OptimEngine.fit()

OptimEngine.fit(*, checkpoint=None, pause_after=None, checkpoint_every=10, allow_version_mismatch=False)

Runs optimization and returns processed results.

Parameters:

• checkpoint (OptimCheckpoint | str | PathLike[str] | None, default: None) – None starts fresh in memory. An OptimCheckpoint resumes in memory. A path selects a persistent run: load it if present, or start fresh if absent, and save subsequent checkpoints there. The parent directory must exist. Only load trusted checkpoint files.

• pause_after (int | None, default: None) – Maximum additional epochs for this call. Pausing preserves optimizer state in result.checkpoint. The stopper still owns the total epoch budget; change engine.stopper.epochs to extend it explicitly.

• checkpoint_every (int, default: 10) – Save every this many completed epochs when a path is supplied, by default 10. Also save at a deliberate pause or normal completion. A NaN failure leaves the last saved file intact.

• allow_version_mismatch (bool, default: False) – Attempt recovery despite differing Liesel, JAX, jaxlib, Optax, or NumPy versions, issuing a warning. Structural checks still apply.

Return type:

OptimResult

Returns:

OptimResult – Processed history, final and best-monitor positions, monitoring source, cumulative active runtime, status, and an independent checkpoint for continuation (None on NaN failure).

Notes

Reconstruct the same model, data, and optimizer settings before resuming. Parameter and observed-variable names, shapes, and dtypes must agree. Compatibility checks cannot detect changed data or learning rates. Resuming a checkpoint that permits no further epochs under the current stopper settings issues a warning and returns its result. Use a new path or call fit() without a checkpoint to start a fresh run. Extending the epoch budget permits continuation only if early stopping does not apply. A failed checkpoint write raises and preserves the previous file. Interruptions recover from the last successful periodic save.