Explainable Learning Rate Regimes for Stochastic Optimization