Approximate Heavily-ConstrainedLearningwith LagrangeMultiplierModels