Training doesn't converge for Mapillary Vistas Dataset training with MaskRCNN

For above log, please refer to below topic and its solution.