Tensorflow Lite 2.2.0 NNAPI memory reak in L5.4.47_2.2.0

[hsaito]

Operating environment is i.mx8qm, using L5.4.47_2.2.0.

We modified label_image.py to check the operation of sensorflow Lite 2.2.0, which uses NNAPI.

The fix is to call Invoke() process continuously.(I have pasted the modified program label_image.py.)

My goal is to use camera footage to perform inference processing, but will using NNAPI cause memory leaks?

Is there any way to improve it?

The starting memory usage is shown in before.png and the memory usage after about an hour is shown in after.png.

[Rita_Wang]

Are you using NXP i.MX8QM reference board?

[hsaito]

This result is our original board.

However, the memory usage continues to increase even when I run it on the evaluation board.