DevTut

# Incompatibilities moving from Python 2 to Python 3

Unlike most languages, Python supports two major versions. Since 2008 when Python 3 was released, many have made the transition, while many have not. In order to understand both, this section covers the important differences between Python 2 and Python 3.

# Integer Division

The standard division symbol (/) operates differently in Python 3 and Python 2 when applied to integers.

When dividing an integer by another integer in Python 3, the division operation x / y represents a true division (opens new window) (uses __truediv__ (opens new window) method) and produces a floating point result. Meanwhile, the same operation in Python 2 represents a classic division (opens new window) that rounds the result down toward negative infinity (also known as taking the floor).

For example:

Code Python 2 output Python 3 output
3 / 2 1 1.5
2 / 3 0 0.6666666666666666
-3 / 2 -2 -1.5

The rounding-towards-zero behavior was deprecated in Python 2.2 (opens new window), but remains in Python 2.7 for the sake of backward compatibility and was removed in Python 3.

Note: To get a float result in Python 2 (without floor rounding) we can specify one of the operands with the decimal point. The above example of 2/3 which gives 0 in Python 2 shall be used as 2 / 3.0 or 2.0 / 3 or 2.0/3.0 to get 0.6666666666666666

Code Python 2 output Python 3 output
3.0 / 2.0 1.5 1.5
2 / 3.0 0.6666666666666666 0.6666666666666666
-3.0 / 2 -1.5 -1.5

There is also the floor division operator (opens new window) (//), which works the same way in both versions: it rounds down to the nearest integer. (although a float is returned when used with floats) In both versions the // operator maps to __floordiv__ (opens new window).

Code Python 2 output Python 3 output
3 // 2 1 1
2 // 3 0 0
-3 // 2 -2 -2
3.0 // 2.0 1.0 1.0
2.0 // 3 0.0 0.0
-3 // 2.0 -2.0 -2.0

One can explicitly enforce true division or floor division using native functions in the operator (opens new window) module:

from operator import truediv, floordiv
assert truediv(10, 8) == 1.25            # equivalent to `/` in Python 3
assert floordiv(10, 8) == 1              # equivalent to `//`

While clear and explicit, using operator functions for every division can be tedious. Changing the behavior of the / operator will often be preferred. A common practice is to eliminate typical division behavior by adding from __future__ import division as the first statement in each module:

# needs to be the first statement in a module
from __future__ import division
Code Python 2 output Python 3 output
3 / 2 1.5 1.5
2 / 3 0.6666666666666666 0.6666666666666666
-3 / 2 -1.5 -1.5

from __future__ import division guarantees that the / operator represents true division and only within the modules that contain the __future__ import, so there are no compelling reasons for not enabling it in all new modules.

Note: Some other programming languages use rounding toward zero (truncation) rather than rounding down toward negative infinity as Python does (i.e. in those languages -3 / 2 == -1). This behavior may create confusion when porting or comparing code.

Note on float operands: As an alternative to from __future__ import division, one could use the usual division symbol / and ensure that at least one of the operands is a float: 3 / 2.0 == 1.5. However, this can be considered bad practice. It is just too easy to write average = sum(items) / len(items) and forget to cast one of the arguments to float. Moreover, such cases may frequently evade notice during testing, e.g., if you test on an array containing floats but receive an array of ints in production. Additionally, if the same code is used in Python 3, programs that expect 3 / 2 == 1 to be True will not work correctly.

See PEP 238 (opens new window) for more detailed rationale why the division operator was changed in Python 3 and why old-style division should be avoided.

See the Simple Math topic (opens new window) for more about division.

# Unpacking Iterables

In Python 3, you can unpack an iterable without knowing the exact number of items in it, and even have a variable hold the end of the iterable. For that, you provide a variable that may collect a list of values. This is done by placing an asterisk before the name. For example, unpacking a list:

first, second, *tail, last = [1, 2, 3, 4, 5]
print(first)
# Out: 1
print(second)
# Out: 2
print(tail)
# Out: [3, 4]
print(last)
# Out: 5

Note: When using the *variable syntax, the variable will always be a list, even if the original type wasn't a list. It may contain zero or more elements depending on the number of elements in the original list.

first, second, *tail, last = [1, 2, 3, 4]
print(tail)
# Out: [3]

first, second, *tail, last = [1, 2, 3]
print(tail)
# Out: []
print(last)
# Out: 3

Similarly, unpacking a str:

begin, *tail = "Hello"
print(begin)
# Out: 'H'
print(tail)
# Out: ['e', 'l', 'l', 'o']

Example of unpacking a date; _ is used in this example as a throwaway variable (we are interested only in year value):

person = ('John', 'Doe', (10, 16, 2016))
*_, (*_, year_of_birth) = person
print(year_of_birth)
# Out: 2016

It is worth mentioning that, since * eats up a variable number of items, you cannot have two *s for the same iterable in an assignment - it wouldn't know how many elements go into the first unpacking, and how many in the second:

*head, *tail = [1, 2]
# Out: SyntaxError: two starred expressions in assignment

So far we have discussed unpacking in assignments. * and ** were extended in Python 3.5 (opens new window). It's now possible to have several unpacking operations in one expression:

