search

pjreddie / darknet

Convolutional Neural Networks
http://pjreddie.com/darknet/
Other
25.69k stars 21.33k forks source link

Speed-Up Yolov3-Tiny #1899

Open ntemperli opened 4 years ago

ntemperli commented 4 years ago

Hi all

I successfully used yolov3-tiny to train with my own dataset for detection of traffic cones. I'm new to darknet. Is it possible to make the weights file smaller, so that the detection speed is higher? If so, how does a sample cfg file look like and what are the training commands?

Thank you very much for your help

AlexeyAB commented 4 years ago

Try to train this cfg-file https://raw.githubusercontent.com/WongKinYiu/PartialResidualNetworks/master/cfg/yolov3-tiny-prn.cfg instead of yolov3-tiny.cfg with the same accuracy.

Also you can set width=320 height=320 in cfg-file and run training and detection, it will be faster, but lower accuracy.

Also you can try to use some of these models: https://github.com/AlexeyAB/darknet/issues/3114#issuecomment-494148968

ntemperli commented 4 years ago

Thank you very much for your response. I'll definitely give it a try.

Do i also have to use the "darknet partial" command from your weights file like in tiny-yolov3 before training? (./darknet partial cfg/yolov3-tiny.cfg yolov3-tiny.weights yolov3-tiny.conv.15 15)

Thanks in advance