Zigrad has a new example of training an AI agent on the cartpole task with Tensorboard integration for logging metrics.
Cartpole is a simple task that involves balancing a pole by applying force to a cart. The agent learns to balance the pole using Deep Q-Learning. Also, this by far the fastest solution I have ever benchmarked (e.g. faster than PyTorch on GPU)! Interestingly, this implementation still stands to gain substantial performance so expect Zigrad to continue to widen the gap.
The tensorboard dashboard showing metrics logged during training:
Thank you to Arwalk (Arwalk) · GitHub for the protobuf bindings.
The example also includes python bindings for the replay buffer and physics simulator to help ensure benchmarks are as fair as possible for the PyTorch solution.
Want to get involved? Join the Discord and introduce yourself