OpenDriveLab / ST-P3

[ECCV 2022] ST-P3, an end-to-end vision-based autonomous driving framework via spatial-temporal feature learning.
Apache License 2.0
280 stars 34 forks source link

Dear author, train loss value <0 for the perceive task #5

Closed emilyemliyM closed 1 year ago

emilyemliyM commented 1 year ago

I just train the perceieve task, But I got the loss with the value <0? Is that is normal?

charleshsc commented 1 year ago

Yeah, you could refer to the "Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics" paper for more theoretical derivation process.