Make it easier to do transfer learning by adding --freeze as an option

[timstokman]

I've been doing some experiments with transfer learning. This small tweak makes it little easier to do transfer learning. It supports the most common use-case for freezing layers: freezing layers up to a point. It's still possible to manually edit the freeze array if more control is required over what exactly to freeze.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improving layer freezing options in YOLOv5 training.

📊 Key Changes

• --freeze argument added to parse_opt(), allowing users to specify the last layer number to freeze during training.
• Model parameter freezing logic updated in train() to handle the new --freeze option.
• Logging enhancement to display which model layers are being frozen or not during initialization.

🎯 Purpose & Impact

• Purpose: This change allows for more precise control over which layers of the model are frozen, which can be beneficial for transfer learning or fine-tuning models where certain layers' weights are desired to remain unchanged.
• Impact: Users now have the flexibility to easily freeze a specific range of layers by providing a layer number, potentially improving training times and preserving pre-trained feature detection capabilities in earlier layers. This can be especially useful when working with limited computational resources or wanting to maintain certain aspects of a pre-trained model.

[timstokman]

Had some merge conflicts, so I fixed those

[IneovaAI]

Hi please have a look at my implementation and tell me what you think. IneovaAI:patch-1 #4238

[glenn-jocher]

/rebase

[glenn-jocher]

@timstokman @IneovaAI closing this PR and merging --freeze PR #4238. Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