Adjust legend labels for classes without instances

[NauchtanRobotics]

See issue #5158

Before

PR Curve legend has labels for "D00" and "D10" but should have labels for "D40" and "EB".
D00 and D10 have no bounding box instances, and no predictions.
'stats' omits classes that have len = 0; therefore 'names' should also omit empty classes.

:

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved class naming handling in precision-recall and F1 score calculations and plots.

📊 Key Changes

• Added filtering to only include class names that correspond to classes with data when generating metrics.
• Converted the filtered class names list to a dictionary with an enumerated index for consistent referencing.

🎯 Purpose & Impact

• Ensures that metric plots only display relevant classes, reducing clutter and potential confusion.
• Impacts users by providing clearer and more accurate visualization of model performance metrics, aiding in better model evaluation and interpretation. 📈🔍

[NauchtanRobotics]

After

The legend of PR curve has labels "D40" and "EB" which is correct (same as labels.jp

g)

[NauchtanRobotics]

@glenn-jocher For future reference, how do you feel about incrementally introducing typing? Would make understanding the code more rapid and development easier. E.g.

from typing import List, Dict


def plot_pr_curve(
    px: np.ndarray, 
    py: np.ndarray, 
    ap: List[float], 
    save_dir: str='pr_curve.png', 
    names: Dict={}
):

[glenn-jocher]

@NauchtanRobotics thanks for the PR! Would this simplified version also work?

    names = [v for k, v in names.items() if k in unique_classes]  # list of unique names
    names = {i: v for i, v in enumerate(names)}  # to dict

Yes, typing is one of the options we are considering for improving the codebase, but for consistency we'd probably want to apply formatting updates across the whole repo, so let's leave it off the plate for now. Thanks!

[NauchtanRobotics]

Yes, that works a treat @glenn-jocher. See all 4 affected charts.

Feel free to squash down to one commit

.

.

[glenn-jocher]

@NauchtanRobotics PR is (squashed and) merged. Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