Increase plot_labels() speed

[glenn-jocher]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved label plotting in training with updated visualization libraries.

📊 Key Changes

• Removed asynchronous thread for label plotting during training in train.py.
• Refactored label plotting in utils/plots.py to utilize Seaborn and Pandas for generating correlograms and histograms.
• Simplified the correlogram plot code by removing exception handling around the import and plotting code.
• Updated scatter plots to use Seaborn's histplot for better visualization and consistency.

🎯 Purpose & Impact

• Purpose: Enhance the data visualization during training for better insights into label distributions and relationships.
• Impact:
  • Developers will benefit from more immediate and potentially more informative visualizations without the complexity and potential errors from asynchronous plotting.
  • Overall improvements to plotting methods can lead to easier identification of issues with data sets, potentially improving model performance.
  • Users should be aware that the update will change how plots are generated and displayed in TensorBoard and other logging tools.

[glenn-jocher]

plot_labels() implementation suffered from speed issues due to some customizations on 2d histograms and daemon thread implementation that seemed to clash with the dataloader threading. This update should reduce label plotting time on COCO from 30 seconds to about 3 seconds.