Update: Note this is a reimplementation based on the Keras implementation (original repo) from one author of the paper. Please create issues in the original repo of the author if you have questions about performance, theory, differences between the implementation and the paper, etc.
PyTorch implementation of the example for the paper:
Alex Kendall, Yarin Gal, and Roberto Cipolla. "Multi-task learning using uncertainty to weigh losses for scene geometry and semantics." In CVPR 2018. [arXiv].