search

kuangliu / pytorch-retinanet

RetinaNet in PyTorch
992 stars 250 forks source link

cls_loss is NAN #47

Closed Brozhao closed 5 years ago

Brozhao commented 5 years ago

loc_loss: 0.096 | cls_loss: 49.812 | train_loss: 49.907 | avg_loss: 312.096 loc_loss: 0.102 | cls_loss: 34.575 | train_loss: 34.676 | avg_loss: 292.280 loc_loss: 0.093 | cls_loss: 22.649 | train_loss: 22.742 | avg_loss: 274.311 loc_loss: 0.095 | cls_loss: 603.826 | train_loss: 603.921 | avg_loss: 294.912 loc_loss: 0.097 | cls_loss: 96.089 | train_loss: 96.186 | avg_loss: 283.222 loc_loss: 0.100 | cls_loss: 268.645 | train_loss: 268.745 | avg_loss: 282.418 loc_loss: 0.246 | cls_loss: 185.135 | train_loss: 185.381 | avg_loss: 277.310 loc_loss: 0.574 | cls_loss: 3271.793 | train_loss: 3272.367 | avg_loss: 427.063 loc_loss: 14.899 | cls_loss: 4664092.368 | train_loss: 4664107.500 | avg_loss: 222507.084 loc_loss: 0.096 | cls_loss: 1018231725413.010 | train_loss: 1018231717888.000 | avg_loss: 46283472297.126 loc_loss: 0.094 | cls_loss: 9635243732500480.000 | train_loss: 9635243732500480.000 | avg_loss: 418967911690913.750 loc_loss: 0.096 | cls_loss: 35978558609519793754739982653718528.000 | train_loss: 35978558370880749703701422304919552.000 | avg_loss: 1499106598786698000397684646608896.000 loc_loss: 0.101 | cls_loss: nan | train_loss: nan | avg_loss: nan loc_loss: 0.101 | cls_loss: nan | train_loss: nan | avg_loss: nan loc_loss: 0.093 | cls_loss: nan | train_loss: nan | avg_loss: nan loc_loss: 0.102 | cls_loss: nan | train_loss: nan | avg_loss: nan loc_loss: 0.118 | cls_loss: nan | train_loss: nan | avg_loss: nan loc_loss: 0.102 | cls_loss: nan | train_loss: nan | avg_loss: nan loc_loss: 0.114 | cls_loss: nan | train_loss: nan | avg_loss: nan

Brozhao commented 5 years ago

how can I solve that would anyone help me

banjuanshua commented 5 years ago

I have fixed this bug, you can see it in my pull request.

Brozhao commented 5 years ago

This problem has been solved according to your method.Thank you！

Brozhao commented 5 years ago

I have fixed this bug, you can see it in my pull request.

NAN of the loss haves never appeared,but the mAP is almost zero.