Adversarial Framework for Unsupervised Learning of Motion Dynamics in Videos