Conference paperDemystifying Learning Rate Policies for High Accuracy Training of Deep Neural Networks