cpython / Doc / library / collections.rst

Full commit

:mod:`collections` --- High-performance container datatypes

Source code: :source:`Lib/` and :source:`Lib/`

This module implements specialized container datatypes providing alternatives to Python's general purpose built-in containers, :class:`dict`, :class:`list`, :class:`set`, and :class:`tuple`.

:func:`namedtuple` factory function for creating tuple subclasses with named fields  
:class:`deque` list-like container with fast appends and pops on either end  
:class:`Counter` dict subclass for counting hashable objects  
:class:`OrderedDict` dict subclass that remembers the order entries were added  
:class:`defaultdict` dict subclass that calls a factory function to supply missing values  

In addition to the concrete container classes, the collections module provides :ref:`abstract base classes <collections-abstract-base-classes>` that can be used to test whether a class provides a particular interface, for example, whether it is hashable or a mapping.

:class:`Counter` objects

A counter tool is provided to support convenient and rapid tallies. For example:

>>> # Tally occurrences of words in a list
>>> cnt = Counter()
>>> for word in ['red', 'blue', 'red', 'green', 'blue', 'blue']:
...     cnt[word] += 1
>>> cnt
Counter({'blue': 3, 'red': 2, 'green': 1})

>>> # Find the ten most common words in Hamlet
>>> import re
>>> words = re.findall(r'\w+', open('hamlet.txt').read().lower())
>>> Counter(words).most_common(10)
[('the', 1143), ('and', 966), ('to', 762), ('of', 669), ('i', 631),
 ('you', 554),  ('a', 546), ('my', 514), ('hamlet', 471), ('in', 451)]

A :class:`Counter` is a :class:`dict` subclass for counting hashable objects. It is an unordered collection where elements are stored as dictionary keys and their counts are stored as dictionary values. Counts are allowed to be any integer value including zero or negative counts. The :class:`Counter` class is similar to bags or multisets in other languages.

Elements are counted from an iterable or initialized from another mapping (or counter):

>>> c = Counter()                           # a new, empty counter
>>> c = Counter('gallahad')                 # a new counter from an iterable
>>> c = Counter({'red': 4, 'blue': 2})      # a new counter from a mapping
>>> c = Counter(cats=4, dogs=8)             # a new counter from keyword args

Counter objects have a dictionary interface except that they return a zero count for missing items instead of raising a :exc:`KeyError`:

>>> c = Counter(['eggs', 'ham'])
>>> c['bacon']                              # count of a missing element is zero

Setting a count to zero does not remove an element from a counter. Use del to remove it entirely:

>>> c['sausage'] = 0                        # counter entry with a zero count
>>> del c['sausage']                        # del actually removes the entry

Counter objects support three methods beyond those available for all dictionaries:

The usual dictionary methods are available for :class:`Counter` objects except for two which work differently for counters.

Common patterns for working with :class:`Counter` objects:

sum(c.values())                 # total of all counts
c.clear()                       # reset all counts
list(c)                         # list unique elements
set(c)                          # convert to a set
dict(c)                         # convert to a regular dictionary
c.items()                       # convert to a list of (elem, cnt) pairs
Counter(dict(list_of_pairs))    # convert from a list of (elem, cnt) pairs
c.most_common()[:-n:-1]         # n least common elements
c += Counter()                  # remove zero and negative counts

Several mathematical operations are provided for combining :class:`Counter` objects to produce multisets (counters that have counts greater than zero). Addition and subtraction combine counters by adding or subtracting the counts of corresponding elements. Intersection and union return the minimum and maximum of corresponding counts. Each operation can accept inputs with signed counts, but the output will exclude results with counts of zero or less.

>>> c = Counter(a=3, b=1)
>>> d = Counter(a=1, b=2)
>>> c + d                       # add two counters together:  c[x] + d[x]
Counter({'a': 4, 'b': 3})
>>> c - d                       # subtract (keeping only positive counts)
Counter({'a': 2})
>>> c & d                       # intersection:  min(c[x], d[x])
Counter({'a': 1, 'b': 1})
>>> c | d                       # union:  max(c[x], d[x])
Counter({'a': 3, 'b': 2})


Counters were primarily designed to work with positive integers to represent running counts; however, care was taken to not unnecessarily preclude use cases needing other types or negative values. To help with those use cases, this section documents the minimum range and type restrictions.

