simulo.Trainer

The training-orchestration contract — run, evaluate, and persist a training job.

A trainer owns the RL loop: it drives an environment, updates the policy, and reads/writes checkpoints. Structural mirror of simulo.core.trainer.Trainer.

train()

Trainer.train(max_iterations: int) -> Dict[str, Any]

Run the training loop for up to max_iterations and return summary metrics.

The returned mapping includes at least the final iteration count and reward statistics (implementation-defined keys).

evaluate()

Trainer.evaluate(checkpoint: Optional[str] = ..., num_episodes: int = ...) -> Dict[str, Any]

Roll out the current (or checkpoint-loaded) policy without learning.

Returns evaluation metrics such as mean episode return over num_episodes episodes.

save()

Trainer.save(path: str) -> None

Write the trainer’s full state (policy + optimizer) to path.

load()

Trainer.load(path: str) -> None

Restore trainer state previously written by save().

---

Note

simulo.Trainer is a mode-aware name: at submit time it resolves to a lean, torch-free stand-in so your app imports instantly; on the GPU worker it resolves to the real runtime class. The contract below is the definitive list of members every simulo.Trainer provides, and the contract your own subclass fulfils.

Declared as:

@runtime_checkable
class TrainerProtocol(Protocol)