Hi everyone ! I’m Glenn Jocher, YOLOv5 author and Ultralytics founder.
I’m super excited about our new forum here at https://community.ultralytics.com! This is a super cool new place for everyone to share YOLO stories, ask for help and provide the Ultralytics team feedback on what you like and what we could do better.
Our mission at Ultralytics is to make AI easy for everyone. But we need your help! Please let us know how we’re doing and how we can help make your AI life easier.
first of all Amazing repository, yolov5 is easily my one of most revisited repository. all the features provided by yolo are amazing. Features like evolve, sweep, tfjs export and those amazing highly generalizable models!
I have used yolov5 extensively, modified the code multiple times to meet my niche requirements. I had high mAP like 99 on some of the projects with yolo small and it to me mention they train very fast.
Now I am still younger and exploring things but few things could be better
data loading is unintuitive:- Dataset.yaml we define base dir for train, cal and yolov5 expects a realative folder structure. This was a bit unintuitive for me, I would have simply preferred giving base dir for both train images and train labels, same for rest of the sets. I also would love to see COCO json support, since COCO json are far more easier and Faster to manipulate than bunch of text files.
I felt that yolov5 is little bit restrictive in terms of modularity, on personal level I would prefer modularity over convenience.
would mentain docs bit better - don’t get me wrong here, documentation is still one of the best overall but some threads are outdated I faced the similar thing recently while I was trying out sweep. The details mentioned are high level and outdated. Whe I read through the entire PR then I understood the feature
torcheserve .mar file export for seamless deployments (I would love to contribute and raise PR for this export feature)
! I’m Glenn Jocher, YOLOv5 author and Ultralytics founder.