  • The :class:`Counter` class itself is a dictionary subclass with no restrictions on its keys and values. The values are intended to be numbers representing counts, but you could store anything in the value field.
  • The :meth:`most_common` method requires only that the values be orderable.
  • For in-place operations such as c[key] += 1, the value type need only support addition and subtraction. So fractions, floats, and decimals would work and negative values are supported. The same is also true for :meth:`update` and :meth:`subtract` which allow negative and zero values for both inputs and outputs.
  • The multiset methods are designed only for use cases with positive values. The inputs may be negative or zero, but only outputs with positive values are created. There are no type restrictions, but the value type needs to support addition, subtraction, and comparison.
  • The :meth:`elements` method requires integer counts. It ignores zero and negative counts.

:class:`deque` objects

Returns a new deque object initialized left-to-right (using :meth:`append`) with data from iterable. If iterable is not specified, the new deque is empty.

Deques are a generalization of stacks and queues (the name is pronounced "deck" and is short for "double-ended queue"). Deques support thread-safe, memory efficient appends and pops from either side of the deque with approximately the same O(1) performance in either direction.

Though :class:`list` objects support similar operations, they are optimized for fast fixed-length operations and incur O(n) memory movement costs for pop(0) and insert(0, v) operations which change both the size and position of the underlying data representation.

If maxlen is not specified or is None, deques may grow to an arbitrary length. Otherwise, the deque is bounded to the specified maximum length. Once a bounded length deque is full, when new items are added, a corresponding number of items are discarded from the opposite end. Bounded length deques provide functionality similar to the tail filter in Unix. They are also useful for tracking transactions and other pools of data where only the most recent activity is of interest.

Deque objects support the following methods:

Deque objects also provide one read-only attribute:

In addition to the above, deques support iteration, pickling, len(d), reversed(d), copy.copy(d), copy.deepcopy(d), membership testing with the :keyword:`in` operator, and subscript references such as d[-1]. Indexed access is O(1) at both ends but slows to O(n) in the middle. For fast random access, use lists instead.


:class:`deque` Recipes

This section shows various approaches to working with deques.

Bounded length deques provide functionality similar to the tail filter in Unix:

def tail(filename, n=10):
    'Return the last n lines of a file'
    return deque(open(filename), n)

Another approach to using deques is to maintain a sequence of recently added elements by appending to the right and popping to the left:

def moving_average(iterable, n=3):
    # moving_average([40, 30, 50, 46, 39, 44]) --> 40.0 42.0 45.0 43.0
    it = iter(iterable)
    d = deque(itertools.islice(it, n-1))
    s = sum(d)
    for elem in it:
        s += elem - d.popleft()
        yield s / float(n)

The :meth:`rotate` method provides a way to implement :class:`deque` slicing and deletion. For example, a pure Python implementation of del d[n] relies on the :meth:`rotate` method to position elements to be popped:

def delete_nth(d, n):

To implement :class:`deque` slicing, use a similar approach applying :meth:`rotate` to bring a target element to the left side of the deque. Remove old entries with :meth:`popleft`, add new entries with :meth:`extend`, and then reverse the rotation. With minor variations on that approach, it is easy to implement Forth style stack manipulations such as dup, drop, swap, over, pick, rot, and roll.

