search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.59k stars 7.95k forks source link

Hard to converge when training yolov4-csp-x-swish #8020

Open erikguo opened 2 years ago

erikguo commented 2 years ago

When training yolov4-csp-x-swish on my database, the avg-loss curve always becomes to a U curve. However, when I change the "objectness_smooth" to 0 in yolo layer, it can converge normally now . Does somebody have same problem? I'm analyzing the source code about objectness_smooth now. Didn't find out the key cause.

midasklr commented 2 years ago

I trained v4-csp-x-swish with my onw dataset and got avg-loss around 2.3 . but test with image and got nothing, have you met the same problem?

erikguo commented 2 years ago

almost the same problem.

ashutoshIITK commented 2 years ago

@midasklr I am also facing the same problem. Loss is really low around training but during inference, there is no output.