The ce_loss became negative when I was using the mask2former to do instance segmentation

command you run:
python train_net.py --num-gpus 8 --config-file configs/coco/instance-segmentation/swi
n/maskformer2_swin_large_IN21k_384_bs16_100ep.yaml MODEL.WEIGHTS "weig
hts/model.pkl"

The data form used in my case is coco.
And the registration of this dataset was strictly followed with the document of our office

Start by reproducing the command in train_net.py with configs/coco/instance-segmentation/swin/maskformer2_swin_large_IN21k_384_bs16_100ep.yaml and weights/model.pkl. Inspect the training output around ce_loss and compare the COCO dataset registration with the referenced documentation. Done means identifying the cause of the negative loss and documenting a verified correction.

Written by the indexing model from the issue text.