This PR adds torch 1.7.0 compatibility updates, and relocates 1.6.0 legacy compatibility updates to attempt_load() in experimental.py. Will update official models as well for 1.7.0 compatibility and attach to a v3.1 release soon.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhancements for model loading and inference compatibility in YOLOv5.

📊 Key Changes

• Added autoshaping, evaluation, and fusing to the YOLOv5 model example in hubconf.py.
• Enriched attempt_load function in experimental.py with compatibility features for different PyTorch versions.
• Removed redundant compatibility code in yolo.py after fusing convolution and batch normalization layers.
• Expanded activation function updates in torch_utils.py for initialization to include nn.Hardswish.

🎯 Purpose & Impact

• Ensuring YOLOv5 models work smoother with various input types and different PyTorch versions.
• 🚀 Potential impact includes easier model deployment, especially for users working with different versions of PyTorch or different image input formats, fostering broader adoption and user convenience.