用于多个参数的python multiprocessing pool.map

Python multiprocessing pool.map for multiple arguments

在python多处理库中,是否有pool.map的变体支持多个参数?

1
2
3
4
5
6
7
8
9
10
11
text ="test"
def harvester(text, case):
    X = case[0]
    text+ str(X)

if __name__ == '__main__':
    pool = multiprocessing.Pool(processes=6)
    case = RAW_DATASET
    pool.map(harvester(text,case),case, 1)
    pool.close()
    pool.join()


is there a variant of pool.map which support multiple arguments?

python 3.3包括pool.starmap()方法:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
#!/usr/bin/env python3
from functools import partial
from itertools import repeat
from multiprocessing import Pool, freeze_support

def func(a, b):
    return a + b

def main():
    a_args = [1,2,3]
    second_arg = 1
    with Pool() as pool:
        L = pool.starmap(func, [(1, 1), (2, 1), (3, 1)])
        M = pool.starmap(func, zip(a_args, repeat(second_arg)))
        N = pool.map(partial(func, b=second_arg), a_args)
        assert L == M == N

if __name__=="__main__":
    freeze_support()
    main()

对于旧版本:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
#!/usr/bin/env python2
import itertools
from multiprocessing import Pool, freeze_support

def func(a, b):
    print a, b

def func_star(a_b):
   """Convert `f([1,2])` to `f(1,2)` call."""
    return func(*a_b)

def main():
    pool = Pool()
    a_args = [1,2,3]
    second_arg = 1
    pool.map(func_star, itertools.izip(a_args, itertools.repeat(second_arg)))

if __name__=="__main__":
    freeze_support()
    main()

产量

1
2
3
1 1
2 1
3 1

注意这里是如何使用itertools.izip()itertools.repeat()的。

由于@unutbu提到的bug,您不能在python 2.6上使用functools.partial()或类似的功能,所以应该明确定义简单的包装函数func_star()。另见uptimebox建议的解决方法。


答案取决于版本和情况。J.F.Sebastian.1首先描述了最近版本的python(自3.3以来)最一般的答案,它使用了Pool.starmap方法,它接受一系列参数元组。然后,它自动解包每个元组中的参数,并将它们传递给给定的函数:

1
2
3
4
5
6
7
8
9
10
11
12
13
import multiprocessing
from itertools import product

def merge_names(a, b):
    return '{} & {}'.format(a, b)

if __name__ == '__main__':
    names = ['Brown', 'Wilson', 'Bartlett', 'Rivera', 'Molloy', 'Opie']
    with multiprocessing.Pool(processes=3) as pool:
        results = pool.starmap(merge_names, product(names, repeat=2))
    print(results)

# Output: ['Brown & Brown', 'Brown & Wilson', 'Brown & Bartlett', ...

对于早期版本的Python,您需要编写一个助手函数来显式地解包参数。如果您想使用with,还需要编写一个包装器来将Pool转换为上下文管理器。(感谢Muon指出这一点。)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
import multiprocessing
from itertools import product
from contextlib import contextmanager

def merge_names(a, b):
    return '{} & {}'.format(a, b)

def merge_names_unpack(args):
    return merge_names(*args)

@contextmanager
def poolcontext(*args, **kwargs):
    pool = multiprocessing.Pool(*args, **kwargs)
    yield pool
    pool.terminate()

if __name__ == '__main__':
    names = ['Brown', 'Wilson', 'Bartlett', 'Rivera', 'Molloy', 'Opie']
    with poolcontext(processes=3) as pool:
        results = pool.map(merge_names_unpack, product(names, repeat=2))
    print(results)

# Output: ['Brown & Brown', 'Brown & Wilson', 'Brown & Bartlett', ...

在更简单的情况下,使用固定的第二个参数,也可以使用partial,但只能在python 2.7+中使用。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
import multiprocessing
from functools import partial
from contextlib import contextmanager

