search

lucifer443 / RecurrentBEV

[ECCV 2024] RecurrentBEV: A Long-term Temporal Fusion Framework for Multi-view 3D Detection
Apache License 2.0
17 stars 1 forks source link

Temporal training takes much longer time than single-frame warmup epochs. #3

Closed synsin0 closed 2 weeks ago

synsin0 commented 2 weeks ago

I train the 256x704 config with 8 L20 and finds that temporal training (after 6 epochs) takes much longer time than single-frame warmup epochs (before 6 epochs). I attach my training log here. Is that normal for this training speed? 20241002_223820.log

lucifer443 commented 2 weeks ago

It is normal for the temporal model to be slower because of the need to load and calculate historical information.