Python 加速技巧总结

转载请注明作者和出处: http://blog.csdn.net/john_bh/

python 使用 concurrent.futures 并行处理数据,速度提高 2~6 倍。对比代码如下:

1. 避免全局变量

许多人刚开始会用 Python 语言写一些简单的脚本,当编写脚本时,通常习惯了直接将其写为全局变量。但是,由于全局变量和局部变量实现方式不同,定义在全局范围内的代码运行速度会比定义在函数中的慢不少。通过将脚本语句放入到函数中,通常可带来 15% - 30% 的速度提升。如下面对比代码:

# -*- coding :utf-8 -*-
import math
import time

start=time.time()
size = 100000
for x in range(size):
    for y in range(size):
        z = math.sqrt(x) + math.sqrt(y)

print("Global Taken Time: {}".format(time.time()-start))

def main():  # 定义到函数中,以减少全部变量使用
    start = time.time()
    size = 100000
    for x in range(size):
        for y in range(size):
            z = math.sqrt(x) + math.sqrt(y)

    print("Local Taken Time: {}".format(time.time() - start))

main()

Global Taken Time: 30.072156190872192
Local Taken Time: 24.563132286071777

2. 避免使用 .

2.1 避免模块和函数属性访问

每次使用.(属性访问操作符时)会触发特定的方法,如__getattribute__()__getattr__(),这些方法会进行字典操作,因此会带来额外的时间开销。通过from import语句,可以消除属性访问

下面代码经过三步优化:

  1. 第一步优化:使用from math import sqrt 替换import math,避免math.sqrt的使用;
  2. 第二部优化:将在局部函数中使用 sqrt = math.sqrt 赋值给局部变量,进行局部化处理;
  3. 第三步优化:优化函数中调用 list 的append方法。通过将该方法赋值给一个局部变量,可以彻底消除函数中for循环内部的.使用。

代码实现以及运行效率如下:

# -*- coding :utf-8 -*-
import math
import time

def computeSqrt1(size: int):
    """
        不推荐写法
    """
    result = []
    for i in range(size):
        result.append(math.sqrt(i))
    return result

from math import sqrt
def computeSqrt2(size: int):
    """
        优化了1.from math import sqrt,避免math.sqrt的使用
    """
    result = []
    for i in range(size):
        result.append(sqrt(i))  # 避免math.sqrt的使用
    return result

def computeSqrt3(size: int):
    """
        优化了1.from math import sqrt,避免math.sqrt的使用
              2.sqrt = math.sqrt  赋值给局部变量
    """
    result = []
    sqrt = math.sqrt  # 赋值给局部变量
    for i in range(size):
        result.append(sqrt(i))  # 避免math.sqrt的使用
    return result

def computeSqrt4(size: int):
    """
        优化了1.from math import sqrt,避免math.sqrt的使用
              2.sqrt = math.sqrt  赋值给局部变量
              3.result.append(sqrt(i))优化为 append = result.append,append(sqrt(i))  避免 result.append 和 math.sqrt 的使用
    """
    result = []
    append = result.append
    sqrt = math.sqrt    # 赋值给局部变量
    for i in range(size):
        append(sqrt(i))  # 避免 result.append 和 math.sqrt 的使用
    return result

def test(computeSqrt_fun):
    size = 10000
    start=time.time()
    for _ in range(size):
        result = computeSqrt_fun(size)
    takenTime=time.time() - start

    print("{} Taken Time: {}".format(computeSqrt_fun.__name__,takenTime))

if __name__ == "__main__":
    test(computeSqrt1)
    test(computeSqrt2)
    test(computeSqrt3)
    test(computeSqrt4)

computeSqrt1 Taken Time: 20.50309944152832
computeSqrt2 Taken Time: 13.954072952270508
computeSqrt3 Taken Time: 13.141065120697021
computeSqrt4 Taken Time: 11.455055952072144

2.2 避免类内属性访问

避免 . 的原则也适用于类内属性,访问self._value的速度会比访问一个局部变量更慢一些。通过将需要频繁访问的类内属性赋值给一个局部变量,可以提升代码运行速度。对比代码如下所示:

