This is the environment bit of CuLE, bundled up into a conda package. Install it with
conda install torchcule -c ajones -c pytorch -c default -c conda-forge
Limitations are that it drags a bit of conda-forge in with it when you install it elsewhere (urgh), and I've only compiled it for my architecture. If you want to build it yourself,
conda-build pkg -c pytorch -c default -c conda-forge
This'll likely take ~30 min. You can test it out on a handful of envs by adding the --fastbuild switch to build.sh.
I do my building and testing in a docker container, which is a version of my standard development env.
All the code in here is from the Nvidia CuLE project. If you use this work, you should cite their paper:
@misc{dalton2019gpuaccelerated,
title={GPU-Accelerated Atari Emulation for Reinforcement Learning},
author={Steven Dalton and Iuri Frosio and Michael Garland},
year={2019},
eprint={1907.08467},
archivePrefix={arXiv},
primaryClass={cs.LG}
}