Dear all,

training yolov4 with my own dataset shows final mAP values around 75% and an almost monotonic loss decay. I am very happy with the results.
However, I wanted to check if I can get even better results with yolov7. Using the same dataset, I see a minimum in the training loss at around 600 iterations, and then the loss settles at a higher value. Nevertheless, the mAP values are very good and approach 77%.

Some additional information:

• My dataset has one class only. I followed the README to adapt the config files yolov4-custom.cfg and yolov7.cfg accordingly
• I used the pretrained weights yolov4.conv.137 and yolov7.conv.132. The minimum in the loss curve is still there if I start training from scratch.
• I used the following commands for training:

./darknet detector train data/obj.data yolov4-custom.cfg yolov4.conv.137 -dont_show -map -mAP_epochs 1
./darknet detector train data/obj.data yolov7.cfg yolov7.conv.132 -dont_show -map -mAP_epochs 1

• I have played around with learning rate, decay, momentum, and burn-in, but this only shifts the minimum horizontally and/or makes it wider.

Does anyone have a clue why the training loss curves of yolov4 and yolov7 look so differently? The minimum in the loss curve really bothers me.

Thanks for your help!