Semantic segmentation loss is stuck, what's the reason?

Viewed 18

I am doing simple tasks on Kaggle, and I came across one problem that I don't know how to explain. See my GitHub repo on this task. Basically, my problem is that the loss is stuck at a certain value and it's not going down. It can wiggle around that value, but no serious improvements. What could be the reason for this?

Also see a screenshot of training to get an idea of what happens to the loss. enter image description here

If you decide to replicate the code and it doesn't work, please let me know (although it should work properly).

0 Answers
Related