InceptionResNetV2 quantisation accuracy issues

General
1

I have fine-tuned various models including InceptionResNetV2 in TF2.X.

i have no issues with most of the models, but with this one i see the following behaviour:

  1. converting from tflite to a simple haar and running emulator with SDK_NATIVE gives same accuracy as original model.

  2. converting the model to har with optimization_level=0 and running emulator with SDK_QUANTIZED gives same accuracy as original model.

  3. converting the model to har with optimization_level=2,3,4 i get about 38 dB in SNR in all cases. but when i run emulator with SDK_QUANTIZED, i get same accuracy at about 2% vs 85%+.

not sure what goes wrong, as inceptionv3/resnet50v1 work ok.