Epoch
One complete pass through the entire training dataset during model training. Epoch count serves as a fundamental measure of training progress and convergence.
An epoch represents one full pass over the entire training dataset. For ImageNet's 1.28 million images with batch size 256, one epoch is approximately 5,000 iterations. Typical image classification models require 90 to 300 epochs to converge.
Epoch count is a hyperparameter controlling training sufficiency. Too few epochs cause underfitting where patterns are not captured. Too many lead to overfitting where the model memorizes training examples at the expense of generalization.
- Learning curves: Plotting training and validation loss per epoch reveals model behavior. When training loss decreases while validation loss rises, overfitting has begun
- Early stopping: Terminates training when validation loss fails to improve for a set number of epochs (patience, e.g., 10), preventing unnecessary computation and overfitting
- Learning rate scheduling: Reduces learning rate as epochs progress. ResNet training uses step decay, dividing by 10 every 30 epochs for fine-grained convergence
Transfer learning reduces required epochs dramatically: fine-tuning pre-trained models typically needs only 10-30 epochs. Data augmentation increases effective dataset size per epoch, improving efficiency and reducing total epochs needed.