A pickleable wrapper for sharing NumPy ndarrays between processes using POSIX shared memory.
SharedNDArrays are designed to be sent over multiprocessing.Pipe and Queue
without serializing or transmitting the underlying ndarray or buffer. While the
associated file descriptor is closed when the SharedNDArray is garbage
collected, the underlying buffer is not released when the process ends: you
must manually call the unlink()
method from the last process to use it.
from __future__ import print_function
import multiprocessing as mp
import numpy as np
from shared_ndarray import SharedNDArray
try:
shm = SharedNDArray((4, 4))
shm.array[0, 0] = 1
p = mp.Process(target=lambda shm: print(shm.array), args=(shm,))
p.start()
p.join()
finally:
shm.unlink()
This should print:
[[ 1. 0. 0. 0.] [ 0. 0. 0. 0.] [ 0. 0. 0. 0.] [ 0. 0. 0. 0.]]
There are also convenience methods to create a new SharedNDArray from an existing NumPy array:
arr = np.array([0, 0])
shm1 = SharedNDArray.copy(arr)
shm2 = SharedNDArray.zeros_like(arr)
shm1.unlink()
shm2.unlink()