CaT: Coaching a Teachable Student

Coaching a Teachable Student
Jimuyang Zhang, Zanming Huang, Eshed Ohn-Bar
CVPR 2023

Introduction

This repository contains the code for CaT: Coaching a Teachable Student.

CaT is a deep knowledge distillation framework for effectively teaching a student agent to drive from the supervision of a privileged teacher agent.

TODO List (comming soon)

• Data collection Code
• Training Code
• Teacher Evaluation Code
• Student Evaluation Code
• Teacher Model Checkpoint
• Student Model Checkpoint

Getting Started

• Please follow Installation to get started.

Evaluation

To run evaluation, first start the CARLA server

cd CARLA_ROOT
./CarlaUE4.sh

Student Evaluation

Make sure all environment varibles are setup correctly.then run

ROUTES=[Path to Route File] bash ./leaderboard/scripts/run_evaluation.sh

where ROUTES points to a route xml file (e.g. leaderboard/data/routes_devtest.xml) .

Teacher Evaluation

To evaluate teacher model on the Longest6 benchmark, run the following after changing the environment varibles accordinly in the script

bash ./leaderboard/scripts/local_evaluation.sh

The teacher model checkpoint can be found here

Citation

If you find this repo useful, please cite

@inproceedings{zhang2023coaching,
        title={Coaching a Teachable Student},
        author={Zhang, Jimuyang and Huang, Zanming and Ohn-Bar, Eshed},
        booktitle={CVPR},
        year={2023}
}

Acknowledgements

Our code references the following repositories:

• CARLA Leaderboard
• Scenario Runner
• Persformer
• LAV