search

liyunsheng13 / BDL

MIT License
222 stars 30 forks source link

problem about loss become nan #46

Closed du841021 closed 3 years ago

du841021 commented 3 years ago

hi,thanks for the code I'm, trying to train GTA2Cityspace on VGG and DeepLab. but no mater with or without self-supervised, after 2 iteration src seg loss will become nan. can you tell me how to fix it? thank you!

my environment: CUDA Version: 11.1, pytorch-gpu:1.2.0 and here is my log on training self-supervised VGG: 擷取選取區域_066

liyunsheng13 commented 3 years ago

I'm not quite sure. Based on the hyper parameters you provide, it should be fine.