_#_'
>>> class _ImmutableMeta(ABCMeta):
...
... """Meta class to construct Immutable."""
...
... def __call__(cls, *args, **kwds):
... obj = cls.__new__(cls, *args, **kwds)
... object.__setattr__(obj, _INIT_MARKER, True)
... cls.__init__(obj, *args, **kwds)
... object.__delattr__(obj, _INIT_MARKER)
... return obj
...
>>> def _setattr(self, name, value):
... if hasattr(self, _INIT_MARKER):
... object.__setattr__(self, name, value)
... else:
... raise AttributeError("Instance of '%s' is immutable."
... % self.__class__.__name__)
...
>>> def _delattr(self, name):
... raise AttributeError("Instance of '%s' is immutable."
... % self.__class__.__name__)
...
>>> _im_dict = {
... '__doc__': "Mix-in class for immutable objects.",
... '__copy__': lambda self: self, # self is immutable, so just return it
... '__setattr__': _setattr,
... '__delattr__': _delattr}
...
>>> Immutable = _ImmutableMeta('Immutable', (), _im_dict)
Note: I'm calling the meta-class directly to make it work both for Python 2.x and 3.x.
>>> class T1(Immutable):
...
... def __init__(self, x=1, y=2):
... self.x = x
... self.y = y
...
>>> t1 = T1(y=8)
>>> t1.x, t1.y
(1, 8)
>>> t1.x = 7
AttributeError: Instance of 'T1' is immutable.
It does work also with slots ...:
>>> class T2(Immutable):
...
... __slots__ = 's1', 's2'
...
... def __init__(self, s1, s2):
... self.s1 = s1
... self.s2 = s2
...
>>> t2 = T2('abc', 'xyz')
>>> t2.s1, t2.s2
('abc', 'xyz')
>>> t2.s1 += 'd'
AttributeError: Instance of 'T2' is immutable.
... and multiple inheritance:
>>> class T3(T1, T2):
...
... def __init__(self, x, y, s1, s2):
... T1.__init__(self, x, y)
... T2.__init__(self, s1, s2)
...
>>> t3 = T3(12, 4, 'a', 'b')
>>> t3.x, t3.y, t3.s1, t3.s2
(12, 4, 'a', 'b')
>>> t3.y -= 3
AttributeError: Instance of 'T3' is immutable.
Note, however, that mutable attributes stay to be mutable:
>>> t3 = T3(12, [4, 7], 'a', 'b')
>>> t3.y.append(5)
>>> t3.y
[4, 7, 5]
One thing that's not really included here is total immutability... not just the parent object, but all the children as well. tuples/frozensets may be immutable for instance, but the objects that it's part of may not be. Here's a small (incomplete) version that does a decent job of enforcing immutability all the way down:
# Initialize lists
a = [1,2,3]
b = [4,5,6]
c = [7,8,9]
l = [a,b]
# We can reassign in a list
l[0] = c
# But not a tuple
t = (a,b)
#t[0] = c -> Throws exception
# But elements can be modified
t[0][1] = 4
t
([1, 4, 3], [4, 5, 6])
# Fix it back
t[0][1] = 2
li = ImmutableObject(l)
li
[[1, 2, 3], [4, 5, 6]]
# Can't assign
#li[0] = c will fail
# Can reference
li[0]
[1, 2, 3]
