About OpenVX Graph Cache

[at_akihito]

Hi.

I'm using a board with i.MX8MPlus.

I have created a model and a Python script to segment humans from captured images using machine learning.

I tried the OpenVX graph cache to reduce the warm-up time before running the inference, but it stopped working on the second and subsequent runs.

I'm attaching the source code and model that reproduces the situation.

logs from the run are as follows:

# tar xf cache_error.tar.gz
# cd image_segmentation
# export VIV_VX_ENABLE_CACHE_GRAPH_BINARY="1"                                                      
# export VIV_VX_CACHE_BINARY_GRAPH_DIR=`pwd`
# python3 image_segmentation.py
INFO: Created TensorFlow Lite delegate for NNAPI.
WARNING: Operator RESIZE_BILINEAR (v2) refused by NNAPI delegate: Operator refused due performance reasons.
WARNING: Operator RESIZE_BILINEAR (v2) refused by NNAPI delegate: Operator refused due performance reasons.
WARNING: Operator RESIZE_BILINEAR (v2) refused by NNAPI delegate: Operator refused due performance reasons.
WARNING: Operator RESIZE_BILINEAR (v2) refused by NNAPI delegate: Operator refused due performance reasons.
Applied NNAPI delegate.

# ls *.nb
7defa710c26bc71d88b66e629b41d231.nb  883db2f3bfce82bede17e5bed2e4a6b5.nb  ca661c9048851a6b2755fa0982d07102.nb
# python3 image_segmentation.py
INFO: Created TensorFlow Lite delegate for NNAPI.
WARNING: Operator RESIZE_BILINEAR (v2) refused by NNAPI delegate: Operator refused due performance reasons.
WARNING: Operator RESIZE_BILINEAR (v2) refused by NNAPI delegate: Operator refused due performance reasons.
WARNING: Operator RESIZE_BILINEAR (v2) refused by NNAPI delegate: Operator refused due performance reasons.
WARNING: Operator RESIZE_BILINEAR (v2) refused by NNAPI delegate: Operator refused due performance reasons.
Applied NNAPI delegate.  # <- freeze

Thanks.

[Zhiming_Liu]

Hi @at_akihito 

Which bsp you use?

I tried it on L5.10.72_2.2.0,fail to use NNAPI even i use libvx_delegate in code.

[at_akihito]

@Zhiming_Liu 

thaks for your reply.

I'm using library based on gatesgarth-5.10.9-1.0.0

Thanks.