In yolov2 , I am able to use batch=64 with subdivisions=8 i.e 8 images per mini-batch now it is giving error and working fine when using subdivisions=16 with batch =64
No. Almost same.
But, It depends on base model(backbone).
I think you used yolo_v2 with darknet19.
yolo_v3 is based on darknet53(like resnet).
darknet53 is more bigger than darknet19. That's why it needs more gpu memory.
@springkim
Can I use yolov3 with darknet19 ?
Most helpful comment
@springkim
Can I use yolov3 with darknet19 ?