# -*- coding :utf-8 -*-
import math
import time
from typing import List

class DemoClass:
    def __init__(self, value: int):
        self._value = value

    def computeSqrt1(self, size: int) -> List[float]:  # 不推荐
        result = []
        append = result.append
        sqrt = math.sqrt
        for _ in range(size):
            append(sqrt(self._value))
        return result

    def computeSqrt2(self, size: int) -> List[float]:
        result = []
        append = result.append
        sqrt = math.sqrt
        value = self._value  # 避免 self._value 的使用
        for _ in range(size):
            append(sqrt(value))  # 避免 self._value 的使用
        return result

def test():
    size = 10000
    start1 = time.time()
    for _ in range(size):
        demo_instance = DemoClass(size)
        result = demo_instance.computeSqrt1(size)
    takenTime1 = time.time() - start1
    print("computeSqrt1 Taken Time: {}".format(takenTime1))

    start2 = time.time()
    for _ in range(size):
        demo_instance = DemoClass(size)
        result = demo_instance.computeSqrt2(size)
    takenTime2=time.time() - start2
    print("computeSqrt2 Taken Time: {}".format(takenTime2))
if __name__ == "__main__":
    test()

computeSqrt1 Taken Time: 13.82009768486023
computeSqrt2 Taken Time: 11.740054607391357

3. 避免不必要的抽象

任何时候当你使用额外的处理层(比如装饰器、属性访问、描述器)去包装代码时,都会让代码变慢。大部分情况下,需要重新进行审视使用属性访问器的定义是否有必要,使用getter/setter函数对属性进行访问通常是 C/C++ 程序员遗留下来的代码风格。如果真的没有必要,就使用简单属性。对比代码如下:

# -*- coding :utf-8 -*-

import time

class DemoClass1:
    """
        不推荐使用
    """
    def __init__(self, value: int):
        self.value = value

    @property
    def value(self) -> int:
        return self._value

    @value.setter
    def value(self, x: int):
        self._value = x

class DemoClass2:
    """
        优化后,避免不必要的抽象
    """
    def __init__(self, value: int):
        self.value = value  # 避免不必要的属性访问器

def test():
    size = 1000000

    start1 = time.time()
    for i in range(size):
        demo_instance = DemoClass1(size)
        value = demo_instance.value
        demo_instance.value = i
    takenTime1 = time.time() - start1
    print("DemoClass1 Taken Time: {}".format(takenTime1))

    start2 = time.time()
    for i in range(size):
        demo_instance = DemoClass2(size)
        value = demo_instance.value
        demo_instance.value = i
    takenTime2 = time.time() - start2
    print("DemoClass2 Taken Time: {}".format(takenTime2))

if __name__ == "__main__":
    test()

DemoClass1 Taken Time: 0.6100289821624756
DemoClass2 Taken Time: 0.3529999256134033

4. 避免数据复制

4.1 避免无意义的数据复制

对比代码如下:

# -*- coding :utf-8 -*-

import time

def fun1():
    """
        不推荐
    """
    size = 10000
    start1 = time.time()
    for _ in range(size):
        value = range(size)
        value_list = [x for x in value]
        square_list = [x * x for x in value_list]
    takenTime1 = time.time() - start1
    print("fun1 Taken Time: {}".format(takenTime1))

def fun2():
    size = 10000
    start2 = time.time()
    for _ in range(size):
        value = range(size)
        square_list = [x * x for x in value]  # 避免无意义的复制
    takenTime2 = time.time() - start2
    print("fun2 Taken Time: {}".format(takenTime2))
if __name__ == "__main__":
    fun1()
    fun2()

fun1 Taken Time: 7.48906683921814
fun2 Taken Time: 5.61405348777771

另外一种情况是对 Python 的数据共享机制过于偏执,并没有很好地理解或信任 Python 的内存模型,滥用 copy.deepcopy()之类的函数。通常在这些代码中是可以去掉复制操作的。

4.2 交换值时不使用中间变量

交换值时创建了一个临时变量temp,如果不借助中间变量,代码更为简洁、且运行速度更快。

# -*- coding :utf-8 -*-

import time

def fun1():
    """
        不推荐
    """
    size = 1000000

    start1 = time.time()

    for _ in range(size):
        a = 3
        b = 5
        temp = a
        a = b
        b = temp

    takenTime1 = time.time() - start1
    print("fun1 Taken Time: {}".format(takenTime1))

def fun2():
    size = 1000000

    start2 = time.time()

    for _ in range(size):
        a = 3
        b = 5
        a, b = b, a  # 不借助中间变量

    takenTime2 = time.time() - start2
    print("fun2 Taken Time: {}".format(takenTime2))

if __name__ == "__main__":
    fun1()
    fun2()

fun1 Taken Time: 0.04600167274475098
fun2 Taken Time: 0.0330042839050293

4.3 字符串拼接用 join 而不是 +

当使用 a + b 拼接字符串时,由于 Python 中字符串是不可变对象,其会申请一块内存空间,将a和b分别复制到该新申请的内存空间中。因此,如果要拼接 n n n 个字符串,会产生 n − 1 n-1 n1 个中间结果,每产生一个中间结果都需要申请和复制一次内存,严重影响运行效率。而使用join()拼接字符串时,会首先计算出需要申请的总的内存空间,然后一次性地申请所需内存,并将每个字符串元素复制到该内存中去。