@contextmanager
def poolcontext(*args, **kwargs):
    pool = multiprocessing.Pool(*args, **kwargs)
    yield pool
    pool.terminate()

def merge_names(a, b):
    return '{} & {}'.format(a, b)

if __name__ == '__main__':
    names = ['Brown', 'Wilson', 'Bartlett', 'Rivera', 'Molloy', 'Opie']
    with poolcontext(processes=3) as pool:
        results = pool.map(partial(merge_names, b='Sons'), names)
    print(results)

# Output: ['Brown & Sons', 'Wilson & Sons', 'Bartlett & Sons', ...

1。这其中的大部分灵感来自于他的回答,而他的回答本应被接受。不过,既然这本书一直位居榜首,似乎最好还是为将来的读者改进一下。


我想下面会更好

1
2
3
4
5
6
7
8
9
def multi_run_wrapper(args):
   return add(*args)
def add(x,y):
    return x+y
if __name__ =="__main__":
    from multiprocessing import Pool
    pool = Pool(4)
    results = pool.map(multi_run_wrapper,[(1,2),(2,3),(3,4)])
    print results

输出

1
[3, 5, 7]


pool.starmap():中使用python 3.3+

1
2
3
4
5
6
7
8
9
10
11
12
from multiprocessing.dummy import Pool as ThreadPool

def write(i, x):
    print(i,"---", x)

a = ["1","2","3"]
b = ["4","5","6"]

pool = ThreadPool(2)
pool.starmap(write, zip(a,b))
pool.close()
pool.join()

结果:

1
2
3
1 --- 4
2 --- 5
3 --- 6

如果您愿意,还可以压缩()更多参数:zip(a,b,c,d,e)

如果您希望将常量值作为参数传递,则必须使用import itertools,然后使用zip(itertools.repeat(constant), a)


在J.F.Sebastian Answer中了解了Itertools之后,我决定更进一步,编写一个负责并行化的parmap包,在python-2.7和python-3.2(以及以后的版本)上提供mapstarmap函数,这些函数可以接受任意数量的位置参数。

安装

1
pip install parmap

如何并行化:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
import parmap
# If you want to do:
y = [myfunction(x, argument1, argument2) for x in mylist]
# In parallel:
y = parmap.map(myfunction, mylist, argument1, argument2)

# If you want to do:
z = [myfunction(x, y, argument1, argument2) for (x,y) in mylist]
# In parallel:
z = parmap.starmap(myfunction, mylist, argument1, argument2)

# If you want to do:
listx = [1, 2, 3, 4, 5, 6]
listy = [2, 3, 4, 5, 6, 7]
param = 3.14
param2 = 42
listz = []
for (x, y) in zip(listx, listy):
        listz.append(myfunction(x, y, param1, param2))
# In parallel:
listz = parmap.starmap(myfunction, zip(listx, listy), param1, param2)

我已经将parmap上传到pypi和github存储库。

例如,问题的答案如下:

1
2
3
4
5
6
7
8
9
import parmap

def harvester(case, text):
    X = case[0]
    text+ str(X)

if __name__ =="__main__":
    case = RAW_DATASET  # assuming this is an iterable
    parmap.map(harvester, case,"test", chunksize=1)

multiprocessing的一个分支称为pathos(注意:使用github上的版本),它不需要starmap——map函数为python的map镜像API,因此map可以接受多个参数。使用pathos时,通常也可以在解释器中进行多处理,而不是卡在__main__块中。Pathos将在一些温和的更新之后发布——主要是转换为python 3.x。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
  Python 2.7.5 (default, Sep 30 2013, 20:15:49)
  [GCC 4.2.1 (Apple Inc. build 5566)] on darwin
  Type"help","copyright","credits" or"license" for more information.
  >>> def func(a,b):
  ...     print a,b
  ...
  >>>
  >>> from pathos.multiprocessing import ProcessingPool    
  >>> pool = ProcessingPool(nodes=4)
  >>> pool.map(func, [1,2,3], [1,1,1])
  1 1
  2 1
  3 1
  [None, None, None]
  >>>
  >>> # also can pickle stuff like lambdas
  >>> result = pool.map(lambda x: x**2, range(10))
  >>> result
  [0, 1, 4, 9, 16, 25, 36, 49, 64, 81]
  >>>
  >>> # also does asynchronous map
  >>> result = pool.amap(pow, [1,2,3], [4,5,6])
  >>> result.get()
  [1, 32, 729]
  >>>
  >>> # or can return a map iterator
  >>> result = pool.imap(pow, [1,2,3], [4,5,6])
  >>> result
  <processing.pool.IMapIterator object at 0x110c2ffd0>
  >>> list(result)
  [1, 32, 729]


