Checkpoints, metrics and models

Checkpoints

Checkpoints save the experiment's state so training can resume. The configuration controls where they are written, how often to save and how many older checkpoints to retain.

Keep the output directory for a run when restarting it. Checkpoints contain more state than an inference model, including learner and experiment data. Use the same current configuration and checkpoint schema when restoring a run.

Metrics

Metrics describe what actors, learners and testers are doing. Console output gives a running view; TensorBoard provides charts for comparing rewards, episode lengths, losses and other measurements over time.

Point TensorBoard at the output directory for the run:

uv run tensorboard --logdir /path/to/training-output

Tester results help distinguish learning progress from noisy training rewards. Review them alongside the learner's losses and the agent's gameplay.

Inspect a checkpoint

The diagnose command runs episodes using a saved model:

uv run python -m experiments.main diagnose Batty ppo /path/to/training-output     --episodes 16

It uses the latest checkpoint by default. Add --checkpoint NUMBER to inspect a specific checkpoint.

Export for inference

Export a trained model to ONNX:

uv run python -m experiments.main export Batty ppo /path/to/training-output

An exported model is intended for choosing actions during inference. Keep the full checkpoints when you want to resume training.

The generated API reference covers checkpoint registration, metric writers and the ONNX exporter.