In early development. Inspired by TF-Agents, I hope to make a similar library for PyTorch - simple enough for me to understand, and powerful enough to express all my crazy ideas.
Currently supported: (Double) Deep Q Networks and Open AI Gym environments.
DDQN Learning Pong after 13 hours (2500 games):
pongdemo.mp4
DDQN Learning Breakout after 16 hours (10k games):
breakoutdemo.mp4
DDQN Learning Space Invaders after 10k games:
spaceinvdemo.mp4
First, install PyTorch with the latest CUDA version for your platform from pytorch.org.
Then install other required Python libraries with pip:
pip3 install tensorboard
pip3 install gym[classic_control]
pip3 install gym[atari]
pip3 install gym[accept-rom-license]
pip3 install gym[other]
pip3 install gym[box2d]