Usage of imx 8m plus NPU

[Vijay_hegde]

I was using tflite model to run object detections and as per the documents I was run it using the libvx_delegate.so. But the warm-up time is very much high and for the videos the things work for each frame.
So is there any other ways to use the NPU on this device.

[Chavira]

Hi @Vijay_hegde!

Thank you for contacting NXP Support!

The only way to use our NPU is using libvx_delegate.so.

You can refer to our i.MX Machine Learning User's Guide to lern more about it.

https://www.nxp.com/docs/en/user-guide/IMX-MACHINE-LEARNING-UG.pdf

Best Regards!

Chavira

[Vijay_hegde]

Can you suggest me how to use that, I have attached my code which uses tflite model and when try to run on the NPU the fps of the frames after object detection in 1.5 which too slow after using NPU.

[Chavira]

Hi @Vijay_hegde!

how you are loading the external delegate in your .py?

You can check one of our examples to apply the proper chanegs in your program.

https://github.com/nxp-imx/tensorflow-imx/blob/lf-6.6.23_2.0.0/tensorflow/lite/examples/python/label...

Best Regards!

Chavira

[Vijay_hegde]

I have referred and loading external delegate using experimental_deligate = tflite.load_delegate().
And I am able to see it running on delegate. But for video its very slow.

[Chavira]

Hi @Vijay_hegde!

To load the external delegate we use the next code:

  # load external delegate
  if args.ext_delegate is not None:
    print('Loading external delegate from {} with args: {}'.format(
        args.ext_delegate, ext_delegate_options))
    ext_delegate = [
        tflite.load_delegate(args.ext_delegate, ext_delegate_options)
    ]

  interpreter = tflite.Interpreter(
      model_path=args.model_file,
      experimental_delegates=ext_delegate,
      num_threads=args.num_threads)
  interpreter.allocate_tensors()

You can check the list of supported operators in the link below:

https://github.com/nxp-imx/tflite-vx-delegate-imx/blob/lf-6.6.23_2.0.0/op_status.md

Best Regards!

Chavira