finish, flush and summary

finish

run.finish(status="finished", *, exit_code=0, error=None)

Flushes, writes an end frame, records the status, exit code and end time in run.json, and compacts the logs into data.parquet when PyArrow is installed. It is safe to call twice.

• It is called at interpreter exit, and when a with block ends.
• An uncaught exception marks the run crashed and stores the traceback as exit.error. KeyboardInterrupt marks it killed.
• If the process is killed outright (SIGKILL, the OOM killer), everything flushed so far stays readable. The run keeps the status running, and the viewer shows it as lost once its heartbeat is two minutes old.

flush

run.flush()

Waits until everything logged so far is on disk. You rarely need it: the background thread writes every flush_interval seconds, and finish flushes.

summary

A dict of the run's scalars so far. Each one maps to last, step, min, max, count and nonfinite. It is empty until something has been logged.

run.summary["train/loss"]["min"]