SLACK: Stable Learning of Augmentations with Cold-start and KL regularization