Add pickle and unpickle support.

[mo-vic]

Picklable objects for multithreading support.

Sometime we have a list of queries to run where each query is independent of each other. In such case, we can create a thread pool and copy the same environment in each thread to run queries in parallel. The following code gives an example of doing this:

import fcl
import numpy as np
from multiprocessing import Pool


class CustomEnv:
    def __init__(self) -> None:
        verts1 = np.array([[0.0, 0.0, 0.0], [1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]])
        tris1 = np.array([[0, 2, 1], [0, 3, 2], [0, 1, 3], [1, 2, 3]])
        self.mesh1 = fcl.BVHModel()
        self.mesh1.beginModel(len(verts1), len(tris1))
        self.mesh1.addSubModel(verts1, tris1)
        self.mesh1.endModel()

        verts2 = verts1 - np.array([[0.0, 1.5, 0.0]])
        tris2 = tris1

        self.mesh2 = fcl.BVHModel()
        self.mesh2.beginModel(len(verts2), len(tris2))
        self.mesh2.addSubModel(verts2, tris2)
        self.mesh2.endModel()

        R = np.eye(3)
        T = np.zeros(3)

        tf = fcl.Transform(R, T)

        self.obj1 = fcl.CollisionObject(self.mesh1, tf)
        self.obj2 = fcl.CollisionObject(self.mesh2, tf)

    def __call__(self, theta) -> bool:
        R = np.array([[np.cos(theta), -np.sin(theta), 0], [np.sin(theta), np.cos(theta), 0], [0, 0, 1]])
        self.obj2.setRotation(R)

        request = fcl.CollisionRequest()
        result = fcl.CollisionResult()

        ret = fcl.collide(self.obj1, self.obj2, request, result)

        return theta, ret


if __name__ == "__main__":
    myEnv = CustomEnv()

    theta = np.linspace(0.0, 2.0 * np.pi, 360)
    with Pool(processes=4) as pool:
        for a, b in pool.map(myEnv, theta):
            print(a / np.pi * 180, b)

However, objects like fcl.BVHModel, fcl.CollisionObject are ==unpicklable==, making it unable to serialize them, and de-serialize them in threads:

TypeError
no default __reduce__ due to non-trivial __cinit__
  File "./python-fcl/tests/test_multithreading.py", line 48, in <module>
    for a, b in pool.map(myEnv, theta):
TypeError: no default __reduce__ due to non-trivial __cinit__

My solution to address this issue is to derive subclasses from those Cython class, add __init__ method to cache input arguments and __reduce__ method to return the cached data for pickling.

Currently only support fcl.Transform, fcl.BVHModel, fcl.CollisionObject, with this commit and a simple magic import: from fcl import fcl2 as fcl, the above example script can run in parallel.

[mikedh]

Yeah it would be nice to make them serializable! It would be nice if it was defined on the original object though (vs a separate fcl2.py + inheritance), what if you defined an __iter__ and then passed them around as dict objects for keyword arguments?

class Thing:
    def __init__(self, a=10, b = 20):
        self.a = a
        self.b = b

    def __iter__(self):
        return iter([('a', self.a),
                     ('b', self.b)])

if __name__ == '__main__':
    t = Thing(a=5, b=2)

    # picklable, json-dumpable, etc.
    thing_serial = dict(t)

[mo-vic]

Yeah it would be nice to make them serializable! It would be nice if it was defined on the original object though (vs a separate fcl2.py + inheritance), what if you defined an __iter__ and then passed them around as dict objects for keyword arguments?

class Thing:
    def __init__(self, a=10, b = 20):
        self.a = a
        self.b = b

    def __iter__(self):
        return iter([('a', self.a),
                     ('b', self.b)])

if __name__ == '__main__':
    t = Thing(a=5, b=2)

    # picklable, json-dumpable, etc.
    thing_serial = dict(t)

But, you still need to create the FCL object in the Process right? then again, this re-occurs when you define the FCL object used the serialized dict. For example, using FCL in PyTorch's dataloader with multiprocessing support, I think it is still necessary to define a reduce method.

[mo-vic]

Yeah it would be nice to make them serializable! It would be nice if it was defined on the original object though (vs a separate fcl2.py + inheritance), what if you defined an __iter__ and then passed them around as dict objects for keyword arguments?

class Thing:
    def __init__(self, a=10, b = 20):
        self.a = a
        self.b = b

    def __iter__(self):
        return iter([('a', self.a),
                     ('b', self.b)])

if __name__ == '__main__':
    t = Thing(a=5, b=2)

    # picklable, json-dumpable, etc.
    thing_serial = dict(t)

Hi mikedh, I tried adding a __iter__ in the original Cython class like below:

def __iter__(self):
    return iter([("a", 1), ("b", 2)])

but the object is not directly picklable:

In [5]: import fcl

In [6]: tf_to_pickle = fcl.Transform(R, T)

In [7]: pickled_tf = pickle.dumps(tf_to_pickle)
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[7], line 1
----> 1 pickled_tf = pickle.dumps(tf_to_pickle)

File stringsource:2, in fcl.fcl.Transform.__reduce_cython__()

TypeError: no default __reduce__ due to non-trivial __cinit__

Also, when I try to cache the data in the __cinit__ method, like this:

def __cinit__(self, *args):
    self.args = args

when creating the object, an AttributeError has been thrown:

In [5]: tf_to_pickle = fcl.Transform(R, T)
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[5], line 1
----> 1 tf_to_pickle = fcl.Transform(R, T)

File /mnt/e/Users/movic/python-fcl/src/fcl/fcl.pyx:56, in fcl.fcl.Transform.__cinit__()
     54
     55     def __cinit__(self, *args):
---> 56         self.args = args
     57         if len(args) == 0:
     58             self.thisptr = new defs.Transform3d()

AttributeError: 'fcl.fcl.Transform' object has no attribute 'args'

so I guess the inheritance solution is still worth considering.