how to train with the diode dataset

thanks for your great work，but I have a question that how can I train with the dataset while there are many zeros in the depth map, in NYUv2 dataset, there is no zero in depthmap, and how should I deal with the pixels without depth? thanks again