Enhancing Visual Domain Adaptation with Source Preparation