Subsections


9. Classes

Python's class mechanism adds classes to the language with a minimum of new syntax and semantics. It is a mixture of the class mechanisms found in C++ and Modula-3. As is true for modules, classes in Python do not put an absolute barrier between definition and user, but rather rely on the politeness of the user not to ``break into the definition.'' The most important features of classes are retained with full power, however: the class inheritance mechanism allows multiple base classes, a derived class can override any methods of its base class or classes, a method can call the method of a base class with the same name. Objects can contain an arbitrary amount of private data.

In C++ terminology, all class members (including the data members) are public, and all member functions are virtual. There are no special constructors or destructors. As in Modula-3, there are no shorthands for referencing the object's members from its methods: the method function is declared with an explicit first argument representing the object, which is provided implicitly by the call. As in Smalltalk, classes themselves are objects, albeit in the wider sense of the word: in Python, all data types are objects. This provides semantics for importing and renaming. But, just like in C++ or Modula-3, built-in types cannot be used as base classes for extension by the user. Also, like in C++ but unlike in Modula-3, most built-in operators with special syntax (arithmetic operators, subscripting etc.) can be redefined for class instances.


9.1 A Word About Terminology

Lacking universally accepted terminology to talk about classes, I will make occasional use of Smalltalk and C++ terms. (I would use Modula-3 terms, since its object-oriented semantics are closer to those of Python than C++, but I expect that few readers have heard of it.)

I also have to warn you that there's a terminological pitfall for object-oriented readers: the word ``object'' in Python does not necessarily mean a class instance. Like C++ and Modula-3, and unlike Smalltalk, not all types in Python are classes: the basic built-in types like integers and lists are not, and even somewhat more exotic types like files aren't. However, all Python types share a little bit of common semantics that is best described by using the word object.

Objects have individuality, and multiple names (in multiple scopes) can be bound to the same object. This is known as aliasing in other languages. This is usually not appreciated on a first glance at Python, and can be safely ignored when dealing with immutable basic types (numbers, strings, tuples). However, aliasing has an (intended!) effect on the semantics of Python code involving mutable objects such as lists, dictionaries, and most types representing entities outside the program (files, windows, etc.). This is usually used to the benefit of the program, since aliases behave like pointers in some respects. For example, passing an object is cheap since only a pointer is passed by the implementation; and if a function modifies an object passed as an argument, the caller will see the change -- this obviates the need for two different argument passing mechanisms as in Pascal.


9.2 Python Scopes and Name Spaces

Before introducing classes, I first have to tell you something about Python's scope rules. Class definitions play some neat tricks with name spaces, and you need to know how scopes and name spaces work to fully understand what's going on. Incidentally, knowledge about this subject is useful for any advanced Python programmer.

Let's begin with some definitions.

A name space is a mapping from names to objects. Most name spaces are currently implemented as Python dictionaries, but that's normally not noticeable in any way (except for performance), and it may change in the future. Examples of name spaces are: the set of built-in names (functions such as abs(), and built-in exception names); the global names in a module; and the local names in a function invocation. In a sense the set of attributes of an object also form a name space. The important thing to know about name spaces is that there is absolutely no relation between names in different name spaces; for instance, two different modules may both define a function ``maximize'' without confusion -- users of the modules must prefix it with the module name.

By the way, I use the word attribute for any name following a dot -- for example, in the expression z.real, real is an attribute of the object z. Strictly speaking, references to names in modules are attribute references: in the expression modname.funcname, modname is a module object and funcname is an attribute of it. In this case there happens to be a straightforward mapping between the module's attributes and the global names defined in the module: they share the same name space!9.1

Attributes may be read-only or writable. In the latter case, assignment to attributes is possible. Module attributes are writable: you can write "modname.the_answer = 42". Writable attributes may also be deleted with the del statement, e.g. "del modname.the_answer".

Name spaces are created at different moments and have different lifetimes. The name space containing the built-in names is created when the Python interpreter starts up, and is never deleted. The global name space for a module is created when the module definition is read in; normally, module name spaces also last until the interpreter quits. The statements executed by the top-level invocation of the interpreter, either read from a script file or interactively, are considered part of a module called __main__, so they have their own global name space. (The built-in names actually also live in a module; this is called __builtin__.)

