Why use train and validation datasets both for training?

[Songyeyaosong]

Normally we train with a training set, then tune parameters with a validation set, and finally test generalization performance with a test set? So... I donot understand why.