8.3. Early Stopping

Early stopping ends training as soon as the validation metric stops improving, which avoids overfitting and saves training time. New in 1.5.0, it is on by default for every time series task type (classification, regression, anomaly detection, forecasting) and for image classification.

8.3.1. Configuration

The three keys live in the training section of the config YAML:

Key	Default	Meaning
`early_stopping`	`True`	Enable or disable early stopping.
`early_stopping_patience`	`5`	Number of consecutive epochs without improvement before training stops.
`early_stopping_min_delta`	`0.0`	Minimum change in the monitored metric that counts as an improvement.

training:
  training_epochs: 100
  early_stopping: True
  early_stopping_patience: 5
  early_stopping_min_delta: 0.0

To always train for the full number of epochs, set early_stopping: False.

8.3.2. How it works

After each validation pass the training script reports the monitored metric. An epoch counts as an improvement only if the metric is better than the best value so far by more than early_stopping_min_delta. Otherwise the counter of bad epochs increases, and when it reaches early_stopping_patience training stops and the log shows Early stopping at epoch N: no improvement for <patience> epochs. The best model found so far is the one that is kept, as in a normal run.

The monitored metric depends on the task type:

Task type	Monitored metric	Better when
Classification (time series, image)	Validation accuracy	higher
Regression	Validation MSE	lower
Anomaly detection	Validation MSE (reconstruction error)	lower
Forecasting	Validation SMAPE	lower

Tip

With a large training_epochs value, the default patience of 5 is usually enough. If your validation curve is noisy (small datasets), raise early_stopping_patience or set a small early_stopping_min_delta instead of disabling early stopping.

Note

The audio and radar classification training scripts in 1.5.0 accept the same three keys in their parameter set but do not apply early stopping in their training loop.

See How to Create a ModelZoo Config YAML for the rest of the training keys.