The local name space for a function is created when the function is called, and deleted when the function returns or raises an exception that is not handled within the function. (Actually, forgetting would be a better way to describe what actually happens.) Of course, recursive invocations each have their own local name space.

A scope is a textual region of a Python program where a name space is directly accessible. ``Directly accessible'' here means that an unqualified reference to a name attempts to find the name in the name space.

Although scopes are determined statically, they are used dynamically. At any time during execution, exactly three nested scopes are in use (i.e., exactly three name spaces are directly accessible): the innermost scope, which is searched first, contains the local names, the middle scope, searched next, contains the current module's global names, and the outermost scope (searched last) is the name space containing built-in names.

Usually, the local scope references the local names of the (textually) current function. Outside of functions, the local scope references the same name space as the global scope: the module's name space. Class definitions place yet another name space in the local scope.

It is important to realize that scopes are determined textually: the global scope of a function defined in a module is that module's name space, no matter from where or by what alias the function is called. On the other hand, the actual search for names is done dynamically, at run time -- however, the language definition is evolving towards static name resolution, at ``compile'' time, so don't rely on dynamic name resolution! (In fact, local variables are already determined statically.)

A special quirk of Python is that assignments always go into the innermost scope. Assignments do not copy data -- they just bind names to objects. The same is true for deletions: the statement "del x" removes the binding of x from the name space referenced by the local scope. In fact, all operations that introduce new names use the local scope: in particular, import statements and function definitions bind the module or function name in the local scope. (The global statement can be used to indicate that particular variables live in the global scope.)


9.3 A First Look at Classes

Classes introduce a little bit of new syntax, three new object types, and some new semantics.


9.3.1 Class Definition Syntax

The simplest form of class definition looks like this:

class ClassName:
    <statement-1>
    .
    .
    .
    <statement-N>

Class definitions, like function definitions (def statements) must be executed before they have any effect. (You could conceivably place a class definition in a branch of an if statement, or inside a function.)

In practice, the statements inside a class definition will usually be function definitions, but other statements are allowed, and sometimes useful -- we'll come back to this later. The function definitions inside a class normally have a peculiar form of argument list, dictated by the calling conventions for methods -- again, this is explained later.

When a class definition is entered, a new name space is created, and used as the local scope -- thus, all assignments to local variables go into this new name space. In particular, function definitions bind the name of the new function here.

When a class definition is left normally (via the end), a class object is created. This is basically a wrapper around the contents of the name space created by the class definition; we'll learn more about class objects in the next section. The original local scope (the one in effect just before the class definitions was entered) is reinstated, and the class object is bound here to the class name given in the class definition header (ClassName in the example).


9.3.2 Class Objects

Class objects support two kinds of operations: attribute references and instantiation.

Attribute references use the standard syntax used for all attribute references in Python: obj.name. Valid attribute names are all the names that were in the class's name space when the class object was created. So, if the class definition looked like this:

class MyClass:
    "A simple example class"
    i = 12345
    def f(x):
        return 'hello world'

then MyClass.i and MyClass.f are valid attribute references, returning an integer and a method object, respectively. Class attributes can also be assigned to, so you can change the value of MyClass.i by assignment. __doc__ is also a valid attribute, returning the docstring belonging to the class: "A simple example class").

Class instantiation uses function notation. Just pretend that the class object is a parameterless function that returns a new instance of the class. For example (assuming the above class):

x = MyClass()

creates a new instance of the class and assigns this object to the local variable x.