可以使用以下两个函数,以避免为每个新函数编写包装器:

1
2
3
4
5
6
7
8
9
import itertools
from multiprocessing import Pool

def universal_worker(input_pair):
    function, args = input_pair
    return function(*args)

def pool_args(function, *args):
    return zip(itertools.repeat(function), zip(*args))

将函数function与参数arg_0arg_1arg_2列表一起使用,如下所示:

1
2
3
4
pool = Pool(n_core)
list_model = pool.map(universal_worker, pool_args(function, arg_0, arg_1, arg_2)
pool.close()
pool.join()

更好的方法是使用decorator而不是手工编写包装函数。特别是当您有很多函数需要映射时,decorator将避免为每个函数编写包装器,从而节省您的时间。通常装饰的功能是不可选择的,但是我们可以使用functools来绕过它。这里有更多的争论。

这里举个例子

1
2
3
4
5
6
7
8
9
10
11
12
13
def unpack_args(func):
    from functools import wraps
    @wraps(func)
    def wrapper(args):
        if isinstance(args, dict):
            return func(**args)
        else:
            return func(*args)
    return wrapper

@unpack_args
def func(x, y):
    return x + y

然后你可以用压缩参数映射它

1
2
3
4
5
np, xlist, ylist = 2, range(10), range(10)
pool = Pool(np)
res = pool.map(func, zip(xlist, ylist))
pool.close()
pool.join()

当然,您可以在python 3(>=3.3)中使用Pool.starmap,如其他答案中所述。


另一个简单的选择是将函数参数包装成一个元组,然后将应该传递给元组的参数包装起来。在处理大数据块时,这可能不理想。我相信它会为每个元组制作副本。

1
2
3
4
5
6
7
8
9
10
11
12
from multiprocessing import Pool

def f((a,b,c,d)):
    print a,b,c,d
    return a + b + c +d

if __name__ == '__main__':
    p = Pool(10)
    data = [(i+0,i+1,i+2,i+3) for i in xrange(10)]
    print(p.map(f, data))
    p.close()
    p.join()

以随机顺序给出输出:

1
2
3
4
5
6
7
8
9
10
11
0 1 2 3
1 2 3 4
2 3 4 5
3 4 5 6
4 5 6 7
5 6 7 8
7 8 9 10
6 7 8 9
8 9 10 11
9 10 11 12
[6, 10, 14, 18, 22, 26, 30, 34, 38, 42]


一种更好的治疗Python2的方法:

1
2
3
4
5
6
from multiprocessing import Pool
def func((i, (a, b))):
    print i, a, b
    return a + b
pool = Pool(3)
pool.map(func, [(0,(1,2)), (1,(2,3)), (2,(3, 4))])

2 3 3

1 2 2

0 1 1

出[ ]:

〔3, 5, 7〕


#"如何接受多个参数"。

1
2
3
4
5
6
7
8
9
10
def f1(args):
    a, b, c = args[0] , args[1] , args[2]
    return a+b+c

if __name__ =="__main__":
    import multiprocessing
    pool = multiprocessing.Pool(4)

    result1 = pool.map(f1, [ [1,2,3] ])
    print(result1)


另一种方法是将列表传递给一个参数例程:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
import os
from multiprocessing import Pool

def task(args):
    print"PID =", os.getpid(),", arg1 =", args[0],", arg2 =", args[1]

pool = Pool()

pool.map(task, [
        [1,2],
        [3,4],
        [5,6],
        [7,8]
    ])

我们可以用自己喜欢的方法构造参数列表。


从python 3.4.4中,可以使用multiprocessing.get_context()获取上下文对象,以使用多个start方法:

