Closed stark-t closed 2 years ago
Wow! @stark-t , this field is moving faster than I can say "object detection"...
Seems that there are even two popular github repositories for YOLOv7:
1) https://github.com/jinfagang/yolov7
2) https://github.com/WongKinYiu/yolov7 This is the one corresponding to the paper link that you put above. Doesn't seem that well documented for now, but hopefully is not much different than the usage of yolov5
I think we should make use of YOLOv7 then. I will start implementing a run on the cluster with yolov7 and see how it goes with setting the environment & dependencies (link 2, the one with the published paper)
@valentinitnelav I'm almost ready to push to yolov7. Some things could be problematic when using v4, v5, v7, we should delete the data cache whenever we test a new model.
yolov7 seems to be quite buggy. eg I needed to manually download the weights etc.
yolov7 seems to be quite buggy. eg I needed to manually download the weights etc.
FYI: On the cluster, I also manually download the weights for yolov5 in a designated folder and I pass the path as an argument to train.py I found that more easy to debug in case a node has internet access restricted at run time.
I will try today to set environments for the other versions of yolo and I will open a discussion as an issue if I encounter difficulties.
YOLOv7 environment set & corresponding cluster job scripts created with commit 530cef0e72321efcb57c68e0042ec9a98255675d
https://arxiv.org/abs/2207.02696