为什么我可以将实例方法传递给multiprocessing.Process，而不是multiprocessing.Pool？

Question

我正在尝试编写一个应用程序，该应用程序将函数与multiprocessing.Pool同时应用。我希望这个函数是一个实例方法（所以我可以在不同的子类中定义它）。这似乎是不可能的；正如我在别处学到的，显然bound methods can't be pickled。那么，为什么以绑定方法作为目标启动multiprocessing.Process？以下代码：

import multiprocessing

def test1():
    print "Hello, world 1"

def increment(x):
    return x + 1

class testClass():
    def process(self):
        process1 = multiprocessing.Process(target=test1)
        process1.start()
        process1.join()
        process2 = multiprocessing.Process(target=self.test2)
        process2.start()
        process2.join()

    def pool(self):
        pool = multiprocessing.Pool(1)
        for answer in pool.imap(increment, range(10)):
            print answer
        print
        for answer in pool.imap(self.square, range(10)):
            print answer

    def test2(self):
        print "Hello, world 2"

    def square(self, x):
        return x * x

def main():
    c = testClass()
    c.process()
    c.pool()

if __name__ == "__main__":
    main()

生成此输出：

Hello, world 1
Hello, world 2
1
2
3
4
5
6
7
8
9
10

Exception in thread Thread-2:
Traceback (most recent call last):
  File "C:\Python27\Lib\threading.py", line 551, in __bootstrap_inner
    self.run()
  File "C:\Python27\Lib\threading.py", line 504, in run
    self.__target(*self.__args, **self.__kwargs)
  File "C:\Python27\Lib\multiprocessing\pool.py", line 319, in _handle_tasks
    put(task)
PicklingError: Can't pickle <type 'instancemethod'>: attribute lookup __builtin__.instancemethod failed

为什么进程可以处理绑定的方法，而不能处理池？

Answer 1

这里有一个我有时使用的替代方法，它在Python2.x中有效：

您可以创建一个顶级的“别名”，按实例方法排序，该别名接受要在池中运行其实例方法的对象，并让该对象为您调用实例方法：

import functools
import multiprocessing

def _instance_method_alias(obj, arg):
    """
    Alias for instance method that allows the method to be called in a 
    multiprocessing pool
    """
    obj.instance_method(arg)
    return

class MyClass(object):
    """
    Our custom class whose instance methods we want to be able to use in a 
    multiprocessing pool
    """

    def __init__(self):
        self.my_string = "From MyClass: {}"

    def instance_method(self, arg):
        """
        Some arbitrary instance method
        """

        print(self.my_string.format(arg))
        return

# create an object of MyClass
obj = MyClass()

# use functools.partial to create a new method that always has the 
# MyClass object passed as its first argument
_bound_instance_method_alias = functools.partial(_instance_method_alias, obj)

# create our list of things we will use the pool to map
l = [1,2,3]

# create the pool of workers
pool = multiprocessing.Pool()

# call pool.map, passing it the newly created function
pool.map(_bound_instance_method_alias, l)

# cleanup
pool.close()
pool.join()

此代码生成此输出：

From MyClass: 1
From MyClass: 2
From MyClass: 3

一个限制是，不能将其用于修改对象的方法。每个进程都会获得它正在调用其方法的对象的副本，因此更改不会传播回主进程。如果不需要从调用的方法修改对象，这可能是一个简单的解决方案。

Answer 2

pickle模块通常无法pickle实例方法：

>>> import pickle
>>> class A(object):
...  def z(self): print "hi"
... 
>>> a = A()
>>> pickle.dumps(a.z)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python2.7/pickle.py", line 1374, in dumps
    Pickler(file, protocol).dump(obj)
  File "/usr/local/lib/python2.7/pickle.py", line 224, in dump
    self.save(obj)
  File "/usr/local/lib/python2.7/pickle.py", line 306, in save
    rv = reduce(self.proto)
  File "/usr/local/lib/python2.7/copy_reg.py", line 70, in _reduce_ex
    raise TypeError, "can't pickle %s objects" % base.__name__
TypeError: can't pickle instancemethod objects

但是，multiprocessing模块has a custom ^{} that adds some code to enable this feature：

#
# Try making some callable types picklable
#

from pickle import Pickler
class ForkingPickler(Pickler):
    dispatch = Pickler.dispatch.copy()

    @classmethod
    def register(cls, type, reduce):
        def dispatcher(self, obj):
            rv = reduce(obj)
            self.save_reduce(obj=obj, *rv)
        cls.dispatch[type] = dispatcher

def _reduce_method(m):
    if m.im_self is None:
        return getattr, (m.im_class, m.im_func.func_name)
    else:
        return getattr, (m.im_self, m.im_func.func_name)
ForkingPickler.register(type(ForkingPickler.save), _reduce_method)

您可以使用^{}模块复制它，以查看它是否适合您自己：

>>> import copy_reg
>>> def _reduce_method(m):
...     if m.im_self is None:
...         return getattr, (m.im_class, m.im_func.func_name)
...     else:
...         return getattr, (m.im_self, m.im_func.func_name)
... 
>>> copy_reg.pickle(type(a.z), _reduce_method)
>>> pickle.dumps(a.z)
"c__builtin__\ngetattr\np0\n(ccopy_reg\n_reconstructor\np1\n(c__main__\nA\np2\nc__builtin__\nobject\np3\nNtp4\nRp5\nS'z'\np6\ntp7\nRp8\n."

