search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.65k stars 7.96k forks source link

training yolov4-csp-swish broken #8340

Open spasserkongen opened 2 years ago

spasserkongen commented 2 years ago

When I train the yolov4-csp-swish.cfg (default settings, except setting classes=1 and filters=18 for the previous layer), my loss quickly converges to almost zero. But for the test map calculation I get zero. When I train a yolov4.cfg using the very same data I get a "normal" loss curve and a corresponding map-calculation.

johndongus commented 2 years ago

This is quite old, but i also have a myriad of issues when it comes to training CSP-swish weights. In my case, training results in a non-changed MAP no matter the amount of time trained. It can be 1000 iterations or 10000 iterations and MAP would still be the same.

To this day i have not been able to successfully train csp-swish