Stochastic and Adversarial Online Learning without Hyperparameters