Work out early stopping patience instantly with clear inputs, formula shown and shareable results.
Early stopping halts training when the validation metric has not improved by more than a minimum delta for a set number of epochs. Patience must be long enough to survive the plateaus that precede a learning-rate drop, but every epoch after the best checkpoint is compute spent on a model you will discard — which is why patience is tuned against epoch cost, not chosen arbitrarily.
Early stopping
epochs remaining = patience - epochs since improvement; wasted compute = epochs since improvement x epoch time
Longer than the interval between learning-rate reductions, since each drop often produces a fresh improvement. With a step schedule every 30 epochs, patience of 10 will stop too early.
The one with the best validation metric, not the final epoch. Set restore_best_weights or equivalent, otherwise you ship the worse model.