Optimizing Tensorflow Models For Serving