Defender-ram-v0 doesn't initialize

[jacekplocharczyk]

Hi,
could you check why Defender-ram-v0 cannot initialize the environment?

I've tested it in the jupyter online https://jupyter.org/try - jupyter lab.

!pip install gym[atari]

import gym
env = gym.make('DemonAttack-ram-v0') # works
print(env)
env = gym.make('Defender-ram-v0') # doesn't work (it hangs)
print(env)

Edit.
I've tested all environments in my docker container and all versions of the Defender-* doesn't work.

[christopherhesse]

I was able to reproduce this on colaboratory, but not on a normal linux machine. This could be a bug in https://github.com/mgbellemare/Arcade-Learning-Environment or else a bug in how colab/jupyter handles something.

Can you try running this directly on a linux machine without docker?

[jacekplocharczyk]

Unfortunately I can't recreate the whole docker setup (there are a lot of CUDA related stuff and I don't want to mess with the NVIDIA drivers).

If that helps here is the Dockerfile.

[christopherhesse]

This looks to be an issue with the defender rom, the following Dockerfile seems to work for me:

FROM continuumio/miniconda:4.7.12
RUN conda install python==3.7.3
RUN pip install gym[atari]
RUN apt-get update && apt-get install -y procps
RUN pip install py-spy
RUN python -c "import gym; env = gym.make('DemonAttack-ram-v0'); print(env.reset().shape)"
RUN md5sum /opt/conda/lib/python3.7/site-packages/atari_py/atari_roms/defender.bin
ADD defender.bin /opt/conda/lib/python3.7/site-packages/atari_py/atari_roms/defender.bin
RUN md5sum /opt/conda/lib/python3.7/site-packages/atari_py/atari_roms/defender.bin
RUN python -c "import gym; env = gym.make('Defender-ram-v0'); print(env.reset().shape)"

The newest version of atari-py doesn't have this issue because it doesn't include the bad rom (https://github.com/openai/atari-py#roms). Your ROM md5 should match the one in this file: https://github.com/mgbellemare/Arcade-Learning-Environment/blob/master/md5.txt

We should upgrade atari-py to the new ale-py when that is available and it should fix this issue in general: Farama-Foundation/Arcade-Learning-Environment#272

[gitanshu]

Bump!
I still see the same issue, can reliably reproduce on ubuntu 18.04 on local, and in docker on CI, even with latest release of gym.
It works fine though if I downgrade atari-py to latest 0.1.x.

[christopherhesse]

I think the new version of ale-py is almost ready, and @JesseFarebro was considering upgrading the version gym uses, but I haven't heard about the status of that recently.

[JesseFarebro]

I am blocked on a couple of things right now but I'll push to have this completed in the next couple of weeks.