A more efficient way to infer with lower memory requirements.