diff --git a/docs/docs.html b/docs/docs.html index 5fd8808..322e000 100644 --- a/docs/docs.html +++ b/docs/docs.html @@ -84,25 +84,29 @@

Training Demo

python demo.py --help # Get help on a specific environment -python demo.py --help --env snake +python demo.py --help --env puffer_snake -# Train breakout with multiprocessing: +# Train breakout with multiprocessing (24 cores): python demo.py --mode train --env breakout --vec multiprocessing -# Run a hyperparameter sweep on Ocean pong: -python demo.py --mode sweep-carbs --env pong --vec multiprocessing +# Run a hyperparameter sweep on Ocean pong. Requires carbs (github pufferai/carbs): +# TODO: Clean up CARBS defaults +python demo.py --mode sweep-carbs --env puffer_pong -# Train Ocean snake with native vectorization and wandb logs: -python demo.py --mode train --env snake --vec native --track +# Train Ocean snake with wandb logs: +python demo.py --env puffer_snake --mode train --track # Set train and env params from cli: -python demo.py --mode train --env snake --vec native --train.learning_rate 0.001 --env.num_snakes 512 +python demo.py --env puffer_snake --mode train --train.learning-rate 0.001 --env.vision 3 # Eval a pretrained baseline model: -python demo.py --mode eval --env snake --vec native --baseline +python demo.py --env puffer_snake --mode eval --baseline + +# Eval an uninitialized policy: +python demo.py --env puffer_snake --mode eval --baseline # Eval a local checkpoint: -python demo.py --mode eval --env snake --vec native --eval-model-path your_model.pt +python demo.py --env puffer_snake --mode eval --eval-model-path your_model.pt

Compared to the original CleanRL code, our demo file (which loads clean_pufferl.py) supports asynchronous on-policy vectorization, better multi-agent training, a convenient cli dashboard, better WandB log and sweeps integration, and more. It's only around 1000 lines of code, most of which is logging.