多处理池：如何在类对象列表上调用任意方法列表

Question

A cleaned up version of the code including the solution to the problem (thanks @JohanL!) can be found as a Gist on GitHub .

---

The following code snipped (CPython 3.[4,5,6]) illustrates my intention (as well as my problem):

from functools import partial
import multiprocessing
from pprint import pprint as pp

NUM_CORES = multiprocessing.cpu_count()

class some_class:
    some_dict = {'some_key': None, 'some_other_key': None}
    def some_routine(self):
        self.some_dict.update({'some_key': 'some_value'})
    def some_other_routine(self):
        self.some_dict.update({'some_other_key': 77})

def run_routines_on_objects_in_parallel_and_return(in_object_list, routine_list):
    func_handle = partial(__run_routines_on_object_and_return__, routine_list)
    with multiprocessing.Pool(processes = NUM_CORES) as p:
        out_object_list = list(p.imap_unordered(
            func_handle,
            (in_object for in_object in in_object_list)
            ))
    return out_object_list

def __run_routines_on_object_and_return__(routine_list, in_object):
    for routine_name in routine_list:
        getattr(in_object, routine_name)()
    return in_object

object_list = [some_class() for item in range(20)]
pp([item.some_dict for item in object_list])

new_object_list = run_routines_on_objects_in_parallel_and_return(
        object_list,
        ['some_routine', 'some_other_routine']
        )
pp([item.some_dict for item in new_object_list])

verification_object_list = [
    __run_routines_on_object_and_return__(
        ['some_routine', 'some_other_routine'],
        item
        ) for item in object_list
    ]
pp([item.some_dict for item in verification_object_list])

I am working with a list of objects of type some_class . some_class has a property, a dictionary, named some_dict and a few methods, which can modify the dict ( some_routine and some_other_routine ). Sometimes, I want to call a sequence of methods on all the objects in the list. Because this is computationally intensive, I intend to distribute the objects over multiple CPU cores (using multiprocessing.Pool and imap_unordered - the list order does not matter).

The routine __run_routines_on_object_and_return__ takes care of calling the list of methods on one individual object. From what I can tell, this is working just fine. I am using functools.partial for simplifying the structure of the code a bit - the multiprocessing pool therefore has to handle the list of objects as an input parameter only.

The problem is ... it does not work. The objects contained in the list returned by imap_unordered are identical to the objects I fed into it. The dictionaries within the objects look just like before. I have used similar mechanisms for working on lists of dictionaries directly without a glitch, so I somehow suspect that there is something wrong with modifying an object property which happens to be a dictionary.

In my example, verification_object_list contains the correct result (though it is generated in a single process/thread). new_object_list is identical to object_list , which should not be the case.

What am I doing wrong?

---

EDIT

I found the following question , which has an actually working and applicable answer . I modified it a bit following my idea of calling a list of methods on every object and it works:

import random
from multiprocessing import Pool, Manager

class Tester(object):
    def __init__(self, num=0.0, name='none'):
        self.num  = num
        self.name = name
    def modify_me(self):
        self.num += random.normalvariate(mu=0, sigma=1)
        self.name = 'pla' + str(int(self.num * 100))
    def __repr__(self):
        return '%s(%r, %r)' % (self.__class__.__name__, self.num, self.name)

def init(L):
    global tests
    tests = L

def modify(i_t_nn):
    i, t, nn = i_t_nn
    for method_name in nn:
        getattr(t, method_name)()
    tests[i] = t # copy back
    return i

def main():
    num_processes = num = 10 #note: num_processes and num may differ
    manager = Manager()
    tests = manager.list([Tester(num=i) for i in range(num)])
    print(tests[:2])

    args = ((i, t, ['modify_me']) for i, t in enumerate(tests))
    pool = Pool(processes=num_processes, initializer=init, initargs=(tests,))
    for i in pool.imap_unordered(modify, args):
        print("done %d" % i)
    pool.close()
    pool.join()
    print(tests[:2])

if __name__ == '__main__':
    main()

