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

Still yolov4 training #5549

Open wdxybhb opened 4 years ago

wdxybhb commented 4 years ago

Hello Alexey, I trained my net based on coco&PAS dataset via using the yolov4-custom.cfg and I changed the class number and filter size in it as well. I found the avg loss had pretty slow decreasing speed and stayed flattened while having large fluctuation around value 3.0 in 15k iterations, not like that yolov3 training in which avg loss curve had an obvious decrease to 1.5 in 15k iterations.

I am using two 1080 ti GPUs, with width=416 batch=64 subdivisions=16.

Can you please tell me where is wrong with my training when you are free?

Thank you so much!

AlexeyAB commented 4 years ago

the avg loss had pretty slow decreasing speed and stayed flattened while having large fluctuation around value 3.0 in 15k iterations,

This is normal.