DetectMultiBackend improvements

[glenn-jocher]

Signed-off-by: Glenn Jocher [email protected]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhancements to model loading and inference processes in YOLOv5.

📊 Key Changes

• 🔧 Optimized output name retrieval for ONNX Runtime sessions.
• ✂️ Simplified metadata loading from YAML files during model initialization.
• 🚀 Improved forward method by caching output names for ONNX and using torch.from_numpy for efficient tensor creation.
• 🧹 Cleaned up conditional checks and default values in _load_metadata method.

🎯 Purpose & Impact

• 🚅 Provides quicker setup for ONNX models by caching output names, leading to faster inference times.
• 📁 Ensures more robust metadata handling, preventing errors and enabling smoother user experiences.
• 🧠 Facilitates better memory management by avoiding unnecessary tensor conversions.
• 💼 Reduces code complexity and potential bugs by refining the metadata loading process, beneficial for maintainers and contributors.