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

yolov4-csp pre-trained weights gave much worse mAP #7362

Open ghost opened 3 years ago

ghost commented 3 years ago

I wasn't paying attention and I trained yolov4-csp with yolov4 pre-trained weights yolov4.conv.137 and I got mAP@0.5=52% which is acceptable with my dataset (its a pretty challenging dataset) and when I noticed my mistake I got excited and thought surely if I change the pre-trained weights to yolov4-csp.conv.142 I'll get better results. With exactly the same cfg file I got mAP@0.5=6%

I'm very confused because I was basically using the wrong values and when I corrected them I got worse results. Can someone explain this phenomenon please ?

arnaud-nt2i commented 3 years ago

initial weight "influence" is lost after every iteration, you could start training from random weight (by providing no weights file) and get the same final mAp (but probably after more iter). Initial weights allow faster convergence, fewer chances of crash and possibly better result.