Failed to run IMALA and MADDPG from examples folder about machin HOT 2 CLOSED

Hi, thank you for trying out and report these bugs, I will answer them one by one:
[MADDPG]
This problem is caused by API change, the initialization part of MADDPG in the example should be changed to (for v0.4.1):

maddpg = MADDPG(
        [deepcopy(actor) for _ in range(agent_num)],
        [deepcopy(actor) for _ in range(agent_num)],
        [deepcopy(critic) for _ in range(agent_num)],
        [deepcopy(critic) for _ in range(agent_num)],
        t.optim.Adam,
        nn.MSELoss(reduction="sum"),
        critic_visible_actors=[list(range(agent_num))] * agent_num,
    )

[IMPALA]
For IMPALA I just found some serious bugs, first, please move learner group creation in the example outside the if-else like:

    learner_group = world.create_collective_group(ranks=[2, 3])
    if rank in (2, 3):
         ...

Then in machin.parallel.distributed.world, add sorted to these methods of class World:

    def get_ranks(self):
        """
        Returns:
            A list of ranks of all processes.
        """
        return sorted(list(self.rank_name_map.keys()))

    def get_members(self):
        """
        Returns:
            A list of names of all processes.
        """
        return sorted(list(self.rank_name_map.values()))

These are temporary fixes and should be able to solve your problem, the "connect()" problem is caused by incoherent member position over all involving processes, to be short.

[Bugs in the test]
These bugs are all from the auto module and they are caused by incorrect returning annotation, I think these type-checking features are introduced in newer PyTorch versions, I will fix that. Besides, there are also bugs caused by PyTorch lightning altering their APIs, I am currently fixing that. These problems will not affect you if auto module is not used.

from machin.

Thanks for the response, with your help now is working properly :D.

from machin.