search

MCZhi / GameFormer-Planner

[ICCV & CVPR Workshop] Learning-enabled Interactive Prediction and Planning Framework for Autonomous Vehicles
https://mczhi.github.io/GameFormer/
MIT License
143 stars 17 forks source link

about interaction loss #8

Closed fshi2006 closed 9 months ago

fshi2006 commented 9 months ago

hello! In the paper, you mentioned using interaction loss to ensure security between level k and k-1. But it seems that this part is not included in the code, can I get the code of this part? Thx a lot!

MCZhi commented 9 months ago

Hi, @fshi2006. Please refer to the code in the main repo GameFormer.

fshi2006 commented 9 months ago

thx a lot! So why didn't use this loss in planner repo?

MCZhi commented 9 months ago

This is because adding the interaction loss would significantly slow the training. Given the large amount of data from the nuPlan dataset, we need to remove this loss to speed up the training and thus train on more data.