Poor PPO 1 CPU performance on the Pong task

[openai]

Poor PPO 1 CPU performance on the Pong task

[shijcnju]

I failed to run PPO, can you tell me how to do it?
Thank you~

[ViktorM]

@shijcnju sorry for delay with reply, traveling at the moment. If you previously had no problem to run DQN. It's pretty straightforward. The only required change if you plan to run it on PC or laptop with 1 CPU- to set num_cpu=1 in:

def main():
    train('PongNoFrameskip-v4', num_timesteps=40e6, seed=0, num_cpu=8)

Except this change PPO for me was running out of the box, except reward growth.

[ViktorM]

Also you should pip install mpi4py first and on Windows a few more actions were required to make mpi4py work.

[iamzoumao]

If you set num_cpu = 4 and run run_atari.py again, you should observe around 17.0 mean reward in about 2k iterations.

However if you are working in Anaconda environment and have installed mpi4py via pip or conda install, this probably will report some errors like 'detected that one or more processes exited with non-zero status'. In this case, you should remove mpi4y and reinstall from source https://bitbucket.org/mpi4py/mpi4py instead.

[ViktorM]

Thanks @iamzoumao for the advice. With 4 num_cpu I see a bit different errors:

 File "run_atari2.py", line 54, in <module>
    main()
  File "run_atari2.py", line 51, in main
    train('PongNoFrameskip-v4', num_timesteps=40e6, seed=0, num_cpu=4)
  File "run_atari2.py", line 20, in train
    whoami  = mpi_fork(num_cpu)
  File "D:\baselines\baselines\common\mpi_fork.py", line 20, in mpi_fork
    subprocess.check_call(args, env=env)
  File "C:\Users\Viktor\Anaconda3\envs\deep-rl\lib\subprocess.py", line 576, in check_call
    retcode = call(*popenargs, **kwargs)
  File "C:\Users\Viktor\Anaconda3\envs\deep-rl\lib\subprocess.py", line 557, in call
    with Popen(*popenargs, **kwargs) as p:
  File "C:\Users\Viktor\Anaconda3\envs\deep-rl\lib\subprocess.py", line 947, in __init__
    restore_signals, start_new_session)
  File "C:\Users\Viktor\Anaconda3\envs\deep-rl\lib\subprocess.py", line 1224, in _execute_child
    startupinfo)
FileNotFoundError: [WinError 2] The system cannot find the file specified

But will try reinstalling from source in any way, thanks.

[nchentanez]

Ran into the same problem as @ViktorM. Tried installing mpi4py from source, but got the same error.

[nchentanez]

Ok I got it to work:
First, install both the mpi run time and mpi SDK from
https://www.microsoft.com/en-us/download/details.aspx?id=55494

Then rename/copy
C:\Program Files\Microsoft MPI\Bin\mpiexec.exe
to
C:\Program Files\Microsoft MPI\Bin\mpirun.exe

or need to modify
baselines\common\mpi_fork.py
to call mpiexec instead, if running on Windows

[KeirSimmons]

Using ubuntu 16.04, and running these from a conda environment. Unable to run PPO or TRPO as I get the error that @iamzoumao stated. Was unable to make mpi4py from source. Any workaround for this?

[ViktorM]

@KeirSimmons for me on Ubuntu 16.04 just conda install mpi4py worked well.

[olegklimov]

We don't officially support Windows. It's great if it works, but we don't have Windows here anywhere. Please post a PR if you're able to fix a specific Windows problem.