search

OpenDriveLab / DriveAdapter

[ICCV 2023 Oral] A New Paradigm for End-to-end Autonomous Driving to Alleviate Causal Confusion
Apache License 2.0
202 stars 16 forks source link

only train the adapter but freeze the student model? #4

Closed eson88866 closed 1 year ago

eson88866 commented 1 year ago

Hello, I have some questions. If we use Roach(same BEV) as the teacher model to train the DriveAdapter model, is it necessary to retrain the student model as well? What I mean is, can we also freeze the student model's weights and only train the adapter? Is this a good approach? It seems like it could significantly reduce training time. Assuming my understanding is correct, would using other teacher models like PlanT result in differences in the student model's perception segmentation task?

I really like your research. Thank you very much.

jiaxiaosong1002 commented 1 year ago

Hi, thanks for your interests.

We have not tried any of your idea.

However, your understand is correct. One significant advantage of the DriveAdapter paradigm is that it decouples the perception, planning, and sim2real part which student, teacher and adapter are responsible for respectively while still maintaining the end-to-end ability.

Thus, you are welcomed to try your idea and report results here. We are glad to hear about more.