Skip to content

Latest commit

 

History

History
48 lines (37 loc) · 2.3 KB

README.md

File metadata and controls

48 lines (37 loc) · 2.3 KB
Raw

Offline RL: AWAC & IQL

PyTorch Black

This repository contains a PyTorch implementation of two offline reinforcement learning algorithms:

  1. AWAC: Accelerating Online Reinforcement Learning with Offline Datasets
  2. Offline Reinforcement Learning with Implicit Q-Learning

Installation

Each algorithm is provided in two versions: a .py file and an .ipynb notebook. It is recommended to run the IPython versions directly in Google Colab. For the .py file to work, you need to install and set up d4rl and mujoco on your local machine. You can use the following command for installation:

 apt-get -qq update
 apt-get -qq install -y libosmesa6-dev libgl1-mesa-glx libglfw3 libgl1-mesa-dev libglew-dev patchelf
 mkdir ~/.mujoco
 wget -q https://mujoco.org/download/mujoco210-linux-x86_64.tar.gz -O mujoco.tar.gz
 tar -zxf mujoco.tar.gz -C "$HOME/.mujoco"
 rm mujoco.tar.gz
 echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/.mujoco/mujoco210/bin' >> ~/.bashrc
 echo 'export LD_PRELOAD=$LD_PRELOAD:/usr/lib/x86_64-linux-gnu/libGLEW.so' >> ~/.bashrc
 echo "/root/.mujoco/mujoco210/bin" > /etc/ld.so.conf.d/mujoco_ld_lib_path.conf
 ldconfig
 pip3 install -U 'mujoco-py<2.2,>=2.1'
 touch .mujoco_setup_complete

Please note that we cannot guarantee that this script will successfully install all the required libraries on your local machine. If any values are missing in your ~/.bashrc file, an error will be reported upon launching.

To set up your local machine, you can also consider using the Dockerfile available at: https://github.com/tinkoff-ai/CORL/blob/main/Dockerfile

Training

To run training locally, execute the following command:

python3 iql.py --yaml_path=configs/iql/antmaze_medium_seed_0.yaml

Ensure that the yaml_path parameter contains the path to the .yaml file with the desired configuration.

Credits

This implementation draws significant inspiration from the following repositories: