Can't pickle : attribute lookup cv2.BRISK failed

Question

I'm trying to run multiple CMT trackers simultaneously. For that reason, I'm setting a Pool of threads:

import argparse
import cv2
from multiprocessing import Pool
import numpy as np
import os
import sys
import time

import VARtracker
import util

CMT1 = VARtracker.CMT()

... # code lines removed

# Clean up
cv2.destroyAllWindows()

if args.inputpath is not None:
    # If a path to a file was given, assume it is a single video file
    if os.path.isfile(args.inputpath):
        cap = cv2.VideoCapture(args.inputpath)
        # Skip first frames if required
        if args.skip is not None:
            cap.set(cv2.cv.CV_CAP_PROP_POS_FRAMES, args.skip)

    # Otherwise assume it is a format string for reading images
    else:
        cap = util.FileVideoCapture(args.inputpath)
        # Skip first frames if required
        if args.skip is not None:
            cap.frame = 1 + args.skip

    # Check if videocapture is working
    if not cap.isOpened():
        print 'Unable to open video input.'
        sys.exit(1)

    # Read first frame
    status, im0 = cap.read()
    im_gray0 = cv2.cvtColor(im0, cv2.COLOR_BGR2GRAY)
    im_draw = np.copy(im0)

# Getting initial bounding boxes
tl1 = [405, 160]
br1 = [450, 275]

VARtracker.initialise(CMT1, im_gray0, tl1, br1)

frame = 1
while True:
    pool = Pool(processes=4)
    print frame

    # Read image
    status, im = cap.read()
    if not status:
        break
    im_gray = cv2.cvtColor(im, cv2.COLOR_BGR2GRAY)
    im_draw = np.copy(im)

    tic = time.time()
    # Serial approach
    #res1 = VARtracker.process_frame(CMT1, im_gray)

    # Parallel approach
    res1 = pool.apply_async(VARtracker.process_frame, (CMT2, im_gray))
    pool.close()
    pool.join()
    res1 = res1.get()
    toc = time.time()

    # Display results
    if res1.has_result:
        cv2.line(im_draw, res1.tl, res1.tr, (255, 0, 0), 4)
        cv2.line(im_draw, res1.tr, res1.br, (255, 0, 0), 4)
        cv2.line(im_draw, res1.br, res1.bl, (255, 0, 0), 4)
        cv2.line(im_draw, res1.bl, res1.tl, (255, 0, 0), 4)

    if not args.quiet:
        cv2.imshow('main', im_draw)
        cv2.waitKey(pause_time)

    # Remember image
    im_prev = im_gray
    frame += 1

Whenever I comment the Serial approach and attempt using threads (Parallel aproach), I come across the following error:

Traceback (most recent call last):

File "/home/rafael/GIT/CMT-Tracker/VaretoCMT/VARmain.py", line 128, in module res1 = res1.get()

File "/usr/lib/python2.7/multiprocessing/pool.py", line 558, in get raise self._value

cPickle.PicklingError: Can't pickle : attribute lookup cv2.BRISK failed

The other files can be encountered on VARmain.py, VARtracker.py and util.py.

I've tried so many ways and I still haven't found a way to overcome this Python limitation. I found out that I cannot serialize class methods, only functions. If possible, I would like to solve it using Python standard libraries.

Answer 1

I managed to solve it. Thanks to @Matt and @Yamaneko. Basically, I moved the block that reads the image into the worker function. Therefore, if the pool size = 6 and there are six bounding boxes, each frame is going to be read six times (within each worker). That's the only way I have found to make it work.

Current version can be found here.

import cv2 as cv
import multiprocessing as mp
import time

def worker(folder_path, list_name, top_left, bot_right, index):
    frame_path = folder_path + '/' + list_name[0]
    image_0 = cv.imread(frame_path)
    gray_0 = cv.cvtColor(image_0, cv.COLOR_BGR2GRAY)

    cmt = VARtracker.CMT()
    cmt.initialise(gray_0, top_left, bot_right)
    box_queue = mp.Queue()

    for name in list_name:
        frame_path = folder_path + '/' + name
        image_now = cv.imread(frame_path)
        gray_now = cv.cvtColor(image_now, cv.COLOR_BGR2GRAY)

        cmt.process_frame(gray_now)
        if cmt.has_result:
            print index, name, zip(cmt.tl, cmt.br)
            output.put((index, name, zip(cmt.tl, cmt.br)))
    print 'Process {} finished'.format(index)

def VARmethod(folder_path, final_frame, top_left, bot_right):
    tic = time.time()

    if len(top_left) == len(bot_right):
        list_frame = [index for index in range(1, final_frame + 1)]
        list_name = [str(index) + '.jpg' for index in list_frame]

        pool = mp.Pool(5)
        for index in range(0, len(top_left)):
            pool.apply_async(worker, args=(folder_path, list_name, top_left[index], bot_right[index], index))
        pool.close()
        pool.join()

        print 'Finished with the script'

    toc = time.time()
    print output.qsize()
    print (toc - tic)

Answer 2

Try this code segment around your classes (this is not my code, credit to Steven Bethard) - this is a workaround to pickle classes; pickle is used by the multiprocessing module to send jobs to workers:

def _pickle_method(method):
    func_name = method.im_func.__name__
    obj = method.im_self
    cls = method.im_class
    return _unpickle_method, (func_name, obj, cls)

def _unpickle_method(func_name, obj, cls):
    for cls in cls.mro():
        try:
            func = cls.__dict__[func_name]
        except KeyError:
            pass
        else:
            break
    return func.__get__(obj, cls)

import copy_reg
import types
copy_reg.pickle(types.MethodType, _pickle_method, _unpickle_method)

Example using it with multiprocessing here Can't pickle <type 'instancemethod'> when using python's multiprocessing Pool.map()

Not saying this would be easy to convert. If you really want multithreading, I suggest Cython with OpenMP. You can just rewrite the parts of the program that need to be parallel with nogil statements and from cython.parallel cimport prange for parallel loops...