Added support for fp16 (half) to export.py

[hodovo]

This adds support for the --half argument when exporting a model.

Considerations
I have only tested this with torchscript. It might be a good idea to only support this for torchscript if you know of any problems with ONNX/CoreML and fp16. My focus is on torchscript. I can easily modify this pull request to load the model separately as half for only torchscript and also create a separate img for the trace.

Let me know your thoughts :)

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Added FP16 half-precision export option to YOLOv5 model exporter. 🚀

📊 Key Changes

• New --half argument to the model export script enabling FP16 half-precision exports. 🎛️
• Added an assertion to ensure FP16 export is not attempted on CPUs, as it's only compatible with GPUs. 🔒
• Conditional logic to convert the model and input image to FP16 when the --half flag is used. 🔧

🎯 Purpose & Impact

• The added --half flag allows users to export their models with reduced precision, which can lead to performance gains on compatible hardware (mostly GPUs) by reducing the computational workload. 📈
• This update emphasizes efficiency, as using FP16 can decrease model size and increase inference speed, which is beneficial for deployment on edge devices or in environments where resources are limited. 🏎️
• The assertion prevents users from inadvertently trying to export an FP16 model on unsupported devices like CPUs, avoiding potential errors and confusion. 🚫🖥️

[glenn-jocher]

@hodovo thanks for the PR! I've minimized the changes a bit but otherwise looks good.

This work conflicts with an earlier PR in progress #2982, so we will merge the earlier PR first hopefully early this week, and then once #2982 is merged we will merge master in this one, resolving conflicts manually.

[hodovo]

Sounds good! Your changes look much nicer :)

Let me know if you need anything. Thank you.

[glenn-jocher]

Tested, works well with ONNX (below) and CoreML (not shown).

[glenn-jocher]

@hodovo PR is merged. Thank you for your contributions!

[hodovo]

My pleasure. Thank you for the awesome platform you have built :)

[xonobo]

Does torchscript work properly in half precision? In this issue it is stated that half precison of convolution is not implemented. A simple testing code for the exported models may be nice for this case.

[hodovo]

@xonobo Yes it does. That issue is specifically for CPU and not GPU. The implementation in this pull request checks to ensure that the export device is a CUDA device. However, that doesn't enforce how the user utilizes the model afterwards.