Skip to content

EarlyStopping — Training Stopping Callback

Stops training when a monitored metric stops improving.


Overview

EarlyStopping monitors a metric (typically val_loss), tracks the best value seen so far, and stops training after patience epochs without improvement. It can optionally restore the best observed weights when stopping.


Key Features

  • Monitor any metric by name (monitor)
  • Configurable patience (patience) and minimum delta (min_delta)
  • Optional best weight restoration (restore_best_weights)
  • Verbose logging support

Parameters

Parameter Type Default Description
monitor str "val_loss" Metric name to watch
patience int 5 Epochs without improvement before stopping
min_delta float 0.0 Minimum improvement required
restore_best_weights bool True Restore best weights when stopping

Methods

on_epoch_end(epoch, logs=None)

Inspect logs; stop if metric hasn't improved for patience epochs.


Example Usage

from machinegnostics.magnet import EarlyStopping
stopper = EarlyStopping(monitor="val_loss", patience=3)

Author: Nirmal Parmar