Optimal Low Rank Tensorflow