:class:`defaultdict` objects

:class:`defaultdict` Examples

Using :class:`list` as the :attr:`default_factory`, it is easy to group a sequence of key-value pairs into a dictionary of lists:

>>> s = [('yellow', 1), ('blue', 2), ('yellow', 3), ('blue', 4), ('red', 1)]
>>> d = defaultdict(list)
>>> for k, v in s:
...     d[k].append(v)
>>> d.items()
[('blue', [2, 4]), ('red', [1]), ('yellow', [1, 3])]

When each key is encountered for the first time, it is not already in the mapping; so an entry is automatically created using the :attr:`default_factory` function which returns an empty :class:`list`. The :meth:`list.append` operation then attaches the value to the new list. When keys are encountered again, the look-up proceeds normally (returning the list for that key) and the :meth:`list.append` operation adds another value to the list. This technique is simpler and faster than an equivalent technique using :meth:`dict.setdefault`:

>>> d = {}
>>> for k, v in s:
...     d.setdefault(k, []).append(v)
>>> d.items()
[('blue', [2, 4]), ('red', [1]), ('yellow', [1, 3])]

Setting the :attr:`default_factory` to :class:`int` makes the :class:`defaultdict` useful for counting (like a bag or multiset in other languages):

>>> s = 'mississippi'
>>> d = defaultdict(int)
>>> for k in s:
...     d[k] += 1
>>> d.items()
[('i', 4), ('p', 2), ('s', 4), ('m', 1)]

When a letter is first encountered, it is missing from the mapping, so the :attr:`default_factory` function calls :func:`int` to supply a default count of zero. The increment operation then builds up the count for each letter.

The function :func:`int` which always returns zero is just a special case of constant functions. A faster and more flexible way to create constant functions is to use :func:`itertools.repeat` which can supply any constant value (not just zero):

>>> def constant_factory(value):
...     return itertools.repeat(value).next
>>> d = defaultdict(constant_factory('<missing>'))
>>> d.update(name='John', action='ran')
>>> '%(name)s %(action)s to %(object)s' % d
'John ran to <missing>'

Setting the :attr:`default_factory` to :class:`set` makes the :class:`defaultdict` useful for building a dictionary of sets:

>>> s = [('red', 1), ('blue', 2), ('red', 3), ('blue', 4), ('red', 1), ('blue', 4)]
>>> d = defaultdict(set)
>>> for k, v in s:
...     d[k].add(v)
>>> d.items()
[('blue', set([2, 4])), ('red', set([1, 3]))]

:func:`namedtuple` Factory Function for Tuples with Named Fields

Named tuples assign meaning to each position in a tuple and allow for more readable, self-documenting code. They can be used wherever regular tuples are used, and they add the ability to access fields by name instead of position index.


Named tuples are especially useful for assigning field names to result tuples returned by the :mod:`csv` or :mod:`sqlite3` modules:

EmployeeRecord = namedtuple('EmployeeRecord', 'name, age, title, department, paygrade')

import csv
for emp in map(EmployeeRecord._make, csv.reader(open("employees.csv", "rb"))):
    print, emp.title

import sqlite3
conn = sqlite3.connect('/companydata')
cursor = conn.cursor()
cursor.execute('SELECT name, age, title, department, paygrade FROM employees')
for emp in map(EmployeeRecord._make, cursor.fetchall()):
    print, emp.title

In addition to the methods inherited from tuples, named tuples support three additional methods and one attribute. To prevent conflicts with field names, the method and attribute names start with an underscore.

To retrieve a field whose name is stored in a string, use the :func:`getattr` function:

>>> getattr(p, 'x')

To convert a dictionary to a named tuple, use the double-star-operator (as described in :ref:`tut-unpacking-arguments`):

>>> d = {'x': 11, 'y': 22}
>>> Point(**d)
Point(x=11, y=22)

Since a named tuple is a regular Python class, it is easy to add or change functionality with a subclass. Here is how to add a calculated field and a fixed-width print format:

>>> class Point(namedtuple('Point', 'x y')):
        __slots__ = ()
        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
Point: x= 3.000  y= 4.000  hypot= 5.000
Point: x=14.000  y= 0.714  hypot=14.018

The subclass shown above sets __slots__ to an empty tuple. This helps keep memory requirements low by preventing the creation of instance dictionaries.

