FP16 quantized TFLite model performs strangely when using XNNPACK or FP16 acceleration

[Mayner0220]

1. System information

• OS Platform and Distribution: Linux Ubuntu 22.04.5 LTS
• TensorFlow installation: pip package
• TensorFlow library

  tensorflow-gpu==2.9.3
  tensorflow-addons==0.19.0
  tensorflow-probability==0.17.0
  

• Mobile device

  ❗ It has been confirmed that it is not an issue arising from the dependence on mobile devices.

  • Samsung Galaxy Tab S7 (Android)
  • Apple iPad M2 Air 13in (iPadOS)

2. Code

• It is difficult to provide the code for personal reasons. 😢

3. Describe the current behavior

• For the purpose of mobile deployment, I converted the Human Pose Estimation(HPE) model into TFLite after training. FP16 quantization was applied for the lightweight of the model at the time of conversion to TFLite. When testing the converted model after deployment to mobile devices, it was found that the inference results of the model were strange, which could be solved by deactivating XNNPACK and FP16 operations. (Examples of this can be found below.)
• The below phenomenon occurred when the model's existing backbone was changed to a new backbone, and the newly applied backbone is ConvFormer-S18.

Activating XNNPACK or FP16	Deactivating XNNPACK or FP16

4. Describe the expected behavior

I hope that the inference result of the model will be normal even when using the XNNPACK and FP16 operations.

Q&A to resolve this issue is welcome. Please feel free to leave your comments or questions.

[pkgoogle]

Yeah that doesn't look right. We'll take a look.