Darknet: YOLOV3-voc.cfg

[mingmmq]

Seem the voc cfg is not provided, but mentioned in the website.

[pjreddie]

Whoops! pushed it b2ae5eea8e26f3826153123ec402013d5154cde0

[sivagnanamn]

Will there be a tiny version of V3 network?

YOLO V3 VOC cfg from https://github.com/pjreddie/darknet/commit/b2ae5eea8e26f3826153123ec402013d5154cde0 - doesn't use yolo layer for detection? whereas YOLO V3 COCO cfg uses yolo layer.

[sivagnanamn]

YOLO V3 VOC from https://github.com/pjreddie/darknet/commit/060faa268de7bf45126fbc1ee775f6af091fc8ed seems to be using yolo layer for detection.

[czero69]

quoting @TheMikeyR user:
"There are no tiny yolov3, yolov3 is just a few improvements to yolov2.
You could try to do the suggestions by Alexey but it seems no one have managed to test it yet AlexeyAB#552 (comment) "

I trained shrinked yolo3 model as AlexeyAB suggest, and it works very fine for me

[Lunaticf]

@czero69 can you share the shrinked yolo3 model and cfg,cause I ran into some problem when I modify it and train it,and I have limited device as a student.sorry,my english is poor.Thank you very much

[czero69]

Yes, of course. Note, that all changes I did are the same as AlexeyAB#552 (comment). Note also, I have classes=6. Please follow AlexeyAB github to set filters accordingly to your class number. I trained it for ripped Coco dataset (person, car, ..) + my own sample. With this dataset, I encountered 'nans' on the console, and the average loss was high ~3.5-4.0. But the results are astonishingly good. I leave defaults anchor boxes defined for Coco set.
yolov3-pedestrians_coco_subset.cfg.zip

[Lunaticf]

@czero69
thanks!