train_envs = SubprocVectorEnv(
            [lambda: gym.make(args.task) for _ in range(args.training_num)])
train_envs.action_space[0]
>Box(8,)
a = train_envs.action_space[0]
b = train_envs.action_space[0]
a is b
>False  # For SubprocVectorEnv, but for DummyVectorEnv the result is true.

This is because that SubprocVectorEnv returns a deep copy but DummyVectorEnv return a reference.
As a result,
If you use SubprocVectorEnv and tianshou/collector, when you use 'random' option in collect() method, you will get different random action everytime you run the scripts even when using same random seeds.

So you have to be very careful about when you call 'train_envs.action_space[0].seed(1)'