obj loss increases during the training

[liamsun2019]

Hi author, Appreicate sharing your excellent work. My use case:

1. one class: person
2. format the label file with yolo style, e.g: 0 0.76328125 0.5339578454332553 0.0546875 0.1592505854800937 where the 5 elements are: clsid, centerx, centery, boxw, boxh, all of which are normalized against original image width/height
3. All the train/val images are filtered out from coco train2017/val2017 with category being "person"
4. Modify flag_sets.py, objectBox_COCO.yaml and coco.yaml, especially set nc to 1 for my case
5. The training shows that the obj loss increases while box loss decreases.

I am curious how this could happen. Any suggestions? Thanks.

[MohsenZand]

Thank you! Check the ground-truths for generated training samples ('train_batchX.jpg') in your project folder to ensure that bounding boxes are only and accurately for 'persons'. In 'config/hyp.yaml', you should additionally tweak the related hyperparameters (such as obj and obj_pw).