search

Sranc3 / M-BEV

[AAAI24]This is the implementation for the paper M-BEV: Masked BEV Perception for Robust Autonomous Driving
26 stars 2 forks source link

Training schedule #3

Closed konyul closed 4 months ago

konyul commented 5 months ago

Thank you for sharing such a great work! I have some questions for the training details.

  1. In the paper, 48 epochs are consumed to train MVR, however, in the github, 24 epoch are consumed for finetuning. which one is right?
  2. how many epochs do you consume for pretraining?
Sranc3 commented 4 months ago

Thanks for your interest. For one missing camera, finetuning for 24 epochs could achieve most of the performance, and for 48 epochs the performance will be saturated. So you can decide how many epochs you want to train with time cost in mind. Note, for more missing views, more training epochs are recommended. In the experiment, the pretraining time is too long while the improvement is limited compared to finetuning, and we suggest to have a short pretraining for rough initialization or go through the finetuning with mvr loss supervision directly. You can have your own try and find the suitable schedule according to your demand.

konyul commented 4 months ago

Thank you for the answer, I am satisfied with your answer.