Closed Yzichen closed 3 weeks ago
We didn't try this trick. Maybe the lr of backbone multiplied by 0.1 can get better performance.
May I ask if you modified it on solofusion? When training the model used for submission (VoVNet backbone), how many epochs did you train and did you use CBGS (number of iterations times 4.5)?
May I ask if you modified it on solofusion? When training the model used for submission (VoVNet backbone), how many epochs did you train and did you use CBGS (number of iterations times 4.5)?
60 epochs without CBGS, other training settings are same with BEVDet
Does backbone's learning rate need to be adjusted when using pre-trained weights? For example, multiply by 0.2.