JIYIK CN >

Preallocating storage for a list or array is a common practice among programmers when they know the number of elements in advance.

Unlike C++and Java, in Python you must initialize all preallocated storage with some value. Typically, developers use dummy values ​​for this purpose, such as None, '', Falseand 0.

Python provides several ways to create fixed-size lists, each with different performance characteristics.

To compare the performance of different approaches, we will use Python's standard module timeit, which provides a convenient way to measure the running time of a small piece of Python code.

Preallocate storage for lists

The first, and fastest, method is to use *the operator, which repeats a list a specified number of times.

>>> [None] * 10
[None, None, None, None, None, None, None, None, None, None]

One million iterations ( timeitthe default iteration value) takes about 117 milliseconds.

>>> timeit("[None] * 10")
0.11655918900214601

Another approach is to rangeuse built-in functions with list comprehensions.

>>> [None for _ in range(10)]
[None, None, None, None, None, None, None, None, None, None]

It is nearly six times slower, taking 612 milliseconds per million iterations.

>>> timeit("[None for _ in range(10)]")
0.6115895550028654

The third method is to use list.append()with fora loop.

>>> a = []
>>> for _ in range(10):
...   a.append(None)
...
>>> a
[None, None, None, None, None, None, None, None, None, None]

Using a loop is the slowest method, requiring 842 milliseconds to complete one million iterations.

>>> timeit("for _ in range(10): a.append(None)", setup="a=[]")
0.8420009529945673

Preallocating storage for other sequential data structures

Since you're pre-allocating storage for a sequential data structure, it arrayprobably makes more sense to use a built-in data structure instead of a list.

>>> from array import array
>>> array('i',(0,)*10)
array('i', [0, 0, 0, 0, 0, 0, 0, 0, 0, 0])

As shown below, this method is second only to [None] * 10.

>>> timeit("array('i',(0,)*10)", setup="from array import array")
0.4557597979946877

Let's NumPycompare the above pure Python approach with Python libraries for scientific computing.

>>> from numpy import empty
>>> empty(10)
array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0.])

The NumPy method takes 589 milliseconds per million iterations.

>>> timeit("empty(10)", setup="from numpy import empty")
0.5890094790011062

However, for larger lists, the NumPy method will be faster.

>>> timeit("[None]*10000")
16.059584009999526
>>> timeit("empty(10000)", setup="from numpy import empty")
1.1065983309963485

The conclusion is that for small lists it is best to use [None] * 10, but switch to NumPy when dealing with larger amounts of sequential data empty().