Segmentation support & other enchancements

[AyushExel]

• Segmentation support
• More robust validator

This PR has not been cleaned up.. Dataset path is hardcoded.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improvements to model CI, dataset handling, and segmentation training.

📊 Key Changes

• CI configuration updated to include latest PyTorch version and removed specified torch versions.
• Dataset processing in train.py now handles COCO and ImageNet more intelligently.
• Added new segmentation training algorithms with smooth loss calculations.
• NotImplementedErrors replaced with concrete implementations in some classes.
• YAML files for YOLOv5 models added, defining network architectures.
• Default training configurations now reflect changes made to dataset handling and model configuration locations.

🎯 Purpose & Impact

• CI changes ensure tests run with the most up-to-date PyTorch framework, improving compatibility.
• Enhanced dataset handling provides a scaffold for more robust training on standard datasets like COCO and ImageNet.
• Segmentation-focused code introduces an expanded usage scope of the Ultralytics framework, catering to more tasks than object detection.
• YAML-based model definitions modularize architecture configurations, facilitating easier experimentation and adjustments.
• Changes impact users by broadening the ease of training custom models and utilzing the framework for tasks beyond traditional object detection.

[glenn-jocher]

@AyushExel resolved conflicts and merged master. You should git pull and make sure the updates didn't break your seg training.

[glenn-jocher]

@AyushExel pre-commit's got a bunch of valid errors here we need to work through following the main merge.
https://results.pre-commit.ci/run/github/535360445/1667852320.wmgI6kxoRCuwwgNUUlI0dg

[AyushExel]

@glenn-jocher now segmentation auto downloads coco128-seg so it can be tested like:

 python yolo/v8/segment/train.py img_size=64

[glenn-jocher]

@AyushExel awesome! I'll test locally and then update CI with a seg training

[glenn-jocher]

@AyushExel I'm getting FileNotFoundError: [Errno 2] No such file or directory: 'yolov5n-seg.yaml', so I added it to the repo under a new v8/models directory.

After this it still doesn't find it, so it looks like check_yaml() and check_file() have not been ported over from v5. The purpose of these functions is to search ROOT for instances of the filename if it is not found in the CWD, which allows simplifying CLI commands. I think we need to add this in and apply check_yaml() to all --cfg and --data arguments.

[AyushExel]

@glenn-jocher should work now. Let's port check functions from v5 in the next step

[glenn-jocher]

@AyushExel PR merged!