cyclical learning rates (clr)

This method allows the model to start learning with a small learning rate, which gradually increases cyclically, leading to faster convergence during training.