The instantiation operation (``calling'' a class object) creates an empty object. Many classes like to create objects in a known initial state. Therefore a class may define a special method named __init__(), like this:

    def __init__(self):
        self.data = []

When a class defines an __init__() method, class instantiation automatically invokes __init__() for the newly-created class instance. So in this example, a new, initialized instance can be obtained by:

x = MyClass()

Of course, the __init__() method may have arguments for greater flexibility. In that case, arguments given to the class instantiation operator are passed on to __init__(). For example,

>>> class Complex:
...     def __init__(self, realpart, imagpart):
...         self.r = realpart
...         self.i = imagpart
... 
>>> x = Complex(3.0,-4.5)
>>> x.r, x.i
(3.0, -4.5)


9.3.3 Instance Objects

Now what can we do with instance objects? The only operations understood by instance objects are attribute references. There are two kinds of valid attribute names.

The first I'll call data attributes. These correspond to ``instance variables'' in Smalltalk, and to ``data members'' in C++. Data attributes need not be declared; like local variables, they spring into existence when they are first assigned to. For example, if x is the instance of MyClass created above, the following piece of code will print the value 16, without leaving a trace:

x.counter = 1
while x.counter < 10:
    x.counter = x.counter * 2
print x.counter
del x.counter

The second kind of attribute references understood by instance objects are methods. A method is a function that ``belongs to'' an object. (In Python, the term method is not unique to class instances: other object types can have methods as well, e.g., list objects have methods called append, insert, remove, sort, and so on. However, below, we'll use the term method exclusively to mean methods of class instance objects, unless explicitly stated otherwise.)

Valid method names of an instance object depend on its class. By definition, all attributes of a class that are (user-defined) function objects define corresponding methods of its instances. So in our example, x.f is a valid method reference, since MyClass.f is a function, but x.i is not, since MyClass.i is not. But x.f is not the same thing as MyClass.f -- it is a method object, not a function object.


9.3.4 Method Objects

Usually, a method is called immediately, e.g.:

x.f()

In our example, this will return the string 'hello world'. However, it is not necessary to call a method right away: x.f is a method object, and can be stored away and called at a later time. For example:

xf = x.f
while 1:
    print xf()

will continue to print "hello world" until the end of time.

What exactly happens when a method is called? You may have noticed that x.f() was called without an argument above, even though the function definition for f specified an argument. What happened to the argument? Surely Python raises an exception when a function that requires an argument is called without any -- even if the argument isn't actually used...

Actually, you may have guessed the answer: the special thing about methods is that the object is passed as the first argument of the function. In our example, the call x.f() is exactly equivalent to MyClass.f(x). In general, calling a method with a list of n arguments is equivalent to calling the corresponding function with an argument list that is created by inserting the method's object before the first argument.

If you still don't understand how methods work, a look at the implementation can perhaps clarify matters. When an instance attribute is referenced that isn't a data attribute, its class is searched. If the name denotes a valid class attribute that is a function object, a method object is created by packing (pointers to) the instance object and the function object just found together in an abstract object: this is the method object. When the method object is called with an argument list, it is unpacked again, a new argument list is constructed from the instance object and the original argument list, and the function object is called with this new argument list.


9.4 Random Remarks

[These should perhaps be placed more carefully...]

Data attributes override method attributes with the same name; to avoid accidental name conflicts, which may cause hard-to-find bugs in large programs, it is wise to use some kind of convention that minimizes the chance of conflicts, e.g., capitalize method names, prefix data attribute names with a small unique string (perhaps just an underscore), or use verbs for methods and nouns for data attributes.

Data attributes may be referenced by methods as well as by ordinary users (``clients'') of an object. In other words, classes are not usable to implement pure abstract data types. In fact, nothing in Python makes it possible to enforce data hiding -- it is all based upon convention. (On the other hand, the Python implementation, written in C, can completely hide implementation details and control access to an object if necessary; this can be used by extensions to Python written in C.)

Clients should use data attributes with care -- clients may mess up invariants maintained by the methods by stamping on their data attributes. Note that clients may add data attributes of their own to an instance object without affecting the validity of the methods, as long as name conflicts are avoided -- again, a naming convention can save a lot of headaches here.

There is no shorthand for referencing data attributes (or other methods!) from within methods. I find that this actually increases the readability of methods: there is no chance of confusing local variables and instance variables when glancing through a method.