{*range(4), 4, *(5, 6, 7)}
# Out: {0, 1, 2, 3, 4, 5, 6, 7}

It is also possible to unpack an iterable into function arguments:

iterable = [1, 2, 3, 4, 5]
print(iterable)
# Out: [1, 2, 3, 4, 5]
print(*iterable)
# Out: 1 2 3 4 5

Unpacking a dictionary uses two adjacent stars ** (PEP 448 (opens new window)):

tail = {'y': 2, 'z': 3}
{'x': 1, **tail}
 # Out: {'x': 1, 'y': 2, 'z': 3}

This allows for both overriding old values and merging dictionaries.

dict1 = {'x': 1, 'y': 1}
dict2 = {'y': 2, 'z': 3}
{**dict1, **dict2}
# Out: {'x': 1, 'y': 2, 'z': 3}

Python 3 removed tuple unpacking in functions. Hence the following doesn't work in Python 3

# Works in Python 2, but syntax error in Python 3:
map(lambda (x, y): x + y, zip(range(5), range(5)))
# Same is true for non-lambdas:
def example((x, y)):
    pass

# Works in both Python 2 and Python 3:
map(lambda x: x[0] + x[1], zip(range(5), range(5)))
# And non-lambdas, too:
def working_example(x_y):
    x, y = x_y
    pass

See PEP 3113 (opens new window) for detailed rationale.

# Strings: Bytes versus Unicode

In Python 2 there are two variants of string: those made of bytes with type (str (opens new window)) and those made of text with type (unicode (opens new window)).

In Python 2, an object of type str is always a byte sequence, but is commonly used for both text and binary data.

A string literal is interpreted as a byte string.

s = 'Cafe'    # type(s) == str

There are two exceptions: You can define a Unicode (text) literal explicitly by prefixing the literal with u:

s = u'Café'   # type(s) == unicode
b = 'Lorem ipsum'  # type(b) == str

Alternatively, you can specify that a whole module's string literals should create Unicode (text) literals:

from __future__ import unicode_literals

s = 'Café'   # type(s) == unicode
b = 'Lorem ipsum'  # type(b) == unicode

In order to check whether your variable is a string (either Unicode or a byte string), you can use:

isinstance(s, basestring)

In Python 3, the str type is a Unicode text type.

s = 'Cafe'           # type(s) == str
s = 'Café'           # type(s) == str (note the accented trailing e)

Additionally, Python 3 added a bytes object (opens new window), suitable for binary "blobs" or writing to encoding-independent files. To create a bytes object, you can prefix b to a string literal or call the string's encode method:

# Or, if you really need a byte string:
s = b'Cafe'          # type(s) == bytes
s = 'Café'.encode()  # type(s) == bytes

To test whether a value is a string, use:

isinstance(s, str)

It is also possible to prefix string literals with a u prefix to ease compatibility between Python 2 and Python 3 code bases. Since, in Python 3, all strings are Unicode by default, prepending a string literal with u has no effect:

u'Cafe' == 'Cafe'

Python 2’s raw Unicode string prefix ur is not supported, however:

>>> ur'Café'
  File "<stdin>", line 1
    ur'Café'
           ^
SyntaxError: invalid syntax

Note that you must encode (opens new window) a Python 3 text (str) object to convert it into a bytes representation of that text. The default encoding of this method is UTF-8 (opens new window).

You can use decode (opens new window) to ask a bytes object for what Unicode text it represents:

>>> b.decode()
'Café'

While the bytes type exists in both Python 2 and 3, the unicode type only exists in Python 2. To use Python 3's implicit Unicode strings in Python 2, add the following to the top of your code file:

from __future__ import unicode_literals
print(repr("hi"))
# u'hi'

Another important difference is that indexing bytes in Python 3 results in an int output like so:

b"abc"[0] == 97

Whilst slicing in a size of one results in a length 1 bytes object:

b"abc"[0:1] == b"a"

In addition, Python 3 fixes some unusual behavior (opens new window) with unicode, i.e. reversing byte strings in Python 2. For example, the following issue (opens new window) is resolved:

# -*- coding: utf8 -*-
print("Hi, my name is Łukasz Langa.")
print(u"Hi, my name is Łukasz Langa."[::-1])
print("Hi, my name is Łukasz Langa."[::-1])

# Output in Python 2
# Hi, my name is Łukasz Langa.
# .agnaL zsakuŁ si eman ym ,iH
# .agnaL zsaku�� si eman ym ,iH

# Output in Python 3
# Hi, my name is Łukasz Langa.
# .agnaL zsakuŁ si eman ym ,iH
# .agnaL zsakuŁ si eman ym ,iH

In Python 2, print (opens new window) is a statement:

print "Hello World"
print                         # print a newline
print "No newline",           # add trailing comma to remove newline 
print >>sys.stderr, "Error"   # print to stderr
print("hello")                # print "hello", since ("hello") == "hello"
print()                       # print an empty tuple "()"
print 1, 2, 3                 # print space-separated arguments: "1 2 3"
print(1, 2, 3)                # print tuple "(1, 2, 3)"

In Python 3, print() (opens new window) is a function, with keyword arguments for common uses:

print "Hello World"              # SyntaxError
print("Hello World")
print()                          # print a newline (must use parentheses)
print("No newline", end="")      # end specifies what to append (defaults to newline)
print("Error", file=sys.stderr)  # file specifies the output buffer
print("Comma", "separated", "output", sep=",")  # sep specifies the separator
print("A", "B", "C", sep="")     # null string for sep: prints as ABC
print("Flush this", flush=True)  # flush the output buffer, added in Python 3.3
print(1, 2, 3)                   # print space-separated arguments: "1 2 3"
print((1, 2, 3))                 # print tuple "(1, 2, 3)"

The print function has the following parameters:

print(*objects, sep=' ', end='\n', file=sys.stdout, flush=False)

sep is what separates the objects you pass to print. For example:

print('foo', 'bar', sep='~') # out: foo~bar
print('foo', 'bar', sep='.') # out: foo.bar

end is what the end of the print statement is followed by. For example:

print('foo', 'bar', end='!') # out: foo bar!

Printing again following a non-newline ending print statement will print to the same line:

print('foo', end='~')
print('bar')
# out: foo~bar

Note : For future compatibility, print (opens new window) function is also available in Python 2.6 onwards; however it cannot be used unless parsing of the print statement is disabled with

from __future__ import print_function

This function has exactly same format as Python 3's, except that it lacks the flush parameter.

See PEP 3105 (opens new window) for rationale.

# Differences between range and xrange functions

In Python 2, range (opens new window) function returns a list while xrange (opens new window) creates a special xrange object, which is an immutable sequence, which unlike other built-in sequence types, doesn't support slicing and has neither index nor count methods:

print(range(1, 10))
# Out: [1, 2, 3, 4, 5, 6, 7, 8, 9]

print(isinstance(range(1, 10), list))
# Out: True

print(xrange(1, 10))
# Out: xrange(1, 10)

print(isinstance(xrange(1, 10), xrange))
# Out: True

In Python 3, xrange was expanded to the range (opens new window) sequence, which thus now creates a range object. There is no xrange type:

print(range(1, 10))
# Out: range(1, 10)

print(isinstance(range(1, 10), range))
# Out: True

# print(xrange(1, 10))
# The output will be:
#Traceback (most recent call last):
#  File "<stdin>", line 1, in <module>
#NameError: name 'xrange' is not defined

Additionally, since Python 3.2, range also supports slicing, index and count:

print(range(1, 10)[3:7])
# Out: range(3, 7)
print(range(1, 10).count(5))
# Out: 1
print(range(1, 10).index(7))
# Out: 6

The advantage of using a special sequence type instead of a list is that the interpreter does not have to allocate memory for a list and populate it:

# range(10000000000000000)
# The output would be:
# Traceback (most recent call last):
#  File "<stdin>", line 1, in <module>
# MemoryError

print(xrange(100000000000000000))
# Out: xrange(100000000000000000)

Since the latter behaviour is generally desired, the former was removed in Python 3. If you still want to have a list in Python 3, you can simply use the list() constructor on a range object:

print(list(range(1, 10)))
# Out: [1, 2, 3, 4, 5, 6, 7, 8, 9]

# Compatibility

In order to maintain compatibility between both Python 2.x and Python 3.x versions, you can use the builtins (opens new window) module from the external package future (opens new window) to achieve both forward-compatiblity and backward-compatiblity:

#forward-compatible
from builtins import range

for i in range(10**8):
    pass


#backward-compatible
from past.builtins import xrange

for i in xrange(10**8):
    pass

The range in future library supports slicing, index and count in all Python versions, just like the built-in method on Python 3.2+.

# Raising and handling Exceptions

This is the Python 2 syntax, note the commas , on the raise and except lines:

try:
    raise IOError, "input/output error"
except IOError, exc:
    print exc

In Python 3, the , syntax is dropped and replaced by parenthesis and the as keyword:

try:
    raise IOError("input/output error")
except IOError as exc:
    print(exc)

For backwards compatibility, the Python 3 syntax is also available in Python 2.6 onwards, so it should be used for all new code that does not need to be compatible with previous versions.

Python 3 also adds exception chaining (opens new window), wherein you can signal that some other exception was the cause for this exception. For example

try:
    file = open('database.db')
except FileNotFoundError as e:
    raise DatabaseError('Cannot open {}') from e

The exception raised in the except statement is of type DatabaseError, but the original exception is marked as the __cause__ attribute of that exception. When the traceback is displayed, the original exception will also be displayed in the traceback:

Traceback (most recent call last):
  File "<stdin>", line 2, in <module>
FileNotFoundError

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "<stdin>", line 4, in <module>
DatabaseError('Cannot open database.db')

If you throw in an except block without explicit chaining:

try:
    file = open('database.db')
except FileNotFoundError as e:
    raise DatabaseError('Cannot open {}')

The traceback is

Traceback (most recent call last):
  File "<stdin>", line 2, in <module>
FileNotFoundError

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "<stdin>", line 4, in <module>
DatabaseError('Cannot open database.db')

Neither one is supported in Python 2.x; the original exception and its traceback will be lost if another exception is raised in the except block. The following code can be used for compatibility:

import sys
import traceback

try:
    funcWithError()
except:
    sys_vers = getattr(sys, 'version_info', (0,))
    if sys_vers < (3, 0):
        traceback.print_exc()
    raise Exception("new exception")

To "forget" the previously thrown exception, use raise from None

try:
    file = open('database.db')
except FileNotFoundError as e:
    raise DatabaseError('Cannot open {}') from None

Now the traceback would simply be

Traceback (most recent call last):
  File "<stdin>", line 4, in <module>
DatabaseError('Cannot open database.db')

Or in order to make it compatible with both Python 2 and 3 you may use the six (opens new window) package like so:

import six
try:
    file = open('database.db')
except FileNotFoundError as e:
    six.raise_from(DatabaseError('Cannot open {}'), None)

# Leaked variables in list comprehension

x = 'hello world!'
vowels = [x for x in 'AEIOU'] 

print (vowels)
# Out: ['A', 'E', 'I', 'O', 'U']
print(x)
# Out: 'U'   


x = 'hello world!'
vowels = [x for x in 'AEIOU']

print (vowels)
# Out: ['A', 'E', 'I', 'O', 'U']
print(x)
# Out: 'hello world!'

As can be seen from the example, in Python 2 the value of x was leaked: it masked hello world! and printed out U, since this was the last value of x when the loop ended.

However, in Python 3 x prints the originally defined hello world!, since the local variable from the list comprehension does not mask variables from the surrounding scope.

Additionally, neither generator expressions (available in Python since 2.5) nor dictionary or set comprehensions (which were backported to Python 2.7 from Python 3) leak variables in Python 2.

Note that in both Python 2 and Python 3, variables will leak into the surrounding scope when using a for loop:

x = 'hello world!'
vowels = []
for x in 'AEIOU':
    vowels.append(x)
print(x)
# Out: 'U'

# True, False and None

In Python 2, True, False and None are built-in constants. Which means it's possible to reassign them.

True, False = False, True
True   # False
False  # True

You can't do this with None since Python 2.4.

None = None  # SyntaxError: cannot assign to None

In Python 3, True, False, and None are now keywords.

True, False = False, True  # SyntaxError: can't assign to keyword

None = None  # SyntaxError: can't assign to keyword

# User Input

In Python 2, user input is accepted using the raw_input function,

user_input = raw_input()

While in Python 3 user input is accepted using the input function.

user_input = input()

In Python 2, the input function will accept input and interpret it. While this can be useful, it has several security considerations and was removed in Python 3. To access the same functionality, eval(input()) can be used.

To keep a script portable across the two versions, you can put the code below near the top of your Python script:

try:
    input = raw_input
except NameError:
    pass

# Comparison of different types

Objects of different types can be compared. The results are arbitrary, but consistent. They are ordered such that None is less than anything else, numeric types are smaller than non-numeric types, and everything else is ordered lexicographically by type. Thus, an int is less than a str and a tuple is greater than a list:

[1, 2] > 'foo'
# Out: False
(1, 2) > 'foo'
# Out: True
[1, 2] > (1, 2)
# Out: False
100 < [1, 'x'] < 'xyz' < (1, 'x')
# Out: True

This was originally done so a list of mixed types could be sorted and objects would be grouped together by type:

l = [7, 'x', (1, 2), [5, 6], 5, 8.0, 'y', 1.2, [7, 8], 'z']
sorted(l)
# Out: [1.2, 5, 7, 8.0, [5, 6], [7, 8], 'x', 'y', 'z', (1, 2)]

An exception is raised when comparing different (non-numeric) types:

1 < 1.5
# Out: True

[1, 2] > 'foo'
# TypeError: unorderable types: list() > str()
(1, 2) > 'foo'
# TypeError: unorderable types: tuple() > str()
[1, 2] > (1, 2)
# TypeError: unorderable types: list() > tuple()

To sort mixed lists in Python 3 by types and to achieve compatibility between versions, you have to provide a key to the sorted function:

>>> list = [1, 'hello', [3, 4], {'python': 2}, 'stackoverflow', 8, {'python': 3}, [5, 6]]
>>> sorted(list, key=str)
# Out: [1, 8, [3, 4], [5, 6], 'hello', 'stackoverflow', {'python': 2}, {'python': 3}]

Using str as the key function temporarily converts each item to a string only for the purposes of comparison. It then sees the string representation starting with either [, ', { or 0-9 and it's able to sort those (and all the following characters).

# .next() method on iterators renamed

In Python 2, an iterator can be traversed by using a method called next on the iterator itself:

g = (i for i in range(0, 3))
g.next()  # Yields 0
g.next()  # Yields 1
g.next()  # Yields 2

In Python 3 the .next method has been renamed to .__next__, acknowledging its “magic” role, so calling .next will raise an AttributeError. The correct way to access this functionality in both Python 2 and Python 3 is to call the next function with the iterator as an argument.

g = (i for i in range(0, 3))
next(g)  # Yields 0
next(g)  # Yields 1
next(g)  # Yields 2

This code is portable across versions from 2.6 through to current releases.

# Renamed modules

A few modules in the standard library have been renamed:

Old name New name
_winreg winreg
ConfigParser configparser
copy_reg copyreg
Queue queue
SocketServer socketserver
_markupbase markupbase
repr reprlib
test.test_support test.support
Tkinter tkinter
tkFileDialog tkinter.filedialog
urllib / urllib2 urllib, urllib.parse, urllib.error, urllib.response, urllib.request, urllib.robotparser

Some modules have even been converted from files to libraries. Take tkinter and urllib from above as an example.

# Compatibility

When maintaining compatibility between both Python 2.x and 3.x versions, you can use the future external package (opens new window) to enable importing top-level standard library packages with Python 3.x names on Python 2.x versions.

# filter(), map() and zip() return iterators instead of sequences

In Python 2 filter (opens new window), map (opens new window) and zip (opens new window) built-in functions return a sequence. map and zip always return a list while with filter the return type depends on the type of given parameter:

>>> s = filter(lambda x: x.isalpha(), 'a1b2c3')
>>> s
'abc'
>>> s = map(lambda x: x * x, [0, 1, 2])
>>> s
[0, 1, 4]
>>> s = zip([0, 1, 2], [3, 4, 5])
>>> s
[(0, 3), (1, 4), (2, 5)]

In Python 3 filter (opens new window), map (opens new window) and zip (opens new window) return iterator instead:

>>> it = filter(lambda x: x.isalpha(), 'a1b2c3')
>>> it
<filter object at 0x00000098A55C2518>
>>> ''.join(it)
'abc'
>>> it = map(lambda x: x * x, [0, 1, 2])
>>> it
<map object at 0x000000E0763C2D30>
>>> list(it)
[0, 1, 4]
>>> it = zip([0, 1, 2], [3, 4, 5])
>>> it
<zip object at 0x000000E0763C52C8>
>>> list(it)
[(0, 3), (1, 4), (2, 5)]

Since Python 2 itertools.izip (opens new window) is equivalent of Python 3 zip izip has been removed on Python 3.

# Removed operators <> and ``, synonymous with != and repr()

In Python 2, <> is a synonym for !=; likewise, foo is a synonym for repr(foo).

>>> 1 <> 2
True
>>> 1 <> 1
False
>>> foo = 'hello world'
>>> repr(foo)
"'hello world'"
>>> `foo`
"'hello world'"


>>> 1 <> 2
  File "<stdin>", line 1
    1 <> 2
       ^
SyntaxError: invalid syntax
>>> `foo`
  File "<stdin>", line 1
    `foo`
    ^
SyntaxError: invalid syntax

# long vs. int

In Python 2, any integer larger than a C ssize_t would be converted into the long data type, indicated by an L suffix on the literal. For example, on a 32 bit build of Python:

>>> 2**31
2147483648L
>>> type(2**31)
<type 'long'>
>>> 2**30
1073741824
>>> type(2**30)
<type 'int'>
>>> 2**31 - 1  # 2**31 is long and long - int is long
2147483647L

However, in Python 3, the long data type was removed; no matter how big the integer is, it will be an int.

2**1024
# Output: 179769313486231590772930519078902473361797697894230657273430081157732675805500963132708477322407536021120113879871393357658789768814416622492847430639474124377767893424865485276302219601246094119453082952085005768838150682342462881473913110540827237163350510684586298239947245938479716304835356329624224137216
print(-(2**1024))
# Output: -179769313486231590772930519078902473361797697894230657273430081157732675805500963132708477322407536021120113879871393357658789768814416622492847430639474124377767893424865485276302219601246094119453082952085005768838150682342462881473913110540827237163350510684586298239947245938479716304835356329624224137216
type(2**1024)
# Output: <class 'int'>

# Reduce is no longer a built-in

In Python 2, reduce is available either as a built-in function or from the functools package (version 2.6 onwards), whereas in Python 3 reduce is available only from functools. However the syntax for reduce in both Python2 and Python3 is the same and is reduce(function_to_reduce, list_to_reduce).

As an example, let us consider reducing a list to a single value by dividing each of the adjacent numbers. Here we use truediv (opens new window) function from the operator library.

In Python 2.x it is as simple as:

>>> my_list = [1, 2, 3, 4, 5]
>>> import operator
>>> reduce(operator.truediv, my_list)
0.008333333333333333

In Python 3.x the example becomes a bit more complicated:

>>> my_list = [1, 2, 3, 4, 5]
>>> import operator, functools
>>> functools.reduce(operator.truediv, my_list)
0.008333333333333333

We can also use from functools import reduce to avoid calling reduce with the namespace name.

# All classes are "new-style classes" in Python 3.

In Python 3.x all classes are new-style classes; when defining a new class python implicitly makes it inherit from object. As such, specifying object in a class definition is a completely optional:

class X: pass
class Y(object): pass

Both of these classes now contain object in their mro (method resolution order):

>>> X.__mro__
(__main__.X, object)

>>> Y.__mro__
(__main__.Y, object)

In Python 2.x classes are, by default, old-style classes; they do not implicitly inherit from object. This causes the semantics of classes to differ depending on if we explicitly add object as a base class:

class X: pass
class Y(object): pass

In this case, if we try to print the __mro__ of Y, similar output as that in the Python 3.x case will appear:

>>> Y.__mro__
(<class '__main__.Y'>, <type 'object'>)

This happens because we explicitly made Y inherit from object when defining it: class Y(object): pass. For class X which does not inherit from object the __mro__ attribute does not exist, trying to access it results in an AttributeError.

In order to ensure compatibility between both versions of Python, classes can be defined with object as a base class:

class mycls(object):
    """I am fully compatible with Python 2/3"""

Alternatively, if the __metaclass__ variable is set to type at global scope, all subsequently defined classes in a given module are implicitly new-style without needing to explicitly inherit from object:

__metaclass__ = type

class mycls:
    """I am also fully compatible with Python 2/3"""

# Absolute/Relative Imports

In Python 3, PEP 404 (opens new window) changes the way imports work from Python 2. Implicit relative imports are no longer allowed in packages and from ... import * imports are only allowed in module level code.

To achieve Python 3 behavior in Python 2:

  • the absolute imports (opens new window) feature can be enabled with from __future__ import absolute_import
  • explicit relative imports are encouraged in place of implicit relative imports

For clarification, in Python 2, a module can import the contents of another module located in the same directory as follows:

import foo

Notice the location of foo is ambiguous from the import statement alone. This type of implicit relative import is thus discouraged in favor of explicit relative imports (opens new window), which look like the following:

from .moduleY import spam
from .moduleY import spam as ham
from . import moduleY
from ..subpackage1 import moduleY
from ..subpackage2.moduleZ import eggs
from ..moduleA import foo
from ...package import bar
from ...sys import path

The dot . allows an explicit declaration of the module location within the directory tree.

# More on Relative Imports

Consider some user defined package called shapes. The directory structure is as follows:

shapes
├── __init__.py
|
├── circle.py
|
├── square.py
|
└── triangle.py

circle.py, square.py and triangle.py all import util.py as a module. How will they refer to a module in the same level?


from . import util # use util.PI, util.sq(x), etc

OR


from .util import * #use PI, sq(x), etc to call functions

The . is used for same-level relative imports.

Now, consider an alternate layout of the shapes module:

shapes
├── __init__.py
|
├── circle
│   ├── __init__.py
│   └── circle.py
|
├── square
│   ├── __init__.py
│   └── square.py
|
├── triangle
│   ├── __init__.py
│   ├── triangle.py
|
└── util.py

Now, how will these 3 classes refer to util.py?


from .. import util # use util.PI, util.sq(x), etc

OR


from ..util import * # use PI, sq(x), etc to call functions

The .. is used for parent-level relative imports. Add more .s with number of levels between the parent and child.

# map()

map() is a builtin that is useful for applying a function to elements of an iterable. In Python 2, map returns a list. In Python 3, map returns a map object, which is a generator.

# Python 2.X
>>> map(str, [1, 2, 3, 4, 5])
['1', '2', '3', '4', '5']
>>> type(_)
>>> <class 'list'>

# Python 3.X
>>> map(str, [1, 2, 3, 4, 5])
<map object at 0x*>
>>> type(_)
<class 'map'>

# We need to apply map again because we "consumed" the previous map....
>>> map(str, [1, 2, 3, 4, 5])
>>> list(_)
['1', '2', '3', '4', '5']

In Python 2, you can pass None to serve as an identity function. This no longer works in Python 3.

>>> map(None, [0, 1, 2, 3, 0, 4])
[0, 1, 2, 3, 0, 4]


>>> list(map(None, [0, 1, 2, 3, 0, 5]))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: 'NoneType' object is not callable

Moreover, when passing more than one iterable as argument in Python 2, map pads the shorter iterables with None (similar to itertools.izip_longest). In Python 3, iteration stops after the shortest iterable.

In Python 2:

>>> map(None, [1, 2, 3], [1, 2], [1, 2, 3, 4, 5])
[(1, 1, 1), (2, 2, 2), (3, None, 3), (None, None, 4), (None, None, 5)]

In Python 3:

>>> list(map(lambda x, y, z: (x, y, z), [1, 2, 3], [1, 2], [1, 2, 3, 4, 5]))
[(1, 1, 1), (2, 2, 2)]

# to obtain the same padding as in Python 2 use zip_longest from itertools
>>> import itertools
>>> list(itertools.zip_longest([1, 2, 3], [1, 2], [1, 2, 3, 4, 5]))
[(1, 1, 1), (2, 2, 2), (3, None, 3), (None, None, 4), (None, None, 5)]

Note: instead of map consider using list comprehensions, which are Python 2/3 compatible. Replacing map(str, [1, 2, 3, 4, 5]):

>>> [str(i) for i in [1, 2, 3, 4, 5]]
['1', '2', '3', '4', '5']

# The round() function tie-breaking and return type

# round() tie breaking

In Python 2, using round() (opens new window) on a number equally close to two integers will return the one furthest from 0. For example:

round(1.5)  # Out: 2.0
round(0.5)  # Out: 1.0
round(-0.5)  # Out: -1.0
round(-1.5)  # Out: -2.0

In Python 3 however, round() (opens new window) will return the even integer (aka bankers' rounding). For example:

round(1.5)  # Out: 2
round(0.5)  # Out: 0
round(-0.5)  # Out: 0
round(-1.5)  # Out: -2

The round() function follows the half to even rounding (opens new window) strategy that will round half-way numbers to the nearest even integer (for example, round(2.5) now returns 2 rather than 3.0).

As per reference in Wikipedia (opens new window), this is also known as unbiased rounding, convergent rounding, statistician's rounding, Dutch rounding, Gaussian rounding, or odd-even rounding.

Half to even rounding is part of the IEEE 754 (opens new window) standard and it's also the default rounding mode in Microsoft's .NET.

This rounding strategy tends to reduce the total rounding error. Since on average the amount of numbers that are rounded up is the same as the amount of numbers that are rounded down, rounding errors cancel out. Other rounding methods instead tend to have an upwards or downwards bias in the average error.

# round() return type

The round() function returns a float type in Python 2.7

round(4.8)
# 5.0

Starting from Python 3.0, if the second argument (number of digits) is omitted, it returns an int.

round(4.8)
# 5

# cmp function removed in Python 3

In Python 3 the cmp (opens new window) built-in function was removed, together with the __cmp__ (opens new window) special method.

From the documentation:

The cmp() (opens new window) function should be treated as gone, and the __cmp__() (opens new window) special method is no longer supported. Use __lt__() (opens new window) for sorting, __eq__() (opens new window) with __hash__() (opens new window), and other rich comparisons as needed. (If you really need the cmp() (opens new window) functionality, you could use the expression (a > b) - (a < b) as the equivalent for cmp(a, b).)

Moreover all built-in functions that accepted the cmp parameter now only accept the key keyword only parameter.

In the functools (opens new window) module there is also useful function cmp_to_key(func) (opens new window) that allows you to convert from a cmp-style function to a key-style function:

Transform an old-style comparison function to a key function. Used with tools that accept key functions (such as sorted() (opens new window), min() (opens new window), max() (opens new window), heapq.nlargest() (opens new window), heapq.nsmallest() (opens new window), itertools.groupby() (opens new window)). This function is primarily used as a transition tool for programs being converted from Python 2 which supported the use of comparison functions.

# File I/O

file is no longer a builtin name in 3.x (open still works).

Internal details of file I/O have been moved to the standard library io module, which is also the new home of StringIO:

import io
assert io.open is open # the builtin is an alias
buffer = io.StringIO()
buffer.write('hello, ') # returns number of characters written
buffer.write('world!\n')
buffer.getvalue() # 'hello, world!\n'

The file mode (text vs binary) now determines the type of data produced by reading a file (and type required for writing):

with open('data.txt') as f:
    first_line = next(f)
    assert type(first_line) is str
with open('data.bin', 'rb') as f:
    first_kb = f.read(1024)
    assert type(first_kb) is bytes

The encoding for text files defaults to whatever is returned by locale.getpreferredencoding(False). To specify an encoding explicitly, use the encoding keyword parameter:

with open('old_japanese_poetry.txt', 'shift_jis') as text:
    haiku = text.read()

# Octal Constants

In Python 2, an octal literal could be defined as

To ensure cross-compatibility, use

# Return value when writing to a file object

In Python 2, writing directly to a file handle returns None:

hi = sys.stdout.write('hello world\n')
# Out: hello world
type(hi)
# Out: <type 'NoneType'>

In Python 3, writing to a handle will return the number of characters written when writing text, and the number of bytes written when writing bytes:

import sys

char_count = sys.stdout.write('hello world 🐍\n')
# Out: hello world 🐍
char_count
# Out: 14

byte_count = sys.stdout.buffer.write(b'hello world \xf0\x9f\x90\x8d\n')
# Out: hello world 🐍
byte_count
# Out: 17

# exec statement is a function in Python 3

In Python 2, exec is a statement, with special syntax: exec code [in globals[, locals]]. In Python 3 exec is now a function: exec(code, [, globals[, locals]]), and the Python 2 syntax will raise a SyntaxError.

As print was changed from statement into a function, a __future__ import was also added. However, there is no from __future__ import exec_function, as it is not needed: the exec statement in Python 2 can be also used with syntax that looks exactly like the exec function invocation in Python 3. Thus you can change the statements

exec 'code'
exec 'code' in global_vars
exec 'code' in global_vars, local_vars

to forms

exec('code')
exec('code', global_vars)
exec('code', global_vars, local_vars)

and the latter forms are guaranteed to work identically in both Python 2 and Python 3.

# encode/decode to hex no longer available

"1deadbeef3".decode('hex')
# Out: '\x1d\xea\xdb\xee\xf3'
'\x1d\xea\xdb\xee\xf3'.encode('hex')
# Out: 1deadbeef3


"1deadbeef3".decode('hex')
# Traceback (most recent call last):
#   File "<stdin>", line 1, in <module>
# AttributeError: 'str' object has no attribute 'decode'

b"1deadbeef3".decode('hex')
# Traceback (most recent call last):
#   File "<stdin>", line 1, in <module>
# LookupError: 'hex' is not a text encoding; use codecs.decode() to handle arbitrary codecs

'\x1d\xea\xdb\xee\xf3'.encode('hex')
# Traceback (most recent call last):
#   File "<stdin>", line 1, in <module>
# LookupError: 'hex' is not a text encoding; use codecs.encode() to handle arbitrary codecs

b'\x1d\xea\xdb\xee\xf3'.encode('hex')
# Traceback (most recent call last):
#  File "<stdin>", line 1, in <module>
# AttributeError: 'bytes' object has no attribute 'encode'

However, as suggested by the error message, you can use the codecs (opens new window) module to achieve the same result:

import codecs
codecs.decode('1deadbeef4', 'hex')
# Out: b'\x1d\xea\xdb\xee\xf4'
codecs.encode(b'\x1d\xea\xdb\xee\xf4', 'hex')
# Out: b'1deadbeef4'

Note that codecs.encode (opens new window) returns a bytes object. To obtain a str object just decode to ASCII:

codecs.encode(b'\x1d\xea\xdb\xee\xff', 'hex').decode('ascii')
# Out: '1deadbeeff'

# Dictionary method changes

In Python 3, many of the dictionary methods are quite different in behaviour from Python 2, and many were removed as well: has_key, iter* and view* are gone. Instead of d.has_key(key), which had been long deprecated, one must now use key in d.