使用Process.start在Windows上生成新进程时，it pickles all the parameters you passed to the child process using this custom ^{}：

#
# Windows
#

else:
    # snip...
    from pickle import load, HIGHEST_PROTOCOL

    def dump(obj, file, protocol=None):
        ForkingPickler(file, protocol).dump(obj)

    #
    # We define a Popen class similar to the one from subprocess, but
    # whose constructor takes a process object as its argument.
    #

    class Popen(object):
        '''
        Start a subprocess to run the code of a process object
        '''
        _tls = thread._local()

        def __init__(self, process_obj):
            # create pipe for communication with child
            rfd, wfd = os.pipe()

            # get handle for read end of the pipe and make it inheritable
            ...
            # start process
            ...

            # set attributes of self
            ...

            # send information to child
            prep_data = get_preparation_data(process_obj._name)
            to_child = os.fdopen(wfd, 'wb')
            Popen._tls.process_handle = int(hp)
            try:
                dump(prep_data, to_child, HIGHEST_PROTOCOL)
                dump(process_obj, to_child, HIGHEST_PROTOCOL)
            finally:
                del Popen._tls.process_handle
                to_child.close()

注意“发送信息给孩子”部分。它使用dump函数，它使用ForkingPickler来pickle数据，这意味着您的实例方法可以被pickle。

现在，当您使用multiprocessing.Pool上的方法向子进程发送方法时，它使用multiprocessing.Pipe来pickle数据。在Python 2.7中，multiprocessing.Pipe是在C，and calls ^{} directly中实现的，因此它不利用ForkingPickler。这意味着pickling实例方法不起作用。

但是，如果使用copy_reg来注册instancemethod类型，而不是自定义的Pickler，所有酸洗尝试都将受到影响。因此，您可以使用它来启用pickling实例方法，甚至可以通过Pool：

import multiprocessing
import copy_reg
import types

def _reduce_method(m):
    if m.im_self is None:
        return getattr, (m.im_class, m.im_func.func_name)
    else:
        return getattr, (m.im_self, m.im_func.func_name)
copy_reg.pickle(types.MethodType, _reduce_method)

def test1():
    print("Hello, world 1")

def increment(x):
    return x + 1

class testClass():
    def process(self):
        process1 = multiprocessing.Process(target=test1)
        process1.start()
        process1.join()
        process2 = multiprocessing.Process(target=self.test2)
        process2.start()
        process2.join()

    def pool(self):
        pool = multiprocessing.Pool(1)
        for answer in pool.imap(increment, range(10)):
            print(answer)
        print
        for answer in pool.imap(self.square, range(10)):
            print(answer)

    def test2(self):
        print("Hello, world 2")

    def square(self, x):
        return x * x

def main():
    c = testClass()
    c.process()
    c.pool()

if __name__ == "__main__":
    main()

输出：

Hello, world 1
Hello, world 2
GOT (0, 0, (True, 1))
GOT (0, 1, (True, 2))
GOT (0, 2, (True, 3))
GOT (0, 3, (True, 4))
GOT (0, 4, (True, 5))
 1GOT (0, 5, (True, 6))

GOT (0, 6, (True, 7))
2
GOT (0, 7, (True, 8))
3
 GOT (0, 8, (True, 9))
GOT (0, 9, (True, 10))
4
5
6
7
8
9
10

GOT (1, 0, (True, 0))
0
GOT (1, 1, (True, 1))
1
GOT (1, 2, (True, 4))
4
GOT (1, 3, (True, 9))
9
 GOT (1, 4, (True, 16))
16
GOT (1, 5, (True, 25))
25
 GOT (1, 6, (True, 36))
36
 GOT (1, 7, (True, 49))
49
 GOT (1, 8, (True, 64))
64
GOT (1, 9, (True, 81))
81
GOT None

还要注意，在Python 3.x中，pickle可以本机pickle实例方法类型，因此这些都不再重要。：）

Answer 3

在Python 2中有一种更简单的工作方式，就是包装原始的实例方法。在MacOSX和Linux上运行良好，而在Windows上运行不好，测试了Python2.7

from multiprocessing import Pool

class Person(object):
    def __init__(self):
        self.name = 'Weizhong Tu'

    def calc(self, x):
        print self.name
        return x ** 5


def func(x, p=Person()):
    return p.calc(x)


pool = Pool()
print pool.map(func, range(10))