Enabling XNNPACK changes the output.

[kafan1986]

System information

• OS Platform and Distribution: Android
• Mobile device (e.g. iPhone 8, Pixel 2, Samsung Galaxy) if the issue happens on mobile device: All android
• TensorFlow installed from (source or binary): Binary
• TensorFlow version (use command below): v2.3
• Python version: N/A
• Bazel version (if compiling from source): N/A
• GCC/Compiler version (if compiling from source): N/A
• CUDA/cuDNN version: N/A
• GPU model and memory: N/A

Describe the current behavior
Running a computer vision model for emotion detection. The output is as expected when XNN is disabled but if I enable tfliteOptions.setUseXNNPACK(true) then the output is incorrect and totally different.

Describe the expected behavior
Output should be same for both XNN enabled and non-XNN option.

I have attached inference class files and model files below.

ClassFiles.zip
emotion_fp16t.tflite.zip

[Cospel]

Having the same issue with XNNPACK. If running delegate with XNNPACK the result is totally different from the disabled one. (TF2 version 2.3.1)

[abattery]

@Maratyszcza could you take a look at this issue?

[Maratyszcza]

@kafan1986 Does it reproduce if you build TFLite from the master branch? There was a bug with asymetric padding a while ago that was fixed in google/XNNPACK@facecc5

[kafan1986]

@Maratyszcza I have tested with 0.0.0-nightly build of tflite and the output is same for both version. But the performance seems to be remain same for both with XNN module true and false, whereas earlier the XNN true would significantly run the inference faster although the output was wrong.