training, validation, and test sets

The three disjoint pieces a labeled dataset is split into before training, each with its own job.

A labeled image dataset is split into three disjoint pieces before training starts. The training set is what the model's weights actually learn from. The validation set is held out, never trained on, and checked periodically during training to catch overfitting. The test set is held out most strictly of all and checked exactly once, after every tuning decision is final, to report an honest number.

Examples

A 2,000-image dataset split 70%/15%/15% gives 1,400 training images, 300 validation images, and 300 test images.