Clarification on YOLO ā†’ TFLite int8 quantization: do I need representative images?

šŸ¤ Community Help
1

Hi everyone ā€” Iā€™m working on exporting a custom YOLO12 model (trained via Roboflow) to TFLite with full integer (int-8) quantization.
My workflow: Roboflow ā†’ download best.pt ā†’ using Ultralytics model.export(format='tflite', int8=True, data='data.yaml').
The question: Do I have to supply a folder of images (representative dataset) manually for calibration, or will the exporter automatically pull sample images from the val path in data.yaml?
Iā€™m seeing uncertain behaviour and need to confirm before deploying.
Any tips/traps I should be aware of (especially for edge-device with 4GB RAM)?
Thanks in advance for any guidance.

  • Project Type: Object Dectction
  • Operating System & Browser: IOS/ Android
  • Project Universe Link or Workspace/Project ID: rah-wound-burn-pxzx3
2

Hi @amirzhou!
I would love to learn more about your use case and the problem you are using Roboflow to solve. We would love to help you get your model into production easier, and would love any feedback or ideas you may have.

Unfortunately, I am unable to support model conversions outside of the Roboflow platform, apologies for the confusion.

Looking forward to learning more about your use case, happy building!