we should have something that allows us to linearly/exponentially shift the learning rate of the optimizer in the training loop