EfficientDet-Lite0 slower with NNAPI

TensorFlow
, ,
1

Hi,

I’ve been trying to use EfficientDet-Lite0 for object detection with TF lite. I tried on Android(Samsung A32) with NNAPI and CPU. Inference time for CPU is <100ms whereas with NNAPI it’s >1000ms. I’m hoping someone can confirm that this is expected and if so why?

Thanks

3

Have you checked?

https://tensorflow-prod.ospodiscourse.com/t/neural-network-fallback-to-cpu-using-nnapi-on-android/7703/3?u=bhack

5

Could be that the CPU is using multiple threads, whereas the NNAPI delegate does nothing but cause the phone stress because this phone does not have any compatibility with the NNAPI delegate. This is an educated guess :slight_smile: