Surprising results with Python timeit: Counter() vs defaultdict() vs dict()

Question

I obtained very surprising results with timeit, can someone tell me if I am doing something wrong ? I am using Python 2.7.

This is the contents of file speedtest_init.py:

import random

to_count = [random.randint(0, 100) for r in range(60)]

These are the contents of speedtest.py:

__author__ = 'BlueTrin'

import timeit

def test_init1():
    print(timeit.timeit('import speedtest_init'))

def test_counter1():
    s = """\
    d = defaultdict(int);
    for i in speedtest_init.to_count:
        d[i] += 1
    """
    print(timeit.timeit(s, 'from collections import defaultdict; import speedtest_init;'))

def test_counter2():
    print(timeit.timeit('d = Counter(speedtest_init.to_count);', 'from collections import Counter; import speedtest_init;'))


if __name__ == "__main__":
    test_init1()
    test_counter1()
    test_counter2()

The console output is:

C:\Python27\python.exe C:/Dev/codility/chlorum2014/speedtest.py
2.71501962931
65.7090444503
91.2953839048

Process finished with exit code 0

I think by default timeit() runs 1000000 times the code, so I need to divide the times by 1000000, but what is surprising is that the Counter is slower than the defaultdict().

Is that expected ?

EDIT:

Also using a dict is faster than a defaultdict(int):

def test_counter3():
    s = """\
    d = {};
    for i in speedtest_init.to_count:
        if i not in d:
            d[i] = 1
        else:
            d[i] += 1
    """
    print(timeit.timeit(stmt=s, setup='from collections import defaultdict; import speedtest_init;')

this last version is faster than the defaultdict(int) meaning that unless you care more about readability you should use the dict() rather than the defaultdict().

Answer 1

Yes, this is expected; the Counter() constructor uses Counter.update() which uses self.get() to load initial values rather than rely on __missing__.

Moreover, the defaultdict __missing__ factory is handled entirely in C code, especially when using another type like int() that is itself implemented in C. The Counter source is pure Python and as such the Counter.__missing__ method requires a Python frame to execute.

Because dict.get() is still handled in C, the constructor approach is the faster approach for a Counter(), provided you use the same trick Counter.update() uses and alias the self.get lookup as a local first:

>>> import timeit
>>> import random
>>> to_count = [random.randint(0, 100) for r in range(60)]
>>> timeit.timeit('for i in to_count: c[i] += 1',
...               'from collections import Counter; from __main__ import to_count; c = Counter()',
...               number=10000)
0.2510359287261963
>>> timeit.timeit('for i in to_count: c[i] = c_get(i, 0) + 1',
...               'from collections import Counter; from __main__ import to_count; c = Counter(); c_get = c.get',
...               number=10000)
0.20978617668151855

Both defaultdict and Counter are helpful classes built for their functionality, not their performance; not relying on the __missing__ hook can be faster still:

>>> timeit.timeit('for i in to_count: d[i] = d_get(i, 0) + 1',
...               'from __main__ import to_count; d = {}; d_get = d.get',
...               number=10000)
0.11437392234802246

That's a regular dictionary using an aliased dict.get() method for maximum speed. But then you'll also have to re-implement the bag behaviour of Counter, or the Counter.most_common() method. The defaultdict use cases go way beyond counting.

In Python 3.2, updating a Counter() got a speed boost by adding a C library that handles this case; see issue 10667. Testing on Python 3.4, the Counter() constructor now beats the aliased dict.get case:

>>> timeit.timeit('Counter(to_count)',
...               'from collections import Counter; from __main__ import to_count',
...               number=100000)
0.8332311600097455
>>> timeit.timeit('for i in to_count: d[i] = d_get(i, 0) + 1',
...               'from __main__ import to_count; d = {}; d_get = d.get',
...               number=100000)
0.961191965994658