传入会话时，Tensorflow Multiprocessing pickle rlock 错误

Question

I created a very simple Tensorflow program to test out multiprocessing (shown below), but I keep getting the error:

TypeError: Can't pickle _thread.Rlock objects

I notice that my below program works perfectly if I don't pass in the tensorflow session, and instead just create it within the test_f2 function, so it must have something to do with passing in Tensorflow sessions in multiprocessing situations.

Does anyone know how I can fix this? Thanks so much!

import tensorflow as tf
import multiprocessing
def test_f2(x, sess):
    import tensorflow as tf
    a = tf.Variable(x, name='a')
    b = tf.Variable(100, name='b')
    c = tf.multiply(a, b, name='c')
    sess.run(tf.global_variables_initializer())
    out = sess.run(c)
    sess.close()
    print("OK: %s" % out)

if __name__ == '__main__':
    num_procs = 2
    procs_list = []
    for i in range(num_procs):
        new_session = tf.Session()
        p = multiprocessing.Process(target=test_f2, args=(i, new_session))
        p.daemon = True
        p.start()
        procs_list.append(p)
    for p in procs_list:
        p.join()

I'm using Tensorflow-GPU 1.14.0 and Python 3.7.4.

Answer 1

For starters, move all of your imports to the top of your file (there is still an import at the top of test_f2() .

But the bigger issue is the way this is trying to share a session across multiple different processes. The session is dependent on the memory addressing in the parent process, which the child processes cannot access ( multiprocessing is true fork and exec type process spawning, it is not threading and the children do not share memory space). You need to create the session in the child processes, it doesn't work to share the way you have it coded.