Early stoppers

The EarlyStopper implementations decide whether Trainer should end before the configured number of epochs. The trainer checks the selected stopper after each available training or validation score.

Available implementations

none or no (NoStopper) never stop early.
basic (BasicStopper) stops after a configured number of epochs without meaningful improvement.
time (TimeStopper) stops after a wall clock time limit (useful for cluster computation).

Common configuration

All implementations accept restore_best (default true). When restore_best is enabled, the trainer restores the best saved model after training stops. Every train() call begins with start(), which resets initial values. Stopper state is therefore not carried across training calls and is not stored in checkpoints.

Example

trainer:
  early_stopper:
    early_stopper_type: basic
    patience: 25
    min_delta: 0.001
    restore_best: true

API reference

See also xanesnet.stoppers and xanesnet/schemas/stoppers/ for complete fields and defaults.