Conventionally, the first argument of methods is often called self. This is nothing more than a convention: the name self has absolutely no special meaning to Python. (Note, however, that by not following the convention your code may be less readable by other Python programmers, and it is also conceivable that a class browser program be written which relies upon such a convention.)

Any function object that is a class attribute defines a method for instances of that class. It is not necessary that the function definition is textually enclosed in the class definition: assigning a function object to a local variable in the class is also ok. For example:

# Function defined outside the class
def f1(self, x, y):
    return min(x, x+y)

class C:
    f = f1
    def g(self):
        return 'hello world'
    h = g

Now f, g and h are all attributes of class C that refer to function objects, and consequently they are all methods of instances of C -- h being exactly equivalent to g. Note that this practice usually only serves to confuse the reader of a program.

Methods may call other methods by using method attributes of the self argument, e.g.:

class Bag:
    def __init__(self):
        self.data = []
    def add(self, x):
        self.data.append(x)
    def addtwice(self, x):
        self.add(x)
        self.add(x)

Methods may reference global names in the same way as ordinary functions. The global scope associated with a method is the module containing the class definition. (The class itself is never used as a global scope!) While one rarely encounters a good reason for using global data in a method, there are many legitimate uses of the global scope: for one thing, functions and modules imported into the global scope can be used by methods, as well as functions and classes defined in it. Usually, the class containing the method is itself defined in this global scope, and in the next section we'll find some good reasons why a method would want to reference its own class!


9.5 Inheritance

Of course, a language feature would not be worthy of the name ``class'' without supporting inheritance. The syntax for a derived class definition looks as follows:

class DerivedClassName(BaseClassName):
    <statement-1>
    .
    .
    .
    <statement-N>

The name BaseClassName must be defined in a scope containing the derived class definition. Instead of a base class name, an expression is also allowed. This is useful when the base class is defined in another module, e.g.,

class DerivedClassName(modname.BaseClassName):

Execution of a derived class definition proceeds the same as for a base class. When the class object is constructed, the base class is remembered. This is used for resolving attribute references: if a requested attribute is not found in the class, it is searched in the base class. This rule is applied recursively if the base class itself is derived from some other class.

There's nothing special about instantiation of derived classes: DerivedClassName() creates a new instance of the class. Method references are resolved as follows: the corresponding class attribute is searched, descending down the chain of base classes if necessary, and the method reference is valid if this yields a function object.

Derived classes may override methods of their base classes. Because methods have no special privileges when calling other methods of the same object, a method of a base class that calls another method defined in the same base class, may in fact end up calling a method of a derived class that overrides it. (For C++ programmers: all methods in Python are effectively virtual.)

An overriding method in a derived class may in fact want to extend rather than simply replace the base class method of the same name. There is a simple way to call the base class method directly: just call "BaseClassName.methodname(self, arguments)". This is occasionally useful to clients as well. (Note that this only works if the base class is defined or imported directly in the global scope.)


9.5.1 Multiple Inheritance

Python supports a limited form of multiple inheritance as well. A class definition with multiple base classes looks as follows:

class DerivedClassName(Base1, Base2, Base3):
    <statement-1>
    .
    .
    .
    <statement-N>

The only rule necessary to explain the semantics is the resolution rule used for class attribute references. This is depth-first, left-to-right. Thus, if an attribute is not found in DerivedClassName, it is searched in Base1, then (recursively) in the base classes of Base1, and only if it is not found there, it is searched in Base2, and so on.

(To some people breadth first -- searching Base2 and Base3 before the base classes of Base1 -- looks more natural. However, this would require you to know whether a particular attribute of Base1 is actually defined in Base1 or in one of its base classes before you can figure out the consequences of a name conflict with an attribute of Base2. The depth-first rule makes no differences between direct and inherited attributes of Base1.)

It is clear that indiscriminate use of multiple inheritance is a maintenance nightmare, given the reliance in Python on conventions to avoid accidental name conflicts. A well-known problem with multiple inheritance is a class derived from two classes that happen to have a common base class. While it is easy enough to figure out what happens in this case (the instance will have a single copy of ``instance variables'' or data attributes used by the common base class), it is not clear that these semantics are in any way useful.


