·
2 commits
to 5.0
since this release
Experiments and pretrained models for PufferLib 5.0
Unzip puffer5_experiments -> copy env folders to logs/. This is a lot of data, even for constellation. The experiments.ini we ship with the repo is a pareto-filtered version. If you want to inspect the full set in constellation, build cache_data and run with --full. We strongly suggest only loading one env at a time in this case.
We ship demo files for a few of the coolest environments. You can reproduce the rest from scratch quite easily on a single GPU. The file name tells you the --policy.hidden_size and --policy.num_layers to use for eval.