Repo is forked from pytorch-deeplab-xception, then modified for the Kaggle challenge.
Pytorch 1.5 + CUDA10.2 on Windows 10 with RTX 2070. (Windows 10 preservses 2G GPU memory for system use so that I only have 6G available GPU memory. Recommend Linux.)
Got 0.99592 with ensembling 2 models.