9.6 Private Variables

There is limited support for class-private identifiers. Any identifier of the form __spam (at least two leading underscores, at most one trailing underscore) is now textually replaced with _classname__spam, where classname is the current class name with leading underscore(s) stripped. This mangling is done without regard of the syntactic position of the identifier, so it can be used to define class-private instance and class variables, methods, as well as globals, and even to store instance variables private to this class on instances of other classes. Truncation may occur when the mangled name would be longer than 255 characters. Outside classes, or when the class name consists of only underscores, no mangling occurs.

Name mangling is intended to give classes an easy way to define ``private'' instance variables and methods, without having to worry about instance variables defined by derived classes, or mucking with instance variables by code outside the class. Note that the mangling rules are designed mostly to avoid accidents; it still is possible for a determined soul to access or modify a variable that is considered private. This can even be useful, e.g. for the debugger, and that's one reason why this loophole is not closed. (Buglet: derivation of a class with the same name as the base class makes use of private variables of the base class possible.)

Notice that code passed to exec, eval() or evalfile() does not consider the classname of the invoking class to be the current class; this is similar to the effect of the global statement, the effect of which is likewise restricted to code that is byte-compiled together. The same restriction applies to getattr(), setattr() and delattr(), as well as when referencing __dict__ directly.

Here's an example of a class that implements its own __getattr__ and __setattr__ methods and stores all attributes in a private variable, in a way that works in Python 1.4 as well as in previous versions:

class VirtualAttributes:
    __vdict = None
    __vdict_name = locals().keys()[0]
     
    def __init__(self):
        self.__dict__[self.__vdict_name] = {}
    
    def __getattr__(self, name):
        return self.__vdict[name]
    
    def __setattr__(self, name, value):
        self.__vdict[name] = value


9.7 Odds and Ends

Sometimes it is useful to have a data type similar to the Pascal ``record'' or C ``struct'', bundling together a couple of named data items. An empty class definition will do nicely, e.g.:

class Employee:
    pass

john = Employee() # Create an empty employee record

# Fill the fields of the record
john.name = 'John Doe'
john.dept = 'computer lab'
john.salary = 1000

A piece of Python code that expects a particular abstract data type can often be passed a class that emulates the methods of that data type instead. For instance, if you have a function that formats some data from a file object, you can define a class with methods read() and readline() that gets the data from a string buffer instead, and pass it as an argument.

Instance method objects have attributes, too: m.im_self is the object of which the method is an instance, and m.im_func is the function object corresponding to the method.


9.7.1 Exceptions Can Be Classes

User-defined exceptions are no longer limited to being string objects -- they can be identified by classes as well. Using this mechanism it is possible to create extensible hierarchies of exceptions.

There are two new valid (semantic) forms for the raise statement:

raise Class, instance

raise instance

In the first form, instance must be an instance of Class or of a class derived from it. The second form is a shorthand for:

raise instance.__class__, instance

An except clause may list classes as well as string objects. A class in an except clause is compatible with an exception if it is the same class or a base class thereof (but not the other way around -- an except clause listing a derived class is not compatible with a base class). For example, the following code will print B, C, D in that order:

class B:
    pass
class C(B):
    pass
class D(C):
    pass

for c in [B, C, D]:
    try:
        raise c()
    except D:
        print "D"
    except C:
        print "C"
    except B:
        print "B"

Note that if the except clauses were reversed (with "except B" first), it would have printed B, B, B -- the first matching except clause is triggered.

When an error message is printed for an unhandled exception which is a class, the class name is printed, then a colon and a space, and finally the instance converted to a string using the built-in function str().



Footnotes

... space!9.1
Except for one thing. Module objects have a secret read-only attribute called __dict__ which returns the dictionary used to implement the module's name space; the name __dict__ is an attribute but not a global name. Obviously, using this violates the abstraction of name space implementation, and should be restricted to things like post-mortem debuggers.


Send comments on this document to python-docs@python.org.