search

strongwolf / DW

A Dual Weighting Label Assignment Scheme for Object Detection
Apache License 2.0
136 stars 17 forks source link

Training in Custom Dataset #4

Closed BlueBlueFF closed 2 years ago

BlueBlueFF commented 2 years ago

When I training in my custom dataset, - mmdet - INFO - Epoch [24][1350/1388] lr: 1.000e-04, eta: 0:00:57, time: 1.529, data_time: 0.023, memory: 7116, loss_cls_pos: 0.0950, loss_loc: 0.2806, loss_cls_neg: 0.1053, loss: 0.4809 the loss_cls_pos seems small than loss_cls_neg, how to only enlarge loss_cls_pos?

strongwolf commented 2 years ago

Is the loss_cls_pos smaller than the loss_cls_neg at the beginning of training or only at the late epochs? I am not sure whether this phenomenon is normal and I think you can have a judgment by checking the mAP.