Find bottlenecks in the workflow to find points to optimize - Maybe using "tf.profiler" or "cProfile"
