Calibrating a Deep Neural Network with Its Predecessors