openai/procgen

Packaging custom environments for use in train-procgen repo

max-reuter opened this issue · 4 comments

As someone less familiar with packaging, I'm having trouble getting started with porting my custom environments such that I can train them using the train-procgen repo's code. This issue seems to be the closest post to my problem. I've opened a similar issue in the train-procgen repo.

I'd appreciate any instructions or tips for getting started.

If you want to create a binary wheel of procgen, I'd suggest using the existing travis CI configuration: https://github.com/openai/procgen/blob/master/.travis.yml which you may want to get running on a local machine first.

A github actions version would be nice, but I haven't had time to port it to github actions.

Does that answer your question? Getting the wheels to build correctly (i.e. in the same way as the wheels on pypi) without using the existing build scripts will likely be a challenge.

I'm not very familiar with travis or wheels. Could you please spell out some steps/checkpoints for packaging my current procgen fork such that I can set up my fork of train-procgen to use that instead of the original procgen repo? My goal is to be able to train on an edited version of bossfight.

Solved by putting the train_procgen directory in my procgen directory (clone of this repo); then, running the training commands in the train-procgen repo automatically rebuilds based on the contents of my edited environments in my procgen directory, using them during training. Thanks @christopher-hesse!