Training Over-parameterized Models with Non-decomposable Objectives