Gym: Fail to render gym environment in Pyenv installed python

[mjlbach]

As context, I'm running all code locally on my Macbook Pro under MacOS Mojave 10.14.

Homebrew recently updated python to 3.7 which is currently not compatible with tensorflow. I reinstalled pyenv so I can manage my active python version and installed tensorflow + ai gym on 3.6.6. When I try to render an environment:

import gym
env = gym.make('CartPole-v0')
env.reset()

for i in range(1000):
    env.step(env.action_space.sample())
    env.render()

I get a blank rendering window when using the pyenv installed interpreter (3.6.6 and 3.7). When I use the hombrew installed intepreter (3.7) the rendering works fine.

Anyone else have this issue? I posted here as any suggestions as to what the majority of the community is using for their development environment (I've used conda, docker, and singularity before) would be welcome.

[damienstanton]

I can confirm the same issue, in a similar environment (Python 3.7.0 and macOS 10.14). I have tried in a standalone script, called by the python and ipython interpreters, as well as a jupyter session.

In my case, the same problem occurs in both the homebrew-installed Python as well as in a venv module virtual environment. I do not use Anaconda.

venv:

system (via homebrew):

It is not pictured, but in both cases the mouse is in an infinite "beach ball" loop. I have to kill the kernel/interpreter in all cases.

[pzhokhov]

@mjlbach @damienstanton your issues actually look different to me.
From what I understand, in @mjlbach 's case, rendering works with python 3.7, but does not with 3.6.6 in pyenv; whereas in @damienstanton 's case rendering does not work with with python 3.7.
Unfortunately, I have no good timeline for the fix (we don't have CI set up for osx - blame travis for lack of python support there, and I don't think anyone working on the gym has switched to Mojave just yet). However, @mjlbach , if you are open to use docker, you should have no problems - we have dockerfiles detailing installation on ubuntu 18.04 and 16.04. Also, if you don't want to deal with X forwarding from docker, but still would like to see what agents are going, it is possible to record and save the videos using VideoRecorder (https://github.com/openai/gym/blob/master/gym/wrappers/monitoring/video_recorder.py)
Sorry I cannot be more helpful here, as soon as I have a way to reproduce this, I'll get back to it.

[mjlbach]

No worries! I just wanted to document the issue. I've been using gym in a docker environment with X forwarding. Thanks for the suggestion about recording videos!

[damienstanton]

@mjlbach sounds like a great workaround in the meantime. Could you share how you use x forwarding in this case for gym?

[imanesh]

Same problem here after upgrading the OSX to Mojave, any fix for this blank screen issue?

[TeaPearce]

+1

[Abdisalan]

Same issue, using Mojave w/ Python 3.7.1

[gauthiermartin]

Same issue using Mojave w/ Python 3.7.1

[damienstanton]

@mjlbach
Can you share how you use

I've been using gym in a docker environment with X forwarding.

[gauthiermartin]

Simple pyenv in my local environment

[Journerist]

same issue here running python 3.6.4 via pyenv

[al-jshen]

Same issue. Mojave 10.14.1 with Python 3.7.1.

Edit: Updating to 10.14.2 solved it for me.

[gauthiermartin]

Does anyone other than @AL-JShen confirm that after updating to 10.14.2 the issue is fixed ?

[al-jshen]

@gauthiermartin

https://github.com/openai/gym/issues/1236

Check out this thread, a few others in there have also fixed their problems by updating.

[damienstanton]

This is great news. Not to be too negative, but I'm very glad that my current Mac laptop is the last Mac I will be using professionally 🙂

[gauthiermartin]

I confirm after updating to 10.14.2 it does work perfectly

[mjlbach]

Closed because 10.14.2 fixes the issue, and alternative solutions (Docker, rendering to video) were available.