Cindy XL
Cindy XL
Same. I am wondering how to fix it.
How did you solve it?
The unexpected nan values usually mean numerical instability issues. Have you tried reducing BASE_LR and/or batch size?
One of the authors here. Can you try updating all the dependencies to the latest version? And yes, training can be slow. There is potential room for optimization.
Seems resolved according to https://github.com/htcr/sam_road/issues/40.
As of today, these Google Drive links in the README still work. SpaceNet https://drive.google.com/uc?id=1FiZVkEEEVir_iUJpEH5NQunrtlG0Ff1W CityScale https://drive.google.com/uc?id=1R8sI1RmFe3rUfWMQaOfsYlBDHpQxFH-H Which dataset link are you referring to?
Thank you for trying SAM-Road! Yes, I think you need to adjust params for your 512x512 dataset- Buffer Size: The TOPO metric uses a buffer around roads to determine matches....
> @TCXX The list is 8 items long, not sure what you mean by volunteers. We're all volunteers here. Haha thanks. I have a list of 30+ and will add...
All suggestions have been committed.