Subclassing is not useful for adding new, stored fields. Instead, simply create a new named tuple type from the :attr:`_fields` attribute:

>>> Point3D = namedtuple('Point3D', Point._fields + ('z',))

Default values can be implemented by using :meth:`_replace` to customize a prototype instance:

>>> Account = namedtuple('Account', 'owner balance transaction_count')
>>> default_account = Account('<owner name>', 0.0, 0)
>>> johns_account = default_account._replace(owner='John')

Enumerated constants can be implemented with named tuples, but it is simpler and more efficient to use a simple class declaration:

>>> Status = namedtuple('Status', 'open pending closed')._make(range(3))
>>>, Status.pending, Status.closed
(0, 1, 2)
>>> class Status:
        open, pending, closed = range(3)

:class:`OrderedDict` objects

Ordered dictionaries are just like regular dictionaries but they remember the order that items were inserted. When iterating over an ordered dictionary, the items are returned in the order their keys were first added.

Return an instance of a dict subclass, supporting the usual :class:`dict` methods. An OrderedDict is a dict that remembers the order that keys were first inserted. If a new entry overwrites an existing entry, the original insertion position is left unchanged. Deleting an entry and reinserting it will move it to the end.

In addition to the usual mapping methods, ordered dictionaries also support reverse iteration using :func:`reversed`.

Equality tests between :class:`OrderedDict` objects are order-sensitive and are implemented as list(od1.items())==list(od2.items()). Equality tests between :class:`OrderedDict` objects and other :class:`Mapping` objects are order-insensitive like regular dictionaries. This allows :class:`OrderedDict` objects to be substituted anywhere a regular dictionary is used.

The :class:`OrderedDict` constructor and :meth:`update` method both accept keyword arguments, but their order is lost because Python's function call semantics pass-in keyword arguments using a regular unordered dictionary.

:class:`OrderedDict` Examples and Recipes

Since an ordered dictionary remembers its insertion order, it can be used in conjuction with sorting to make a sorted dictionary:

>>> # regular unsorted dictionary
>>> d = {'banana': 3, 'apple':4, 'pear': 1, 'orange': 2}

>>> # dictionary sorted by key
>>> OrderedDict(sorted(d.items(), key=lambda t: t[0]))
OrderedDict([('apple', 4), ('banana', 3), ('orange', 2), ('pear', 1)])

>>> # dictionary sorted by value
>>> OrderedDict(sorted(d.items(), key=lambda t: t[1]))
OrderedDict([('pear', 1), ('orange', 2), ('banana', 3), ('apple', 4)])

>>> # dictionary sorted by length of the key string
>>> OrderedDict(sorted(d.items(), key=lambda t: len(t[0])))
OrderedDict([('pear', 1), ('apple', 4), ('orange', 2), ('banana', 3)])

The new sorted dictionaries maintain their sort order when entries are deleted. But when new keys are added, the keys are appended to the end and the sort is not maintained.

It is also straight-forward to create an ordered dictionary variant that remembers the order the keys were last inserted. If a new entry overwrites an existing entry, the original insertion position is changed and moved to the end:

class LastUpdatedOrderedDict(OrderedDict):
    'Store items in the order the keys were last added'

    def __setitem__(self, key, value):
        if key in self:
            del self[key]
        OrderedDict.__setitem__(self, key, value)

An ordered dictionary can be combined with the :class:`Counter` class so that the counter remembers the order elements are first encountered:

class OrderedCounter(Counter, OrderedDict):
     'Counter that remembers the order elements are first encountered'

     def __repr__(self):
         return '%s(%r)' % (self.__class__.__name__, OrderedDict(self))

     def __reduce__(self):
         return self.__class__, (OrderedDict(self),)

Collections Abstract Base Classes

The collections module offers the following :term:`ABCs <abstract base class>`:

