Python多进程池（Multiprocessing pool）如何使用多个参数和空函数？

Question

我试图在一个不返回任何内容的void函数上使用Python多进程库，并传递多个参数。以下是我的最小工作示例。

import numpy as np
from multiprocessing import Pool

dim1 = 2
dim2 = 2

test1 = np.zeros((dim1,dim2))
test2 = np.zeros((dim1,dim2))

iteration = []
for i in range(0,dim1):
    for j in range(0,dim2):
        iteration.append((i,j))
        
def testing(num1,num2):
    test1[num1,num2] = 1
    test2[num1,num2] = 2
    
if __name__ == '__main__':
    pool = Pool(processes=4)  
    pool.starmap(testing, iteration)
    
print(test1)
print(test2)

问题在于变量test1和test2在初始化时打印了一个零数组。相反，我想要test1是由1组成的数组，test2是由2组成的数组。我希望代码可以实现这一点。

if __name__ == '__main__':
    pool = Pool(processes=4)  
    pool.starmap(testing, iteration)

需要做的是:

testing(0,0)
testing(1,0)
testing(0,1)
testing(1,1)

我看过一些相关的帖子，比如这篇。这篇文章和我的区别在于，我的函数是一个空函数，而且我想让函数改变变量的值，而不是返回变量。

- Hosea

Answer 1

要在多个进程中使用全局数组更新数组而不返回结果：

• 使用multiprocessing.Array类存储数组数据。
• 创建进程池时使用initializer参数将数组传递给进程。

请注意，Array是一维的，因此必须对其进行重塑以进行更新和显示。

尝试这段代码：

import numpy as np
from multiprocessing import Pool, Array

dim1 = 2
dim2 = 2

def init(tt1,tt2):  # receive shared arrays
   global test1,test2
   test1,test2 = tt1,tt2

def testing(num1,num2):
    t1 = np.frombuffer(test1.get_obj()).reshape((dim1, dim2))  # need to reshape to 2D array
    t2 = np.frombuffer(test2.get_obj()).reshape((dim1, dim2))
    t1[num1,num2] = 1
    t2[num1,num2] = 2
   
if __name__ == '__main__':
    tt1 = Array('d', dim1*dim2)  # 1 dimensional arrays
    tt2 = Array('d', dim1*dim2)

    iteration = []
    for i in range(0,dim1):
        for j in range(0,dim2):
            iteration.append((i,j))
            
    pool = Pool(processes=4, initializer=init, initargs=(tt1,tt2))   # pass shared arrays to processes
    pool.starmap(testing, iteration)
    
    # still have access to the shared arrays
    t1final = np.frombuffer(tt1.get_obj()).reshape((dim1, dim2))
    t2final = np.frombuffer(tt2.get_obj()).reshape((dim1, dim2))
    print(t1final, t2final, sep='\n')

输出

[[1. 1.]
 [1. 1.]]
[[2. 2.]
 [2. 2.]]

- Mike67