Favor EMA over model in train checkpoints

[glenn-jocher]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improvements in model saving and configuration handling in Ultralytics training engine.

📊 Key Changes

• Removed de_parallel utility from imports, streamlining the library's dependencies.
• Changed how models are saved: now explicitly sets "model" to None in the checkpoint, emphasizing reliance on EMA (Exponential Moving Average) for model restoration.
• Adjusted model setup to derive configuration directly from weights rather than through checkpoint dictionary, simplifying the configuration process.

🎯 Purpose & Impact

• Efficiency Boost: Removing unnecessary utilities and streamlining saving processes makes the code cleaner and potentially faster. 🚀
• Simplification: By directly linking model configuration to weights and emphasizing EMA for checkpoints, this update simplifies the model restoration process for users, making it more intuitive. 🌐
• User Experience: These changes enhance user experience by making model saving and loading both more straightforward and more reliable, likely reducing common issues or confusion during these critical steps. 🛠️

This update is part of ongoing efforts to refine and optimize the Ultralytics framework, ensuring it remains at the forefront of efficiency, usability, and reliability in the AI and machine learning community. 🌟

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 76.98%. Comparing base (479afce) to head (e91c97b).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #9433      +/-   ##
==========================================
+ Coverage   76.94%   76.98%   +0.03%     
==========================================
  Files         117      117              
  Lines       14854    14854              
==========================================
+ Hits        11430    11435       +5     
+ Misses       3424     3419       -5     

Flag	Coverage Δ
Benchmarks	36.78% <0.00%> (+0.05%)	⬆️
GPU	38.84% <100.00%> (-0.02%)	⬇️
Tests	72.05% <100.00%> (-0.03%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.