Now, I went a bit further and introduced my original some_class into the game, which contains a the described dictionary property some_dict . It does NOT work:

import random
from multiprocessing import Pool, Manager
from pprint import pformat as pf

class some_class:
    some_dict = {'some_key': None, 'some_other_key': None}
    def some_routine(self):
        self.some_dict.update({'some_key': 'some_value'})
    def some_other_routine(self):
        self.some_dict.update({'some_other_key': 77})
    def __repr__(self):
        return pf(self.some_dict)

def init(L):
    global tests
    tests = L

def modify(i_t_nn):
    i, t, nn = i_t_nn
    for method_name in nn:
        getattr(t, method_name)()
    tests[i] = t # copy back
    return i

def main():
    num_processes = num = 10 #note: num_processes and num may differ
    manager = Manager()
    tests = manager.list([some_class() for i in range(num)])
    print(tests[:2])

    args = ((i, t, ['some_routine', 'some_other_routine']) for i, t in enumerate(tests))
    pool = Pool(processes=num_processes, initializer=init, initargs=(tests,))
    for i in pool.imap_unordered(modify, args):
        print("done %d" % i)
    pool.close()
    pool.join()
    print(tests[:2])

if __name__ == '__main__':
    main()

The diff between working and not working is really small, but I still do not get it:

diff --git a/test.py b/test.py
index b12eb56..0aa6def 100644
--- a/test.py
+++ b/test.py
@@ -1,15 +1,15 @@
 import random
 from multiprocessing import Pool, Manager
+from pprint import pformat as pf

-class Tester(object):
-       def __init__(self, num=0.0, name='none'):
-               self.num  = num
-               self.name = name
-       def modify_me(self):
-               self.num += random.normalvariate(mu=0, sigma=1)
-               self.name = 'pla' + str(int(self.num * 100))
+class some_class:
+       some_dict = {'some_key': None, 'some_other_key': None}
+       def some_routine(self):
+               self.some_dict.update({'some_key': 'some_value'})
+       def some_other_routine(self):
+               self.some_dict.update({'some_other_key': 77})
        def __repr__(self):
-               return '%s(%r, %r)' % (self.__class__.__name__, self.num, self.name)
+               return pf(self.some_dict)

 def init(L):
        global tests
@@ -25,10 +25,10 @@ def modify(i_t_nn):
 def main():
        num_processes = num = 10 #note: num_processes and num may differ
        manager = Manager()
-       tests = manager.list([Tester(num=i) for i in range(num)])
+       tests = manager.list([some_class() for i in range(num)])
        print(tests[:2])

-       args = ((i, t, ['modify_me']) for i, t in enumerate(tests))
+       args = ((i, t, ['some_routine', 'some_other_routine']) for i, t in enumerate(tests))

What is happening here?

Answer 1

Your problem is due to two things; namely that you are using a class variable and that you are running your code in different processes.

Since different processes do not share memory, all objects and parameters must be pickled and sent from the original process to the process that executes it. When the parameter is an object, its class is not sent with it. Instead the receiving process uses its own blueprint (ie class ).

In your current code, you pass the object as a parameter, update it and return it. However, the updates are not made to the object, but rather to the class itself, since you are updating a class variable. However, this update is not sent back to your main process, and therefore you are left with your not updated class.

What you want to do, is to make some_dict a part of your object, rather than of your class. This is easily done by an __init__() method. Thus modify some_class as:

class some_class:
    def __init__(self):
        self.some_dict = {'some_key': None, 'some_other_key': None}
    def some_routine(self):
        self.some_dict.update({'some_key': 'some_value'})
    def some_other_routine(self):
        self.some_dict.update({'some_other_key': 77})

This will make your program work as you intend it to. You almost always want to setup your object in an __init__() call, rather than as class variables, since in the latter case the data will be shared between all instances (and can be updated by all). That is not normally what you want, when you encapsulate data and state in an object of a class.

EDIT: It seems I was mistaken in whether the class is sent with the pickled object. After further inspection of what happens, I think also the class itself, with its class variables are pickled. Since, if the class variable is updated before sending the object to the new process, the updated value is available. However it is still the case that the updates done in the new process are not relayed back to the original class .