In Python 2, dictionary methods keys, values and items return lists. In Python 3 they return view objects instead; the view objects are not iterators, and they differ from them in two ways, namely:

  • they have size (one can use the len function on them)
  • they can be iterated over many times

Additionally, like with iterators, the changes in the dictionary are reflected in the view objects.

Python 2.7 has backported these methods from Python 3; they're available as viewkeys, viewvalues and viewitems. To transform Python 2 code to Python 3 code, the corresponding forms are:

  • d.keys(), d.values() and d.items() of Python 2 should be changed to list(d.keys()), list(d.values()) and list(d.items())
  • d.iterkeys(), d.itervalues() and d.iteritems() should be changed to iter(d.keys()), or even better, iter(d); iter(d.values()) and iter(d.items()) respectively
  • and finally Python 2.7 method calls d.viewkeys(), d.viewvalues() and d.viewitems() can be replaced with d.keys(), d.values() and d.items().

Porting Python 2 code that iterates over dictionary keys, values or items while mutating it is sometimes tricky. Consider:

d = {'a': 0, 'b': 1, 'c': 2, '!': 3}
for key in d.keys():
    if key.isalpha():
        del d[key]

The code looks as if it would work similarly in Python 3, but there the keys method returns a view object, not a list, and if the dictionary changes size while being iterated over, the Python 3 code will crash with RuntimeError: dictionary changed size during iteration. The solution is of course to properly write for key in list(d).

Similarly, view objects behave differently from iterators: one cannot use next() on them, and one cannot resume iteration; it would instead restart; if Python 2 code passes the return value of d.iterkeys(), d.itervalues() or d.iteritems() to a method that expects an iterator instead of an iterable, then that should be iter(d), iter(d.values()) or iter(d.items()) in Python 3.

# Class Boolean Value

In Python 2, if you want to define a class boolean value by yourself, you need to implement the __nonzero__ method on your class. The value is True by default.

class MyClass:
    def __nonzero__(self):
        return False

my_instance = MyClass()
print bool(MyClass)       # True
print bool(my_instance)   # False

In Python 3, __bool__ is used instead of __nonzero__

class MyClass:
    def __bool__(self):
        return False

my_instance = MyClass()
print(bool(MyClass))       # True
print(bool(my_instance))   # False

# hasattr function bug in Python 2

In Python 2, when a property raise a error, hasattr will ignore this property, returning False.

class A(object):
    @property
    def get(self):
        raise IOError


class B(object):
    @property
    def get(self):
        return 'get in b'

a = A()
b = B()

print 'a hasattr get: ', hasattr(a, 'get')
# output False in Python 2 (fixed, True in Python 3)
print 'b hasattr get', hasattr(b, 'get')
# output True in Python 2 and Python 3

This bug is fixed in Python3. So if you use Python 2, use

try:
    a.get
except AttributeError:
    print("no get property!")

or use getattr instead

p = getattr(a, "get", None)
if p is not None:
    print(p)
else:
    print("no get property!")

# Remarks

There are currently two supported versions of Python: 2.7 (Python 2) and 3.6 (Python 3). Additionally versions 3.3 and 3.4 receive security updates in source format.

Python 2.7 is backwards-compatible with most earlier versions of Python, and can run Python code from most 1.x and 2.x versions of Python unchanged. It is broadly available, with an extensive collection of packages. It is also considered deprecated by the CPython developers, and receives only security and bug-fix development. The CPython developers intend to abandon this version of the language in 2020 (opens new window).

According to Python Enhancement Proposal 373 (opens new window) there are no planned future releases of Python 2 after 25 June 2016, but bug fixes and security updates will be supported until 2020. (It doesn't specify what exact date in 2020 will be the sunset date of Python 2.)

Python 3 intentionally broke backwards-compatibility, to address concerns the language developers had with the core of the language. Python 3 receives new development and new features. It is the version of the language that the language developers intend to move forward with.

Over the time between the initial release of Python 3.0 and the current version, some features of Python 3 were back-ported into Python 2.6, and other parts of Python 3 were extended to have syntax compatible with Python 2. Therefore it is possible to write Python that will work on both Python 2 and Python 3, by using future imports and special modules (like six).

Future imports have to be at the beginning of your module:

from __future__ import print_function
# other imports and instructions go after __future__
print('Hello world')

For further information on the __future__ module, see the relevant page in the Python documentation (opens new window).

The 2to3 tool (opens new window) is a Python program that converts Python 2.x code to Python 3.x code, see also the Python documentation (opens new window).

The package six (opens new window) provides utilities for Python 2/3 compatibility:

  • unified access to renamed libraries
  • variables for string/unicode types
  • functions for method that got removed or has been renamed

A reference for differences between Python 2 and Python 3 can be found here (opens new window).

← 2to3 tool →