search

Dootmaan / DSRL

Implementation of CVPR 2020 Dual Super-Resolution Learning for Semantic Segmentation
MIT License
95 stars 17 forks source link

about half the amount of fea_seg are zero #8

Open nickin1023 opened 3 years ago

nickin1023 commented 3 years ago

Hi there,

thank you for your works. i have git clone your code, and i have cancel the comment of line 16, 18, 23, 25 in 'util/fa_loss.py' and run 'train.py' with cityscapes. Then, i get NaN or Inf caused by feature of deeplabv3. About half amount of it are zero. What should i do for preparation?

Thank you.

Dootmaan commented 3 years ago

Hi there,

thank you for your works. i have git clone your code, and i have cancel the comment of line 16, 18, 23, 25 in 'util/fa_loss.py' and run 'train.py' with cityscapes. Then, i get NaN or Inf caused by feature of deeplabv3. About half amount of it are zero. What should i do for preparation?

Thank you.

Hi there, thank you for your question. I have also encountered NaN at the beginning of this experiment and I found it may be caused by the normalization in FA module, so as you can see this repo doesn't use normalization by default. I'd like to know that does this problem still exist when you are not using normalization or when you switch to using subpixel branch?

nickin1023 commented 3 years ago

Thank you for your help! Sorry, I misunderstood that to cancel the comment of that lines required to use FA module. When I do not use normalization, I could run your code.