Yolov3: How different is this library when using YOLOv4 weights with YOLOv4 performance?
Wanted to know how different the performance of actual YOLOv4 implementation including some other tricks in the recently released paper and this library when just using the YOLOv4 weights and cfg file
suarezjessie
All 5 comments
@suarezjessie well, it seems that when trained from scratch with this repo, yolov4 produces a higher mAP (without the tricks) here than on darknet with the tricks.
glenn-jocher
on 23 May 2020
@glenn-jocher Oh cool. And how different is the architecture? Is there a really big difference with the one in the original paper?
suarezjessie
on 23 May 2020
@suarezjessie there are no architecture differences. You can use train here with the same cfg, the only difference is that many of the special tricks in the yolo layer attributes are not implemented in this repo:
python train.py --cfg yolov4-relu.cfg --weights ''
glenn-jocher
on 23 May 2020
Got it. Thanks!
suarezjessie
on 24 May 2020
@suarezjessie there are no architecture differences. You can use train here with the same cfg, the only difference is that many of the special tricks in the yolo layer attributes are not implemented in this repo:
python train.py --cfg yolov4-relu.cfg --weights ''
Hi, Can you share the pre-trained yolov4-relu weights on the coco dataset? Training from scratch will consume a lot of time. Thanks!
buaazsj
on 1 Jul 2020
Related issues
Arunavameister
路
3Comments
leeyunhome
路
3Comments
MichaelCong
路
4Comments
yoga-0125
路
4Comments
khalidw
路
5Comments
Most helpful comment
@suarezjessie there are no architecture differences. You can use train here with the same cfg, the only difference is that many of the special tricks in the yolo layer attributes are not implemented in this repo:
python train.py --cfg yolov4-relu.cfg --weights ''