1
2
3
4
5
6
7
8
9
10
11
12
13
import multiprocessing as mp

def foo(q, h, w):
    q.put(h + ' ' + w)
    print(h + ' ' + w)

if __name__ == '__main__':
    ctx = mp.get_context('spawn')
    q = ctx.Queue()
    p = ctx.Process(target=foo, args=(q,'hello', 'world'))
    p.start()
    print(q.get())
    p.join()

或者你只是简单地替换

1
pool.map(harvester(text,case),case, 1)

通过:

1
pool.apply_async(harvester(text,case),case, 1)

这里有很多答案,但似乎没有一个提供可以在任何版本上工作的python2/3兼容代码。如果您希望您的代码能够正常工作,这对于任何一个Python版本都是有效的:

1
2
3
4
5
6
7
8
9
10
11
# For python 2/3 compatibility, define pool context manager
# to support the 'with' statement in Python 2
if sys.version_info[0] == 2:
    from contextlib import contextmanager
    @contextmanager
    def multiprocessing_context(*args, **kwargs):
        pool = multiprocessing.Pool(*args, **kwargs)
        yield pool
        pool.terminate()
else:
    multiprocessing_context = multiprocessing.Pool

之后,您可以使用多处理常规的python3方式,不管您喜欢什么。例如:

1
2
3
4
def _function_to_run_for_each(x):
       return x.lower()
with multiprocessing_context(processes=3) as pool:
    results = pool.map(_function_to_run_for_each, ['Bob', 'Sue', 'Tim'])    print(results)

将在python2或python3中工作。


在官方文件中,它只支持一个不可辩驳的论点。我喜欢在这种情况下使用Apply_Async。在你的情况下,我会这样做:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
from multiprocessing import Process, Pool, Manager

text ="test"
def harvester(text, case, q = None):
 X = case[0]
 res = text+ str(X)
 if q:
  q.put(res)
 return res


def block_until(q, results_queue, until_counter=0):
 i = 0
 while i < until_counter:
  results_queue.put(q.get())
  i+=1

if __name__ == '__main__':
 pool = multiprocessing.Pool(processes=6)
 case = RAW_DATASET
 m = Manager()
 q = m.Queue()
 results_queue = m.Queue() # when it completes results will reside in this queue
 blocking_process = Process(block_until, (q, results_queue, len(case)))
 blocking_process.start()
 for c in case:
  try:
   res = pool.apply_async(harvester, (text, case, q = None))
   res.get(timeout=0.1)
  except:
   pass
 blocking_process.join()

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
text ="test"

def unpack(args):
    return args[0](*args[1:])

def harvester(text, case):
    X = case[0]
    text+ str(X)

if __name__ == '__main__':
    pool = multiprocessing.Pool(processes=6)
    case = RAW_DATASET
    # args is a list of tuples
    # with the function to execute as the first item in each tuple
    args = [(harvester, text, c) for c in case]
    # doing it this way, we can pass any function
    # and we don't need to define a wrapper for each different function
    # if we need to use more than one
    pool.map(unpack, args)
    pool.close()
    pool.join()


这是我用于将多个参数传递给pool.imap fork中使用的单参数函数的例程示例:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
from multiprocessing import Pool

# Wrapper of the function to map:
class makefun:
    def __init__(self, var2):
        self.var2 = var2
    def fun(self, i):
        var2 = self.var2
        return var1[i] + var2

# Couple of variables for the example:
var1 = [1, 2, 3, 5, 6, 7, 8]
var2 = [9, 10, 11, 12]

# Open the pool:
pool = Pool(processes=2)

# Wrapper loop
for j in range(len(var2)):
    # Obtain the function to map
    pool_fun = makefun(var2[j]).fun

    # Fork loop
    for i, value in enumerate(pool.imap(pool_fun, range(len(var1))), 0):
        print(var1[i], '+' ,var2[j], '=', value)

# Close the pool
pool.close()

对于python2,你可以使用这个技巧

1
2
3
4
5
6
def fun(a,b):
    return a+b

pool = multiprocessing.Pool(processes=6)
b=233
pool.map(lambda x:fun(x,b),range(1000))