# -*- coding :utf-8 -*-

import time
import string
from typing import List

def concatString1(string_list: List[str]) -> str:
    """
        不推荐
    """
    result = ''
    for str_i in string_list:
        result += str_i
    return result

def concatString2(string_list: List[str]) -> str:
    return ''.join(string_list)  # 使用 join 而不是 +

def test():
    string_list = list(string.ascii_letters * 100)

    start1 = time.time()
    for _ in range(10000):
        result = concatString1(string_list)

    takenTime1 = time.time() - start1
    print("concatString1 by + Taken Time: {}".format(takenTime1))

    start2 = time.time()
    for _ in range(10000):
        result = concatString2(string_list)

    takenTime2 = time.time() - start2
    print("concatString2 by join Taken Time: {}".format(takenTime2))


if __name__ == "__main__":
    test()

concatString1 by + Taken Time: 9.188046216964722
concatString2 by join Taken Time: 0.3179972171783447

5. 利用if条件的短路特性

if 条件的短路特性是指对if a and b这样的语句, 当aFalse时将直接返回,不再计算b;对于if a or b这样的语句,当aTrue时将直接返回,不再计算b。因此, 为了节约运行时间,对于or语句,应该将值为True可能性比较高的变量写在or前,而and应该推后。

# -*- coding :utf-8 -*-

import time
import string
from typing import List

def concatString1(string_list: List[str]) -> str:
    abbreviations = {'cf.', 'e.g.', 'ex.', 'etc.', 'flg.', 'i.e.', 'Mr.', 'vs.'}
    result = ''
    for str_i in string_list:
        if str_i in abbreviations:
            result += str_i
    return result

def concatString2(string_list: List[str]) -> str:
    abbreviations = {'cf.', 'e.g.', 'ex.', 'etc.', 'flg.', 'i.e.', 'Mr.', 'vs.'}
    result = ''
    for str_i in string_list:
        if str_i[-1] == '.' and str_i in abbreviations:  # 利用 if 条件的短路特性
            result += str_i
    return result

def test():
    string_list = ['Mr.', 'Hat', 'is', 'Chasing', 'the', 'black', 'cat', '.']

    start1 = time.time()
    for _ in range(100000):
        result = concatString1(string_list)
    takenTime1 = time.time() - start1
    print("concatString1 Taken Time: {}".format(takenTime1))
    
    start2 = time.time()
    for _ in range(100000):
        result = concatString2(string_list)
    takenTime2 = time.time() - start2
    print("concatString2 Taken Time: {}".format(takenTime2))

if __name__ == "__main__":
    test()

理论上应该concatString2 比concatString1 快,但是结果是:

concatString1 Taken Time: 0.051025390625
concatString2 Taken Time: 0.07299995422363281

6. 循环优化

6.1 用for循环代替while循环,使用隐式for循环代替显式for循环

# -*- coding :utf-8 -*-

import time

def computeSum_while(size: int) -> int:
    """
        不推荐
    """
    sum_ = 0
    i = 0
    while i < size:
        sum_ += i
        i += 1
    return sum_

def computeSum_for(size: int) -> int:
    sum_ = 0
    for i in range(size):  # for 循环代替 while 循环
        sum_ += i
    return sum_

def computeSum(size: int) -> int:
    return sum(range(size))  # 隐式 for 循环代替显式 for 循环

def test():
    size = 10000

    start1 = time.time()
    for _ in range(size):
        sum_ = computeSum_while(size)
    takenTime1 = time.time() - start1
    print("computeSum_while Taken Time: {}".format(takenTime1))

    start2 = time.time()
    for _ in range(size):
        sum_ = computeSum_for(size)
    takenTime2 = time.time() - start2
    print("computeSum_for Taken Time: {}".format(takenTime2))

    start3 = time.time()
    for _ in range(size):
        sum_ = computeSum(size)
    takenTime3 = time.time() - start3
    print("computeSum Taken Time: {}".format(takenTime3))

if __name__ == "__main__":
    test()

computeSum_while Taken Time: 8.544030666351318
computeSum_for Taken Time: 4.945058107376099
computeSum Taken Time: 1.5659689903259277

