joblib: 简单易懂的平行世界

Joblib是一款轻量级的Python工作流实现,可以简单快速地把相互独立的循环并行起来,从而节省时间。

安装

1
pip install joblib

使用方法

以打印出0-9的平方根为例,如果采用直接循环方法,我们可以很简单地写出如下代码。这里用 sleep 来模拟任务本身的延迟:

1
2
3
4
5
6
7
from math import sqrt
from time import sleep
from datetime import datetime

for i in range(10):
    print(datetime.now(), sqrt(i))
    sleep(1)

输出如下,可见是每秒吐出一个值:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
2023-06-15 11:02:01.200605 0.0
2023-06-15 11:02:02.203232 1.0
2023-06-15 11:02:03.207490 1.4142135623730951
2023-06-15 11:02:04.211349 1.7320508075688772
2023-06-15 11:02:05.211646 2.0
2023-06-15 11:02:06.215645 2.23606797749979
2023-06-15 11:02:07.218109 2.449489742783178
2023-06-15 11:02:08.220789 2.6457513110645907
2023-06-15 11:02:09.225884 2.8284271247461903
2023-06-15 11:02:10.229875 3.0

若利用 joblib.Parallel,将循环体写成函数,并把结果作为返回值,则可改写如下:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
from math import sqrt
from time import sleep
from datetime import datetime
from joblib import Parallel, delayed

def func(i):
    result = sqrt(i)
    print(datetime.now(), result)
    sleep(1)
    return result

Parallel(n_jobs=2)(delayed(func)(i) for i in range(10))

可以看到输出如下,由于 n_jobs 设置为2,故每次有2个worker同时运行,每秒得到两个输出:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
2023-06-15 11:07:45.843004 0.0
2023-06-15 11:07:45.873286 1.0
2023-06-15 11:07:46.864319 1.4142135623730951
2023-06-15 11:07:46.891164 1.7320508075688772
2023-06-15 11:07:47.871781 2.0
2023-06-15 11:07:47.896673 2.23606797749979
2023-06-15 11:07:48.875107 2.449489742783178
2023-06-15 11:07:48.902218 2.6457513110645907
2023-06-15 11:07:49.880168 2.8284271247461903
2023-06-15 11:07:49.907760 3.0

且最终返回值以一个列表形式列出:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
[0.0,
 1.0,
 1.4142135623730951,
 1.7320508075688772,
 2.0,
 2.23606797749979,
 2.449489742783178,
 2.6457513110645907,
 2.8284271247461903,
 3.0]

可见上述代码的效果等价于如下的列表生成式:

1
[func(i) for i in range(10)]

用这种方法,可以简单快捷地把彼此相互独立的循环体拆解成多个平行任务。当然 n_jobs 的设置请遵循自己使用环境的情况,避免影响到其他人的使用。

Licensed under CC BY-NC-SA 4.0
comments powered by Disqus