Tensorflow Lite operator refused by NNAPI Delegate

[Creative]

Hello!

I'm trying to run real-time face detection on I.MX 8M Plus. I have trained my .tflite model on pre-trained model ssd_mobilenet_v2_320x320_coco, converted and quantized it with TensorflowLite converter with interference_input_type=uint8, interference_output_type=float32

When I run the model I have this warnings:

WARNING: Operator PACK (v2) refused by NNAPI delegate: Unsupported operation type.
WARNING: Operator PACK (v2) refused by NNAPI delegate: Unsupported operation type.
WARNING: Operator PACK (v2) refused by NNAPI delegate: Unsupported operation type.
WARNING: Operator PACK (v2) refused by NNAPI delegate: Unsupported operation type.
WARNING: Operator CUSTOM (v1) refused by NNAPI delegate: Unsupported operation type.

Because of this my model does not fully run on NPU and more slower than I wanted.

I have tried other conversion types - none of them helped.

How could I get rid of this and have my model run faster fully on NPU?

I attach my model for you to examine.

[Bio_TICFSL]

Hello Creative,

1. that's from an underlying NNAPI driver. When you delegate your model to NNAPI, the NNAPI runtime asks all the drivers to decide which driver / hardware is the right one to dispatch the an op to run on it.
2. you can try

adb shell setprop debug.nn.vlog 1

and then adb logcat | grep -i best to see how your model is handled by NNAPI. Check https://developer.android.com/ndk/guides/neuralnetworks for more NNAPI related information

3. use either one of them, not both.
4. you may want to read related guides first, https://www.tensorflow.org/lite/performance/model_optimization

Regards