python

multiprocessing

python-multiprocessing

In the Python multiprocessing library, is there a variant of pool.map which supports multiple arguments?

import multiprocessing

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()

Solution 1

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

Python 3.3 includes pool.starmap() method:

#!/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()

For older versions:

#!/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()

Output

1 1
2 1
3 1

Notice how itertools.izip() and itertools.repeat() are used here.

Due to the bug mentioned by @unutbu you can't use functools.partial() or similar capabilities on Python 2.6, so the simple wrapper function func_star() should be defined explicitly. See also the workaround suggested by uptimebox.

Solution 2

The answer to this is version- and situation-dependent. The most general answer for recent versions of Python (since 3.3) was first described below by J.F. Sebastian.1 It uses the Pool.starmap method, which accepts a sequence of argument tuples. It then automatically unpacks the arguments from each tuple and passes them to the given function:

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', ...

For earlier versions of Python, you'll need to write a helper function to unpack the arguments explicitly. If you want to use with, you'll also need to write a wrapper to turn Pool into a context manager. (Thanks to muon for pointing this out.)

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', ...

In simpler cases, with a fixed second argument, you can also use partial, but only in Python 2.7+.

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. Much of this was inspired by his answer, which should probably have been accepted instead. But since this one is stuck at the top, it seemed best to improve it for future readers.

Solution 3

I think the below will be better:

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

Output

[3, 5, 7]

Solution 4

Using Python 3.3+ with pool.starmap():

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()

Result:

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

You can also zip() more arguments if you like: zip(a,b,c,d,e)

In case you want to have a constant value passed as an argument:

import itertools

zip(itertools.repeat(constant), a)

In case your function should return something:

results = pool.starmap(write, zip(a,b))

This gives a List with the returned values.

Solution 5

How to take multiple arguments:

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)

Solution 6

Having learnt about itertools in J.F. Sebastian's answer I decided to take it a step further and write a parmap package that takes care about parallelization, offering map and starmap functions in Python 2.7 and Python 3.2 (and later also) that can take any number of positional arguments.

Installation

pip install parmap

How to parallelize:

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)

I have uploaded parmap to PyPI and to a GitHub repository.

As an example, the question can be answered as follows:

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)

Solution 7

There's a fork of multiprocessing called pathos (note: use the version on GitHub) that doesn't need starmap -- the map functions mirror the API for Python's map, thus map can take multiple arguments.

With pathos, you can also generally do multiprocessing in the interpreter, instead of being stuck in the __main__ block. Pathos is due for a release, after some mild updating -- mostly conversion to Python 3.x.

  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]

pathos has several ways that that you can get the exact behavior of starmap.

>>> def add(*x):
...   return sum(x)
...
>>> x = [[1,2,3],[4,5,6]]
>>> import pathos
>>> import numpy as np
>>> # use ProcessPool's map and transposing the inputs
>>> pp = pathos.pools.ProcessPool()
>>> pp.map(add, *np.array(x).T)
[6, 15]
>>> # use ProcessPool's map and a lambda to apply the star
>>> pp.map(lambda x: add(*x), x)
[6, 15]
>>> # use a _ProcessPool, which has starmap
>>> _pp = pathos.pools._ProcessPool()
>>> _pp.starmap(add, x)
[6, 15]
>>>

Solution 8

A better solution for Python 2:

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))])

Output

2 3 4

1 2 3

0 1 2

out[]:

[3, 5, 7]

Solution 9

Another way is to pass a list of lists to a one-argument routine:

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]
    ])

One can then construct a list lists of arguments with one's favorite method.

Solution 10

A better way is using a decorator instead of writing a wrapper function by hand. Especially when you have a lot of functions to map, a decorator will save your time by avoiding writing a wrapper for every function. Usually a decorated function is not picklable, however we may use functools to get around it. More discussions can be found here.

Here is the example:

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

Then you may map it with zipped arguments:

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

Of course, you may always use Pool.starmap in Python 3 (>=3.3) as mentioned in other answers.

Solution 11

You can use the following two functions so as to avoid writing a wrapper for each new function:

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))

Use the function function with the lists of arguments arg_0, arg_1 and arg_2 as follows:

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

Solution 12

Another simple alternative is to wrap your function parameters in a tuple and then wrap the parameters that should be passed in tuples as well. This is perhaps not ideal when dealing with large pieces of data. I believe it would make copies for each tuple.

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()

Gives the output in some random order:

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]

Solution 13

Here is another way to do it that IMHO is more simple and elegant than any of the other answers provided.

This program has a function that takes two parameters, prints them out and also prints the sum:

import multiprocessing

def main():

    with multiprocessing.Pool(10) as pool:
        params = [ (2, 2), (3, 3), (4, 4) ]
        pool.starmap(printSum, params)
    # end with

# end function

def printSum(num1, num2):
    mySum = num1 + num2
    print('num1 = ' + str(num1) + ', num2 = ' + str(num2) + ', sum = ' + str(mySum))
# end function

if __name__ == '__main__':
    main()

output is:

num1 = 2, num2 = 2, sum = 4
num1 = 3, num2 = 3, sum = 6
num1 = 4, num2 = 4, sum = 8

See the python docs for more info:

https://docs.python.org/3/library/multiprocessing.html#module-multiprocessing.pool

In particular be sure to check out the starmap function.

