EarlyStoppingHook

class mmengine.hooks.EarlyStoppingHook(monitor, rule=None, min_delta=0.1, strict=False, check_finite=True, patience=5, stopping_threshold=None)

Early stop the training when the monitored metric reached a plateau.

Parameters:

• monitor (str) – The monitored metric key to decide early stopping.

• rule (str, optional) – Comparison rule. Options are ‘greater’, ‘less’. Defaults to None.

• min_delta (float, optional) – Minimum difference to continue the training. Defaults to 0.01.

• strict (bool, optional) – Whether to crash the training when monitor is not found in the metrics. Defaults to False.

• check_finite (bool) – Whether to stop training when the monitor becomes NaN or infinite. Defaults to True.

• patience (int, optional) – The times of validation with no improvement after which training will be stopped. Defaults to 5.

• stopping_threshold (float, optional) – Stop training immediately once the monitored quantity reaches this threshold. Defaults to None.

Note

New in version 0.7.0.

after_val_epoch(runner, metrics)

Decide whether to stop the training process.

Parameters:

• runner (Runner) – The runner of the training process.

• metrics (dict) – Evaluation results of all metrics

before_run(runner)

Check stop_training variable in runner.train_loop.

Parameters:

runner (Runner) – The runner of the training process.

Return type:

None