search

geoelements / gns

Graph Network Simulator
https://www.geoelements.org/gns/
Other
124 stars 30 forks source link

Training issue with noise_std #26

Closed qilinli closed 2 months ago

qilinli commented 2 years ago

Hey guys,

Many thanks for sharing a Pytorch version of GNS. It makes life much easier, compared to using the official tf version.

I'm currently using this code to train on my own dataset and one strange thing is that the training loss (one-step prediction) degrade quicker with higher noise_std and does not converge with noise_std=0, which is weird. My understanding is that this should only happen to the rollout prediction where more noise means more stable prediction.

Do you have similar observations or any clue on this?

arjun-mani commented 2 years ago

Some thoughts here based on my experience -- this might have to do with normalization (value of noise_std affects magnitude of targets). However, it should still converge with noise_std=0, that's unusual.