I'm using Python 3.6, I'm not sure if this will work with older Python versions

Why there is not a very straight-forward example like this in the docs, I'm not sure.

Solution 14

From Python 3.4.4, you can use multiprocessing.get_context() to obtain a context object to use multiple start methods:

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()

Or you just simply replace

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

with:

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

Solution 15

In the official documentation states that it supports only one iterable argument. I like to use apply_async in such cases. In your case I would do:

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()

Solution 16

There are many answers here, but none seem to provide Python 2/3 compatible code that will work on any version. If you want your code to just work, this will work for either Python version:

# 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

After that, you can use multiprocessing the regular Python 3 way, however you like. For example:

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)

will work in Python 2 or Python 3.

Solution 17

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()

Solution 18

This is an example of the routine I use to pass multiple arguments to a one-argument function used in a pool.imap fork:

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()

Solution 19

This might be another option. The trick is in the wrapper function that returns another function which is passed in to pool.map. The code below reads an input array and for each (unique) element in it, returns how many times (ie counts) that element appears in the array, For example if the input is

np.eye(3) = [ [1. 0. 0.]
              [0. 1. 0.]
              [0. 0. 1.]]

then zero appears 6 times and one 3 times

import numpy as np
from multiprocessing.dummy import Pool as ThreadPool
from multiprocessing import cpu_count


def extract_counts(label_array):
    labels = np.unique(label_array)
    out = extract_counts_helper([label_array], labels)
    return out

def extract_counts_helper(args, labels):
    n = max(1, cpu_count() - 1)
    pool = ThreadPool(n)
    results = {}
    pool.map(wrapper(args, results), labels)
    pool.close()
    pool.join()
    return results

def wrapper(argsin, results):
    def inner_fun(label):
        label_array = argsin[0]
        counts = get_label_counts(label_array, label)
        results[label] = counts
    return inner_fun

def get_label_counts(label_array, label):
    return sum(label_array.flatten() == label)

if __name__ == "__main__":
    img = np.ones([2,2])
    out = extract_counts(img)
    print('input array: \n', img)
    print('label counts: ', out)
    print("========")
           
    img = np.eye(3)
    out = extract_counts(img)
    print('input array: \n', img)
    print('label counts: ', out)
    print("========")
    
    img = np.random.randint(5, size=(3, 3))
    out = extract_counts(img)
    print('input array: \n', img)
    print('label counts: ', out)
    print("========")

You should get:

input array: 
 [[1. 1.]
 [1. 1.]]
label counts:  {1.0: 4}
========
input array: 
 [[1. 0. 0.]
 [0. 1. 0.]
 [0. 0. 1.]]
label counts:  {0.0: 6, 1.0: 3}
========
input array: 
 [[4 4 0]
 [2 4 3]
 [2 3 1]]
label counts:  {0: 1, 1: 1, 2: 2, 3: 2, 4: 3}
========

Solution 20

import time
from multiprocessing import Pool


def f1(args):
    vfirst, vsecond, vthird = args[0] , args[1] , args[2]
    print(f'First Param: {vfirst}, Second value: {vsecond} and finally third value is: {vthird}')
    pass


if __name__ == '__main__':
    p = Pool()
    result = p.map(f1, [['Dog','Cat','Mouse']])
    p.close()
    p.join()
    print(result)

Solution 21

Store all your arguments as an array of tuples.

The example says normally you call your function as:

def mainImage(fragCoord: vec2, iResolution: vec3, iTime: float) -> vec3:

Instead pass one tuple and unpack the arguments:

def mainImage(package_iter) -> vec3:
    fragCoord = package_iter[0]
    iResolution = package_iter[1]
    iTime = package_iter[2]

Build up the tuple by using a loop beforehand:

package_iter = []
iResolution = vec3(nx, ny, 1)
for j in range((ny-1), -1, -1):
    for i in range(0, nx, 1):
        fragCoord: vec2 = vec2(i, j)
        time_elapsed_seconds = 10
        package_iter.append((fragCoord, iResolution, time_elapsed_seconds))

Then execute all using map by passing the array of tuples:

array_rgb_values = []

with concurrent.futures.ProcessPoolExecutor() as executor:
    for val in executor.map(mainImage, package_iter):
        fragColor = val
        ir = clip(int(255* fragColor.r), 0, 255)
        ig = clip(int(255* fragColor.g), 0, 255)
        ib = clip(int(255* fragColor.b), 0, 255)

        array_rgb_values.append((ir, ig, ib))

I know Python has * and ** for unpacking, but I haven't tried those yet.

Also better to use the higher-level library concurrent futures than the low level multiprocessing library.

Solution 22

For Python 2, you can use this trick

def fun(a, b):
    return a + b

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