YOLOv4-Mish Loss very high and negative iou.

[andrelapian]

this is chart for pre-trained models using yolov4x-mish

This is chart for pre-trained models using yolov4

My question is why yolov4-mish give high loss even though I use the same dataset? and why yolov4-mish give negative value for iou_loss?

v3 (iou loss, Normalizer: (iou: 0.05, obj: 4.00, cls: 0.50) Region 168 Avg (IOU: 0.136990), count: 2, class_loss = 39.872097, iou_loss = -29.828011, total_loss = 10.044086 v3 (iou loss, Normalizer: (iou: 0.05, obj: 1.00, cls: 0.50) Region 185 Avg (IOU: 0.421453), count: 16, class_loss = 5.155934, iou_loss = 0.427526, total_loss = 5.583461 v3 (iou loss, Normalizer: (iou: 0.05, obj: 0.40, cls: 0.50) Region 202 Avg (IOU: 0.360355), count: 20, class_loss = 1.842278, iou_loss = 2.920083, total_loss = 4.762361 total_bbox = 257324, rewritten_bbox = 0.000000 % v3 (iou loss, Normalizer: (iou: 0.05, obj: 4.00, cls: 0.50) Region 168 Avg (IOU: 0.000000), count: 1, class_loss = 2.948591, iou_loss = -2.211443, total_loss = 0.737148 v3 (iou loss, Normalizer: (iou: 0.05, obj: 1.00, cls: 0.50) Region 185 Avg (IOU: 0.449434), count: 17, class_loss = 4.662511, iou_loss = 0.263417, total_loss = 4.925928 v3 (iou loss, Normalizer: (iou: 0.05, obj: 0.40, cls: 0.50) Region 202 Avg (IOU: 0.467637), count: 27, class_loss = 2.409369, iou_loss = 3.892585, total_loss = 6.301954 total_bbox = 257368, rewritten_bbox = 0.000000 % v3 (iou loss, Normalizer: (iou: 0.05, obj: 4.00, cls: 0.50) Region 168 Avg (IOU: 0.079922), count: 1, class_loss = 21.372038, iou_loss = -15.975222, total_loss = 5.396816 v3 (iou loss, Normalizer: (iou: 0.05, obj: 1.00, cls: 0.50) Region 185 Avg (IOU: 0.433072), count: 11, class_loss = 2.610621, iou_loss = 0.336940, total_loss = 2.947561 v3 (iou loss, Normalizer: (iou: 0.05, obj: 0.40, cls: 0.50) Region 202 Avg (IOU: 0.389636), count: 18, class_loss = 1.677491, iou_loss = 2.662407, total_loss = 4.339898 total_bbox = 257398, rewritten_bbox = 0.000000 %

[molyswu]

same issues

[akashAD98]

same issue. @WongKinYiu @AlexeyAB your help is really appreciate.