Regularizing Class-wise Predictions via Self-knowledge Distillation