dniku / rl-attention

rl-attention

Instructions

To use the launcher, run in the cloned repository:

python -m venv .env
source .env/bin/activate
pip install -r requirements.txt

To use the launcher, just run main.py. All parameters are stored in config.json.

Adding a new model architecture is essentially replacing a Policy. Currently config.json specifies a CnnPolicy which comes bundled with stable-baselines. See stable_baselines/common/policies.py for examples of how to define custom policies.

We also include a copy of the code for training algorithms here so that it can be modified more easily.

The complete trained model is stored in stored under saved_models as env_name-model_name-policy_type.pkl. The config file and 100-step reward averages are stored under saved_metrics as env_name-model_name-policy_type.txt.

Jupyter instructions

To install Jupyter, register a new kernel, and start a notebook, run in the virtual environment:

pip install jupyter
ipython kernel install --user --name=.env
jupyter notebook

Then activate the .env kernel in the notebook.

To make the logging work, add and execute after importing the logging module in the notebook:

logger = logging.getLogger()
logger.setLevel(logging.DEBUG)