AttributeError: 'SharedArray' object has no attribute 'dtype'

[SOMEAIDI]

when I try the tutorial of pistonball, I met this attribute error. But I have no idea that what should do. So what should I do to deal with the error?

[jkterry1]

Could you please provide system specs (OS, python verison, etc.) and give me your pip freeze? And did you use this file exactly? https://github.com/PettingZoo-Team/PettingZoo/blob/master/tutorials/13_lines.py

[benblack769]

Just a heads up, this is a common problem for windows users, see #100 for someone who had a similar issue.

[SOMEAIDI]

Could you please provide system specs (OS, python verison, etc.) and give me your pip freeze? And did you use this file exactly? https://github.com/PettingZoo-Team/PettingZoo/blob/master/tutorials/13_lines.py

Thanks for your reply.
I perform my code on windows.And my code is :

    env = pistonball_v4.parallel_env(n_pistons=20, time_penalty=-0.1, continuous=True, random_drop=True,
                                     random_rotate=True, ball_mass=0.75, ball_friction=0.3, ball_elasticity=1.5, max_cycles=125)

    env = ss.color_reduction_v0(env, mode='B')

    env = ss.resize_v0(env, x_size=84, y_size=84)

    env = ss.frame_stack_v1(env, 3)

    env = ss.concat_vec_envs_v0(env, 8, num_cpus=1, base_class='stable_baselines3')

It doesn't work at all.And the error is about stable_baseline3 , "You should use natureCNN" is the attribute error.
But I find that when I remove some code :

    env = pistonball_v4.parallel_env(n_pistons=20, time_penalty=-0.1, continuous=True, random_drop=True,
                                     random_rotate=True, ball_mass=0.75, ball_friction=0.3, ball_elasticity=1.5, max_cycles=125)

    # env = ss.color_reduction_v0(env, mode='B')

    env = ss.resize_v0(env, x_size=84, y_size=84)

    # env = ss.frame_stack_v1(env, 3)

    env = ss.concat_vec_envs_v0(env, 8, num_cpus=1, base_class='stable_baselines3')

Now it works correctly, but I have no idea why this happen . So is Supersuit not suitable for windows ?

[benblack769]

So when you set num_cpus to 1, that should have solved the error in #100. Is that the case?

This error about natureCNN is a bit les clear to me, but it sounds like something you are doing to the preprocessing is confusing SB3. If you could post some complete (with imports and everything) code that you expect to work, but intsead the code fails, I can help you see what is going wrong.

[jkterry1]

Closing as it's been awhile, please let us know if you need anything else.

[ic]

For your information, same problem today on MacOS, with the same tutorial.

PettingZoo==1.13.1
SuperSuit==3.3.0

python 3.9.7

#100 solves the problem by reducing num_cpus to 1 (sadly 16 cores eagerly available :-)).

[jkterry1]

@benblack769 I was getting this error too

[benblack769]

@jkterry1 I'll prioritize this. Of course, in the meantime, you can simply turn off multiprocessing by setting num_cpus=1

[benblack769]

This was a change in python3.8 caused by changes in the core MacOS system. See the following trail:

• python 3.8 changes default: https://docs.python.org/3/library/multiprocessing.html#contexts-and-start-methods
• due to problem layed out here: https://bugs.python.org/issue33725
• Objective-C makes it hard to just pretend it isn't a problem and hope for the best: http://sealiesoftware.com/blog/archive/2017/6/5/Objective-C_and_fork_in_macOS_1013.html

It might be necessary to make changes to the 13 lines tutorial as well as changes to supersuit to get this to work (putting the code after a __name__ == "__main__" guard).

[jkterry1]

@benblack769 this was only happening to me in Ubuntu

[benblack769]

@jkterry1 If you could give more details as to what your setup was when this error was triggered that would be great. Typically, this would happen with "spawn" multiprocessing, which is a default on MacOS or Windows, but "fork" multiprocessing is the default on linux.

[wittenator]

Out of curiosity:
Why is the loading and setting for pickling of the SharedArray asymmetrical at this part of the code:

SuperSuit/supersuit/vector/utils/shared_array.py

Line 15 in be8dcfc

def __getstate__(self):

?

I had a similar error and if I am not entirely mistaken that fixed that error for me.

[benblack769]

Ah. That would be a bug.

[benblack769]

@wittenator Thanks for pointing this out! I didn't know that it was possible for process shared data to be passed to spawn environments, but I guess they can. I went ahead and tested it in the PR and it seems to work.