# But immutability conferred on returned object too
#li[0][1] = 4 will throw an exception
# Full solution should wrap all the comparison e.g. decorators.
# Also, you'd usually want to add a hash function, i didn't put
# an interface for that.
class ImmutableObject(object):
def __init__(self, inobj):
self._inited = False
self._inobj = inobj
self._inited = True
def __repr__(self):
return self._inobj.__repr__()
def __str__(self):
return self._inobj.__str__()
def __getitem__(self, key):
return ImmutableObject(self._inobj.__getitem__(key))
def __iter__(self):
return self._inobj.__iter__()
def __setitem__(self, key, value):
raise AttributeError, 'Object is read-only'
def __getattr__(self, key):
x = getattr(self._inobj, key)
if callable(x):
return x
else:
return ImmutableObject(x)
def __hash__(self):
return self._inobj.__hash__()
def __eq__(self, second):
return self._inobj.__eq__(second)
def __setattr__(self, attr, value):
if attr not in ['_inobj', '_inited'] and self._inited == True:
raise AttributeError, 'Object is read-only'
object.__setattr__(self, attr, value)
You can just override setAttr in the final statement of init. THen you can construct but not change. Obviously you can still override by usint object.setAttr but in practice most languages have some form of reflection so immutablility is always a leaky abstraction. Immutability is more about preventing clients from accidentally violating the contract of an object. I use:
=============================
The original solution offered was incorrect, this was updated based on the comments using the solution from here
The original solution is wrong in an interesting way, so it is included at the bottom.
===============================
class ImmutablePair(object):
__initialised = False # a class level variable that should always stay false.
def __init__(self, a, b):
try :
self.a = a
self.b = b
finally:
self.__initialised = True #an instance level variable
def __setattr__(self, key, value):
if self.__initialised:
self._raise_error()
else :
super(ImmutablePair, self).__setattr__(key, value)
def _raise_error(self, *args, **kw):
raise NotImplementedError("Attempted To Modify Immutable Object")
if __name__ == "__main__":
immutable_object = ImmutablePair(1,2)
print immutable_object.a
print immutable_object.b
try :
immutable_object.a = 3
except Exception as e:
print e
print immutable_object.a
print immutable_object.b
Output :
1
2
Attempted To Modify Immutable Object
1
2
======================================
Original Implementation:
It was pointed out in the comments, correctly, that this does not in fact work, as it prevents the creation of more than one object as you are overriding the class setattr method, which means a second cannot be created as self.a = will fail on the second initialisation.
class ImmutablePair(object):
def __init__(self, a, b):
self.a = a
self.b = b
ImmutablePair.__setattr__ = self._raise_error
def _raise_error(self, *args, **kw):
raise NotImplementedError("Attempted To Modify Immutable Object")
I've created a small class decorator decorator to make class immutable (except inside __init__). As part of https://github.com/google/etils.
from etils import epy
#epy.frozen
class A:
def __init__(self):
self.x = 123 # Inside `__init__`, attribute can be assigned
a = A()
a.x = 456 # AttributeError
This support inheritance too.
Implementation:
_Cls = TypeVar('_Cls')
def frozen(cls: _Cls) -> _Cls:
"""Class decorator which prevent mutating attributes after `__init__`."""
if not isinstance(cls, type):
raise TypeError(f'{cls.__name__} is not a class.')
cls.__init__ = _wrap_init(cls.__init__)
cls.__setattr__ = _wrap_setattr(cls.__setattr__)
return cls
def _wrap_init(init_fn):
"""`__init__` wrapper."""
#functools.wraps(init_fn)
def new_init(self, *args, **kwargs):
if hasattr(self, '_epy_is_init_done'):
# `_epy_is_init_done` already created, so it means we're
# a `super().__init__` call.
return init_fn(self, *args, **kwargs)
object.__setattr__(self, '_epy_is_init_done', False)
init_fn(self, *args, **kwargs)
object.__setattr__(self, '_epy_is_init_done', True)
return new_init
def _wrap_setattr(setattr_fn):
"""`__setattr__` wrapper."""
#functools.wraps(setattr_fn)
def new_setattr(self, name, value):
if not hasattr(self, '_epy_is_init_done'):
raise ValueError(
'Child of `#epy.frozen` class should be `#epy.frozen` too. (Error'
f' raised by {type(self)})'
)
if not self._epy_is_init_done: # pylint: disable=protected-access
return setattr_fn(self, name, value)
else:
raise AttributeError(
f'Cannot assign {name!r} in `#epy.frozen` class {type(self)}'
)
return new_setattr