the normal training loss profile

[Wolfybox]

hi, author. I have tried training MapTRv2 with the given three auxiliary tasks: depth, pv_seg, bev_seg. However, after training for 24 epochs, I observed that the loss of these auxiliary tasks only slightly decrease. Their loss curves also show strong fluentuation.

loss_depth: from 6.8 to 6.64 loss_pv_seg: from 0.265 to 0.215 loss_bev_seg: from 0.138 to 0.118

Note that the depth loss is still large at the end.

Other losses also look potentially unconverged. loss_pts at the last dec layer: from 12 to 4 loss_cls at the last dec layer: 0.65 to 0.35

I wonder whether the above phenomenons are normal for training.