conductor

Yolo v3 본문

카테고리 없음

Yolo v3

Hemos 2019. 10. 28. 12:00

* Joseph Redmon - pjreddie.com / darknet
AlexeyAB - github.com/AlexeyAB / darknet
(https://github.com/AlexeyAB/darknet)

 

darknet data

 

coco.names : 80 (car...)

openimages.names : 601 (sunglasses / helmet...)

voc.names : 20 (bus/car/horse/monitor/sofa...)

9k.names : 9418 (wave/hair/weather...) / 9k.labels / 9k.tree

imagenet.shortnames.list : 21842 (tree/carbon/soil/vitamin E) / vi imagenet.labels.list

 

filters = ( classes + 5 ) * 3
classes

599, 610  689, 696  776,783

Yolo_mark/x64/Release/data (img, obj.data, obj.names, train.txt) to darknet/data

 

./darknet detector train data/obj.data cfg/yolov3.cfg darknet53.conv.74 -gpus 0,1

-gpu 0
-gpus 0,1

 

 

detector.c
138
& make

 

gpus0,1

대략

 

 

 

training ---

weight 경로
~.data 파일 -> backup 경로 ../

training 관련 및  학습효과 업 관련
: https://murra.tistory.com/18

cfg random BL
cfg stride, layers 717,720
flip 17
stopbackward 548
anchors size
--
cfg height,width
subdivision 16,32,64

weights 성능파악 (win)
: darknet.exe detector map data/obj.data yolo-obj.cfg backup\\yolo-obj\_(~).weights
-map flag (train)

2000 cycles / class
avg
avg loss