ABC Inherits from Abstract Methods Mixin Methods
:class:`Container`   __contains__  
:class:`Hashable`   __hash__  
:class:`Iterable`   __iter__  
:class:`Iterator` :class:`Iterable` next __iter__
:class:`Sized`   __len__  
:class:`Callable`   __call__  
:class:`Sequence` :class:`Sized`, :class:`Iterable`, :class:`Container` __getitem__ __contains__, __iter__, __reversed__, index, and count
:class:`MutableSequence` :class:`Sequence` __setitem__, __delitem__, insert Inherited :class:`Sequence` methods and append, reverse, extend, pop, remove, and __iadd__
:class:`Set` :class:`Sized`, :class:`Iterable`, :class:`Container`   __le__, __lt__, __eq__, __ne__, __gt__, __ge__, __and__, __or__, __sub__, __xor__, and isdisjoint
:class:`MutableSet` :class:`Set` add, discard Inherited :class:`Set` methods and clear, pop, remove, __ior__, __iand__, __ixor__, and __isub__
:class:`Mapping` :class:`Sized`, :class:`Iterable`, :class:`Container` __getitem__ __contains__, keys, items, values, get, __eq__, and __ne__
:class:`MutableMapping` :class:`Mapping` __setitem__, __delitem__ Inherited :class:`Mapping` methods and pop, popitem, clear, update, and setdefault
:class:`MappingView` :class:`Sized`   __len__
:class:`ItemsView` :class:`MappingView`, :class:`Set`   __contains__, __iter__
:class:`KeysView` :class:`MappingView`, :class:`Set`   __contains__, __iter__
:class:`ValuesView` :class:`MappingView`   __contains__, __iter__

ABCs for classes that provide respectively the methods :meth:`__contains__`, :meth:`__hash__`, :meth:`__len__`, and :meth:`__call__`.

ABC for classes that provide the :meth:`__iter__` method. See also the definition of :term:`iterable`.

ABC for classes that provide the :meth:`__iter__` and :meth:`next` methods. See also the definition of :term:`iterator`.

ABCs for read-only and mutable :term:`sequences <sequence>`.

ABCs for read-only and mutable sets.

ABCs for read-only and mutable :term:`mappings <mapping>`.

ABCs for mapping, items, keys, and values :term:`views <view>`.

These ABCs allow us to ask classes or instances if they provide particular functionality, for example:

size = None
if isinstance(myvar, collections.Sized):
    size = len(myvar)

Several of the ABCs are also useful as mixins that make it easier to develop classes supporting container APIs. For example, to write a class supporting the full :class:`Set` API, it only necessary to supply the three underlying abstract methods: :meth:`__contains__`, :meth:`__iter__`, and :meth:`__len__`. The ABC supplies the remaining methods such as :meth:`__and__` and :meth:`isdisjoint`

class ListBasedSet(collections.Set):
     ''' Alternate set implementation favoring space over speed
         and not requiring the set elements to be hashable. '''
     def __init__(self, iterable):
         self.elements = lst = []
         for value in iterable:
             if value not in lst:
     def __iter__(self):
         return iter(self.elements)
     def __contains__(self, value):
         return value in self.elements
     def __len__(self):
         return len(self.elements)

s1 = ListBasedSet('abcdef')
s2 = ListBasedSet('defghi')
overlap = s1 & s2            # The __and__() method is supported automatically

Notes on using :class:`Set` and :class:`MutableSet` as a mixin:

  1. Since some set operations create new sets, the default mixin methods need a way to create new instances from an iterable. The class constructor is assumed to have a signature in the form ClassName(iterable). That assumption is factored-out to an internal classmethod called :meth:`_from_iterable` which calls cls(iterable) to produce a new set. If the :class:`Set` mixin is being used in a class with a different constructor signature, you will need to override :meth:`_from_iterable` with a classmethod that can construct new instances from an iterable argument.
  2. To override the comparisons (presumably for speed, as the semantics are fixed), redefine :meth:`__le__` and then the other operations will automatically follow suit.
  3. The :class:`Set` mixin provides a :meth:`_hash` method to compute a hash value for the set; however, :meth:`__hash__` is not defined because not all sets are hashable or immutable. To add set hashabilty using mixins, inherit from both :meth:`Set` and :meth:`Hashable`, then define __hash__ = Set._hash.