search

fudan-zvg / PolarFormer

[AAAI 2023] PolarFormer: Multi-camera 3D Object Detection with Polar Transformers
MIT License
160 stars 15 forks source link

How much the loss can be convergence？ #10

Closed mayechi closed 1 year ago

mayechi commented 1 year ago

I have trained the vovnet version for 18 epcochs, but the total loss is still about 12.

yanqinJiang commented 1 year ago

Sorry, I cannot find the log of polarformer-vovnet. But the log of polarformer-r101-w/o-bev_aug indicates that the loss converges at around 3.8 after 24 epochs, and I remember polarformer-vovnet converges much faster and ends up with a lower value than polarformer-r101, maybe around 3 or below 3. (The vovnet backbone I use was pretrained on nuScenes trainval set #8 , so polarformer-vovnet is supposed to converge faster.)

Maybe you can share your training log here and I will check if there is anything wrong. : )