How to use NPU on Android

[re_roy]

Hello Khadas team

I have Edge2 that has powerful NPU

I hope to use that powerful source on Edge2 with Android

Could you support “how to convert model” & “how to use converted model on Android”?

I already tested rknn tool kit 2 demo source and model convert script

so i have a converted model (some model can’t convert)

when i see converted model content by netron

I think it is not successful converting (because data type is “?” )

and android demo app replace yolov5 model to ours is not work

I hope to use this model

google/mediapipe/blob/master/docs/solutions/models.md#face_detection:~:text=com/mediapipe/solutions-,Face Detection,-Short-range model

[Electr1]

@re_roy

I already tested rknn tool kit 2 demo source and model convert script

so i have a converted model (some model can’t convert)

What kind of errors are you getting from conversion ?

Its suggested to raise issue in https://redmine.rock-chips.com/

@re_roy

when i see converted model content by netron

I think it is not successful converting (because data type is “?” )

RKNN conversion is still experimental in Netron, for knowing for the details of the model, make sure you do the conversion with enableing the verbose in the RKNN object such as RKNN(verbose=True)

This will give detailed information of the graph/layers summary with data type information.

@re_roy

I hope to use this model

mediapipe/docs/solutions/models.md at master · google/mediapipe · GitHub

Do you mean face_detection_short_range.tflite ? Need to check the version of tensorflow lite ops set. RKNN conversion tool is still based on Tensorflow 2.8.0 Respectively.

Regards.