This repository contains the code for running the baselines from the Craftax paper. For packaging reasons, this is separate to the main repository.
git clone https://github.com/MichaelTMatthews/Craftax_Baselines.git
cd Craftax_Baselines
pip install -r requirements.txt -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
pre-commit install
python ppo.py
python ppo_rnn.py
python ppo.py --train_icm
python ppo.py --train_icm --use_e3b --icm_reward_coeff 0
python ppo_rnd.py
You can save trained policies with the --save_policy
flag. These can then be viewed with the view_ppo_agent
script (pass in the path up to the files
directory).