will the android platform support tflite fp16 & fp32 , and extend the image size to 640?

[dingniaoji]

as title described.

[UltralyticsAssistant]

👋 Hello @dingniaoji, thank you for submitting a ultralytics/yolo-flutter-app 🚀 Issue. This is an automated response to help guide you, and an Ultralytics engineer will assist soon.

To address your concern efficiently, please provide the following information:

1. For questions:
   • A clear and concise description of the issue or feature mentioned
   • Any specific details regarding TensorFlow Lite support for FP16 & FP32 on Android
   • Whether you’re experiencing any specific limitations with image size
   • Include any research you've already done on the topic
   • Specify which parts of the documentation, if any, you've already consulted

For any additional context or updates, please comment on this issue. Make sure you've searched existing issues to avoid duplicates.

Thank you for helping improve our project! 🌟

[ice6]

actually, it does supports!

you just need to make sure that the imgsz in your metadata.yaml is same with your model.

although fp16 and fp32 is quite slow in some mobile device with tensorflow, but yolo-flutter-app does support it.

so the instruction in the README.md file is really misleading. it does not only support imgsz=320 with int8.

this issue can close now. after debugging the android code, I found that is does support fp16 and fp32 and more image size, not only int8 with imgsz=320 according to the README.md file.

[ice6]

more information for others : int8 is ok for s and n model arch. m l x not works. maybe m l x is not suitable for mobile device and some complex operation is not support for the conversion.
after yolo11 release, I found n and s mAP50-90 is better than yolov8, now I do not need to use a x model for mobile.

[pderrenger]

Thank you for sharing your insights! It's great to hear about your experience with the models. If you have any further questions or need assistance, feel free to reach out.

[sakura-xiamu]

how to get metadata.yaml ?

[pderrenger]

@sakura-xiamu the metadata.yaml file is typically generated during the model export process. If you're exporting a model using Ultralytics tools, it should be automatically created. If it's missing, ensure you're using the latest version of the tools and try re-exporting your model. If you continue to experience issues, please let us know.