Improved OpenVINO quantization code

[adrianboguszewski]

Small fixes + better code layout

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhancement and simplification of OpenVINO export in YOLOv5.

📊 Key Changes

• Updated the filename for OpenVINO export to include int8_ prefix when INT8 quantization is used.
• Upgraded the required version of NNCF for post-training quantization to >= 2.5.0.
• Streamlined the quantization code by removing redundant preprocessing function and directly incorporating preprocessing within the transformation function.
• Fixed assertion to ensure input images are in UINT8 format for quantization preprocessing.

🎯 Purpose & Impact

• 🚀 Purpose: To improve the OpenVINO export functionality of YOLOv5 by making the output filenames more descriptive and ensuring compatibility with the latest post-training quantization tools.
• 📈 Impact: Users can now have clearer naming for models with INT8 quantization and benefit from newer quantization features with the updated NNCF requirement. The streamlined preprocessing should make the quantization process more efficient and less error-prone.
• 🔍 Non-Expert Use: For users who are not familiar with machine learning quantization, the update means better performance and reliability when using YOLOv5 with OpenVINO, especially on systems that support INT8 operations.

[glenn-jocher]

@adrianboguszewski hi there!

Thanks for your interest in contributing to the YOLOv5 project! 🚀

For any code improvements, including enhancements to OpenVINO quantization, we welcome pull requests. Please ensure your changes are well-documented and tested for compatibility. Once you've submitted a PR, our team will review it and provide feedback or merge it if everything checks out.

Looking forward to seeing your contributions!

[glenn-jocher]

@adrianboguszewski PR merged! I checked the benchmarking CI and it looks like the new changes are working correctly :)