Optimizing Deep Learning Models with Xaver Initialization and Average Magnitude Scaling Factor in MxNet