Deep Video Generation, Prediction and Completion of Human Action Sequences