Update dataset names from array to dictionary

[glenn-jocher]

Created with

    from utils.general import yaml_load
    for f in list(Path('data').glob('*.yaml')):
        d = yaml_load(f)
        d['names'] = {i: v for i, v in enumerate(d['names'])}
        yaml_save(f.with_stem(f.stem + '_new'), d)

@AyushExel @kalenmike implements names dict in yamls as discussed.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Refinement of class representation and minor code adjustments in Ultralytics YOLOv5.

📊 Key Changes

• 🔢 YAML files updated from list to dictionary format for class names.
• 📝 Code modified to handle the YAML changes and ensure compatibility with dictionary format.
• 🧹 Small tweak in predict.py to convert tensor to list when logging.

🎯 Purpose & Impact

• 🧑‍💻 Helps developers with more flexible and accessible class indexing.
• 📈 Ensures that downstream code correctly interprets the new class name format.
• 🛠 Facilitates potential future extensions to class handling.

These changes create a more intuitive way of handling class names and data, positively impacting maintainability and scalability of the model configurations.

[AyushExel]

Just to confirm, this will be backwards compatible right?

[glenn-jocher]

@AyushExel yeah that's the idea. All data yamls pass through check_dataset(), so I'll add some logic to check for isinstance(names, list) and convert to dict.

[glenn-jocher]

@AyushExel @kalenmike PR is merged, backwards compatible with older data yamls and should output the same results for HUB dataset stats.

Let me know if you run into any problems.

[austin1howard]

FYI @glenn-jocher this doesn't break loading/training, but because of the change at https://github.com/ultralytics/yolov5/pull/9000/files#diff-cfb1ff087a99a34369673c9f34bdcd22f2d429ab3599a89f386c5de1fd9a2566R452 (which sets model.names to the dictionary now) any code (like ours) that was using model.names breaks now. In our case, because we were iterating over it, our iteration went from class1, class2, class3 to 0,1,2.

[glenn-jocher]

@austin1howard thanks for bringing this to our attention. I've made a note to check for this scenario and will make sure to address it in the next update. We aim to maintain compatibility and apologize for any inconvenience caused.