Question Parallelization of Environments

Hi!

So I have some questions regarding parallelization in ml-agents. Maybe someone can clarify these things.
With the python API you are able to run multiple environments inside one singe executable.
Is this faster then spawining multiple instances of the Environment, like with the gym wrapper and for example baselines SupProcVec class?
Also is there a similar class (like supprocvec env) for the python api or an example how to run multiple instances of an executable, with for example mpi4py? Or would it make sense to just increase the number of environments of the executable?
Thanks in advance!

 

Also another thing: I read somewhere, that you cannot use the parameter no-graphics=True when you are using physics inside the game. Is that right? So I could not use collison checks/Rigidbodys?

 

I have asked the similiar question before. there is a way for multiple instances of an executable to use "--num-envs".
no-graphics=True when not using visual observation. you can have a try and I am not sure about it.

 

I am using my own implementations of Algorithms so I cannot use --num-envs. That works only for ppo and sac I think, which are already implemented.