6.2 减少内层for循环的计算

代码中尽量少用内循环,如下面函数 for_in 中 sqrt(x)位于内侧for循环, 每次训练过程中都会重新计算一次,增加了时间开销。

# -*- coding :utf-8 -*-

import time
import math

def for_in():
    """
        不推荐
    """
    start1 = time.time()
    size = 10000
    sqrt = math.sqrt
    for x in range(size):
        for y in range(size):
            z = sqrt(x) + sqrt(y)

    takenTime1 = time.time() - start1
    print("for_in Taken Time: {}".format(takenTime1))

def for_out():
    size = 10000
    sqrt = math.sqrt
    start2 = time.time()
    for x in range(size):
        sqrt_x = sqrt(x)  # 减少内层 for 循环的计算
        for y in range(size):
            z = sqrt_x + sqrt(y)

    takenTime2 = time.time() - start2
    print("for_out Taken Time: {}".format(takenTime2))

if __name__ == "__main__":
    for_in()
    for_out()

for_in Taken Time: 19.739089250564575
for_out Taken Time: 11.243038654327393

7. 使用numba.jit

numba可以将 Python 函数 JIT 编译为机器码执行,大大提高代码运行速度。关于numba的更多信息见下面的主页:http://numba.pydata.org/

# -*- coding :utf-8 -*-

import time
import numba

ef computeSum_while(size: int) -> int:
    """
        不推荐
    """
    sum_ = 0
    i = 0
    while i < size:
        sum_ += i
        i += 1
    return sum_

def computeSum_for(size: int) -> int:
    sum_ = 0
    for i in range(size):  # for 循环代替 while 循环
        sum_ += i
    return sum_

def computeSum(size: int) -> int:
    return sum(range(size))  # 隐式 for 循环代替显式 for 循环

@numba.jit
def computeSum_jit(size: float) -> int:
    sum = 0
    for i in range(size):
        sum += i
    return sum

def test():
    size = 10000

    start1 = time.time()
    for _ in range(size):
        sum_ = computeSum_while(size)
    takenTime1 = time.time() - start1
    print("computeSum_while Taken Time: {}".format(takenTime1))

    start2 = time.time()
    for _ in range(size):
        sum_ = computeSum_for(size)
    takenTime2 = time.time() - start2
    print("computeSum_for Taken Time: {}".format(takenTime2))

    start3 = time.time()
    for _ in range(size):
        sum_ = computeSum(size)
    takenTime3 = time.time() - start3
    print("computeSum Taken Time: {}".format(takenTime3))

    start4 = time.time()
    for _ in range(size):
        sum_ = computeSum_jit(size)
    takenTime4 = time.time() - start4
    print("computeSum_jit Taken Time: {}".format(takenTime4))

if __name__ == "__main__":
    test()

computeSum_while Taken Time: 8.01906418800354
computeSum_for Taken Time: 4.3460166454315186
computeSum Taken Time: 1.5610063076019287
computeSum_jit Taken Time: 0.4790055751800537

8. 选择合适的数据结构

Python 内置的数据结构如str, tuple, list, set, dict底层都是 C 实现的,速度非常快,自己实现新的数据结构想在性能上达到内置的速度几乎是不可能的。

list类似于 C++ 中的std::vector,是一种动态数组。其会预分配一定内存空间,当预分配的内存空间用完,又继续向其中添加元素时,会申请一块更大的内存空间,然后将原有的所有元素都复制过去,之后销毁之前的内存空间,再插入新元素。删除元素时操作类似,当已使用内存空间比预分配内存空间的一半还少时,会另外申请一块小内存,做一次元素复制,之后销毁原有大内存空间。因此,如果有频繁的新增、删除操作,新增、删除的元素数量又很多时,list的效率不高。此时,应该考虑使用 collections.dequecollections.deque是双端队列,同时具备栈和队列的特性,能够在两端进行 O ( 1 ) O(1) O(1) 复杂度的插入和删除操作。

list 的查找操作也非常耗时。当需要在list频繁查找某些元素,或频繁有序访问这些元素时,可以使用bisect维护list对象有序并在其中进行二分查找,提升查找的效率。

另外一个常见需求是查找极小值或极大值,此时可以使用heapq模块list转化为一个堆,使得获取最小值的时间复杂度是 O ( 1 ) O(1) O(1)

下面的网页给出了常用的 Python 数据结构的各项操作的时间复杂度:https://wiki.python.org/moin/TimeComplexity

  • list 操作:
    在这里插入图片描述
  • deque操作:
    在这里插入图片描述
  • set操作: 在这里插入图片描述
  • dict操作:
    在这里插入图片描述

参考文章:

  1. Python加速运行技巧
  2. David Beazley & Brian K. Jones. Python Cookbook, Third edition. O’Reilly Media, ISBN: 9781449340377, 2013
  3. 张颖 & 赖勇浩. 编写高质量代码:改善Python程序的91个建议. 机械工业出版社, ISBN: 9787111467045, 2014.
  4. 常用的 Python 数据结构的各项操作的时间复杂度
  5. numba 主页
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值