如何将二维数组作为 multiprocessing.Array 传递给 multiprocessing.Pool？

Posted 2023-02-16

【中文标题】如何将二维数组作为 multiprocessing.Array 传递给 multiprocessing.Pool？

【英文标题】：How to pass 2d array as multiprocessing.Array to multiprocessing.Pool?

【发布时间】：2021-01-21 03:45:13

【问题描述】：

我的目标是将父数组传递给mp.Pool 并用2s 填充它，同时将其分配给不同的进程。这适用于一维数组：

import numpy as np
import multiprocessing as mp
import itertools


def worker_function(i=None):
    global arr
    val = 2
    arr[i] = val
    print(arr[:])


def init_arr(arr=None):
    globals()['arr'] = arr

def main():
    arr = mp.Array('i', np.zeros(5, dtype=int), lock=False)
    mp.Pool(1, initializer=init_arr, initargs=(arr,)).starmap(worker_function, zip(range(5)))
    print(arr[:])


if __name__ == '__main__':
    main()

输出：

[2, 0, 0, 0, 0]
[2, 2, 0, 0, 0]
[2, 2, 2, 0, 0]
[2, 2, 2, 2, 0]
[2, 2, 2, 2, 2]
[2, 2, 2, 2, 2]

但是我怎样才能对 x 维数组做同样的事情呢？给arr添加维度：

arr = mp.Array('i', np.zeros((5, 5), dtype=int), lock=False)

产生错误：

Traceback (most recent call last):
  File "C:/Users/Artur/Desktop/RL_framework/test2.py", line 23, in <module>
    main()
  File "C:/Users/Artur/Desktop/RL_framework/test2.py", line 17, in main
    arr = mp.Array('i', np.zeros((5, 5), dtype=int), lock=False)
  File "C:\Users\Artur\anaconda3\envs\RL_framework\lib\multiprocessing\context.py", line 141, in Array
    ctx=self.get_context())
  File "C:\Users\Artur\anaconda3\envs\RL_framework\lib\multiprocessing\sharedctypes.py", line 88, in Array
    obj = RawArray(typecode_or_type, size_or_initializer)
  File "C:\Users\Artur\anaconda3\envs\RL_framework\lib\multiprocessing\sharedctypes.py", line 67, in RawArray
    result.__init__(*size_or_initializer)
TypeError: only size-1 arrays can be converted to Python scalars

更改arr 的dtype 也无济于事。

【参考方案1】：

你不能直接将multiprocessing.Array用作二维数组，但在一维内存中，第二维无论如何只是一种错觉:)。

幸运的是，numpy 允许从 buffer 读取数组并对其进行整形而无需复制它。在下面的演示中，我只是使用了一个单独的锁，这样我们就可以逐步观察所做的更改，目前它正在做的事情没有竞争条件。

import multiprocessing as mp
import numpy as np    

def worker_function(i):
    global arr, arr_lock
    val = 2
    with arr_lock:
        arr[i, :i+1] = val
        print(f"mp.current_process().name\narr[:]")


def init_arr(arr, arr_lock=None):
    globals()['arr'] = np.frombuffer(arr, dtype='int32').reshape(5, 5)
    globals()['arr_lock'] = arr_lock


def main():
    arr = mp.Array('i', np.zeros(5 * 5, dtype='int32'), lock=False)
    arr_lock = mp.Lock()

    mp.Pool(2, initializer=init_arr, initargs=(arr, arr_lock)).map(
        worker_function, range(5)
    )

    arr = np.frombuffer(arr, dtype='int32').reshape(5, 5)
    print(f"mp.current_process().name\narr")


if __name__ == '__main__':
    main()

输出：

ForkPoolWorker-1
[[2 0 0 0 0]
 [0 0 0 0 0]
 [0 0 0 0 0]
 [0 0 0 0 0]
 [0 0 0 0 0]]
ForkPoolWorker-2
[[2 0 0 0 0]
 [2 2 0 0 0]
 [0 0 0 0 0]
 [0 0 0 0 0]
 [0 0 0 0 0]]
ForkPoolWorker-1
[[2 0 0 0 0]
 [2 2 0 0 0]
 [2 2 2 0 0]
 [0 0 0 0 0]
 [0 0 0 0 0]]
ForkPoolWorker-2
[[2 0 0 0 0]
 [2 2 0 0 0]
 [2 2 2 0 0]
 [2 2 2 2 0]
 [0 0 0 0 0]]
ForkPoolWorker-1
[[2 0 0 0 0]
 [2 2 0 0 0]
 [2 2 2 0 0]
 [2 2 2 2 0]
 [2 2 2 2 2]]
MainProcess
[[2 0 0 0 0]
 [2 2 0 0 0]
 [2 2 2 0 0]
 [2 2 2 2 0]
 [2 2 2 2 2]]

Process finished with exit code 0