search

NVIDIA / semantic-segmentation

Nvidia Semantic Segmentation monorepo
BSD 3-Clause "New" or "Revised" License
1.76k stars 388 forks source link

Loss value is negative #155

Open ed-Im opened 2 years ago

ed-Im commented 2 years ago

image

I am learning by custom dataset. The problem is that the loss value becomes negative after a few iterations after learning. May I know the cause of your prediction?

ajtao commented 2 years ago

It's natural for RMI loss to go negative: https://github.com/ZJULearning/RMI/issues/5

But there's another issue which is that you're out of memory in the GPU. Try to shrink the crop size to something like 800