multiprocessing.Pool.map() drops attribute of subclassed ndarray

问题

When using map() from multiprocessing.Pool() on a list of instances from a numpy.ndarray-subclass, the new attributes of the own class are dropped.

The following minimal example based on the numpy docs subclassing example reproduces the problem:

from multiprocessing import Pool
import numpy as np


class MyArray(np.ndarray):

    def __new__(cls, input_array, info=None):
        obj = np.asarray(input_array).view(cls)
        obj.info = info
        return obj

    def __array_finalize__(self, obj):
        if obj is None: return
        self.info = getattr(obj, 'info', None)

def sum_worker(x):
    return sum(x) , x.info

if __name__ == '__main__':
    arr_list = [MyArray(np.random.rand(3), info=f'foo_{i}') for i in range(10)]
    with Pool() as p:
        p.map(sum_worker, arr_list)

The attribute info is dropped

AttributeError: 'MyArray' object has no attribute 'info'

Using the builtin map() works fine

arr_list = [MyArray(np.random.rand(3), info=f'foo_{i}') for i in range(10)]
list(map(sum_worker, arr_list2))

The purpose of the method __array_finalize__() is that the object keeps the attribute after slicing

arr = MyArray([1,2,3], info='foo')
subarr = arr[:2]
print(subarr.info)

But for Pool.map() this method is somehow not working...

回答1:

Because multiprocessing uses pickle to serialize data to/from separate processes, this is essentially a duplicate of this question.

Adapting the accepted solution from that question, your example becomes:

from multiprocessing import Pool
import numpy as np

class MyArray(np.ndarray):

    def __new__(cls, input_array, info=None):
        obj = np.asarray(input_array).view(cls)
        obj.info = info
        return obj

    def __array_finalize__(self, obj):
        if obj is None: return
        self.info = getattr(obj, 'info', None)

    def __reduce__(self):
        pickled_state = super(MyArray, self).__reduce__()
        new_state = pickled_state[2] + (self.info,)
        return (pickled_state[0], pickled_state[1], new_state)

    def __setstate__(self, state):
        self.info = state[-1]
        super(MyArray, self).__setstate__(state[0:-1])

def sum_worker(x):
    return sum(x) , x.info

if __name__ == '__main__':
    arr_list = [MyArray(np.random.rand(3), info=f'foo_{i}') for i in range(10)]
    with Pool() as p:
        p.map(sum_worker, arr_list)

Note, the second answer suggests you might be able to use pathos.multiprocessing with your unadapted original code since pathos uses dill instead of pickle. This did not work when I tested it however.

来源：https://stackoverflow.com/questions/46813375/multiprocessing-pool-map-drops-attribute-of-subclassed-ndarray