Gym: conda installation support

[jimdowling]

It would be great to have an anaconda package for installing gym in conda environments using conda (not pip).

Background: I'm running a large DL course with 150 students, and much of my time has gone on getting 'pip install gym[...]' to work for conda enviornments. There have been issues with gcc incompatability, the usual python 2.7/3.6 issues. I expect my situation is not unique.

[FirefoxMetzger]

This is not a solution, but out of sheer curiosity: What swayed you away from using tools like Docker for teaching?

[jimdowling]

They work on real big data (40+ TB of different ML datasets) - we use Hops Hadoop as a multi-tenant platform that supports GPUs. It supports conda and runs Tensorflow/SparkML jobs on YARN. Docker isn't there yet with YARN. Afaik, we're the only distributed DL course where students work with massive datasets.

[FirefoxMetzger]

Hm, I suppose you already looked into conda skeleton pypi gym to try and generate the anaconda package from the available pypi one. Did that cause any problems?

[jimdowling]

We're limited in the conda channels we can use - 'default', 'conda-forge'. So, no didn't try it. I suppose i could try and get it in 'conda-forge'.

[andrewgough94]

Any update here ?

[kevindavenport]

Curious of the same

[netkachev]

+1, this feature would be nice to have

[MulderNick]

I am pretty new to python including anaconda an pip. For me pip was installed in anaconda3 and i was able to successfully use pip install gym in the anaconda terminal. It installed successfully into my environment. I do now know if this is a good solution however.
Perhaps this is of use as well

[Ketki-Savle]

Thank you. I tried conda install --upgrade gym and conda install gym and got an error. pip install gym went through quickly and could work on it easily.

[kota8]

pip install gym works good, but supporting conda install gym will be great. An advantage is that it can resolve dependency with external software. Such as libgcc #543 and ffmpeg. https://github.com/openai/gym/blob/5404b39d06f72012f562ec41f60734bd4b5ceb4b/gym/wrappers/monitoring/video_recorder.py#L249-L254
Both could be resolved by conda packaging by linking to conda install libgcc and conda install -c conda-forge ffmpeg.

[christopherhesse]

Conda support would be great, but I think we can get a lot of the benefit by making the pip install more reliable for everyone. I think video_recorder.py should probably use imageio which already has a package, imageio-ffmpeg, and then we don't need a ffmpeg dependency. For gcc, I think it would be best if we have compiled wheels for gym dependencies.

What packages are people having the most trouble with? From https://github.com/openai/gym/blob/master/setup.py#L9 it looks like atari-py and box2d-py are the most likely to fail (besides mujoco licensing issues, but we've recently released binary wheels for roboschool https://github.com/openai/roboschool). Those dependencies now have binary wheels on mac and linux that should work for most people.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[josephcarmack]

I tried creating a conda package version of gym with conda skeleton pypi gym but could not get it to work. My failure is likely due to my lack of experience with this feature of conda so any help with getting a conda package version of gym would be much appreciated.

[AdamStelmaszczyk]

That would not work, because gym is not present in the default conda channel. Right now you can install it with pip.

I will probably look into creating conda package, it would be very handy.

[justinkterry]

@AdamStelmaszczyk Any movement here? If not I'd be interested in working on doing official support for it.

[AdamStelmaszczyk]

Hey @justinkterry, sorry I didn't have time. I remember I looked briefly into it and seemed doable, but quite a bit of work. Feel welcome to do it :)

[justinkterry]

So it's already on conda: https://anaconda.org/powerai/gym, so porting it isn't the issue. The main challenge is that we'd want the release to be official, and to happen when it's pushed to PyPI. That would require involvement on the part of Gym maintainers and is not something I could fix on my own. @ppzhokhov ?