search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.68k stars 7.96k forks source link

Transfer-Learning #5529

Open buimanhlinh96 opened 4 years ago

buimanhlinh96 commented 4 years ago

Hi AlexeyAB, thanks for your contributions. I am using Yolo-v4 for my custom object detection task. I used transfer-learning based on your pre-train model. However, I struggle to understand the value of stopbackward in this version and where i need to put it in my cfg file. I read some issues but i dont actually have a deep understand about the value (800, 2000, 6000, etc.). Can you explain it for me? Thanks for your support!

AlexeyAB commented 4 years ago

stopbackward=800 - for some layer means that all the layers' weights before the current layer will not be updated only for the first 800 iterations.

buimanhlinh96 commented 4 years ago

Yeah, i got it! Great parameter for us. To facilicate training strategy, can I put more than one 'stopbackward' in my cfg file? So it will work bottom-up?