search

lhoyer / DAFormer

[CVPR22] Official Implementation of DAFormer: Improving Network Architectures and Training Strategies for Domain-Adaptive Semantic Segmentation
Other
470 stars 93 forks source link

src.loss_imnet_feat_dist is nan #17

Closed Haochen-Wang409 closed 2 years ago

Haochen-Wang409 commented 2 years ago

Hi, thanks for your excellent work! However, when I tried to train DAFormer on GTA5 to Cityscapes benchmark, I found src.loss_imnet_feat_dist is nan while training. Since the final result is expected (68.3 mIoU), I'm a little bit confused.

lhoyer commented 2 years ago

Thanks for your interest in our work! Please, have a look at issue #11. It should clarify your question.

Haochen-Wang409 commented 2 years ago

Thanks a lot for your answer!

kaigelee commented 2 years ago

Hi, thanks for your excellent work! However, when I tried to train DAFormer on GTA5 to Cityscapes benchmark, I found src.loss_imnet_feat_dist is nan while training. Since the final result is expected (68.3 mIoU), I'm a little bit confused.

My decode.loss_seg is also nan, could you help me fix it?

lhoyer commented 2 years ago

Hi, thanks for your excellent work! However, when I tried to train DAFormer on GTA5 to Cityscapes benchmark, I found src.loss_imnet_feat_dist is nan while training. Since the final result is expected (68.3 mIoU), I'm a little bit confused.

My decode.loss_seg is also nan, could you help me fix it?

This is a separate issue. Please continue the discussion in issue #22.