How to make python script work when using openvino on arm64

[lynn1050]

I wanted to use NCS2, and installed openvino on Nvida TX2, following this guide:

https://community.intel.com/t5/Intel-Distribution-of-OpenVINO/How-to-install-Openvino-on-Ubuntu-Arm64-architecture/m-p/1222442

https://software.intel.com/content/www/us/en/develop/articles/arm64-sbc-and-ncs2.html

Installation was successful, but when I ran script ./setupvars.sh , I got this error:

[setupvars.sh] WARNING: Can not find OpenVINO Python binaries by path /home/dajian/Code/openvino/scripts/python

[setupvars.sh] WARNING: OpenVINO Python environment does not set properly

[setupvars.sh] OpenVINO environment initialized

In a result, the command below works well:

./bin/aarch64/Release/object_detection_sample_ssd -m face-detection-adas-0001.xml -d MYRIAD -i jialin.jpg

But  I didn't succeed in running my python script that used openvino api.  "No module named 'openvino' " occurred.

Any suggetion will be appreciated.

[Rizal_Intel]

Hi Lynn,

Are you able to get the benchmark app to run?

I do not think the instructions you followed build the python libraries.

You may need to build and use existing C++ examples or develop them.

Regards,

Rizal

[lynn1050]

 Yes, I am able to get the benchmark app to run, because the excutable binary file benchmark was build from c++ code. I may build a c++ project to get my app.

But there is another problem. When I run benchmark app with option '-d CPU',  I got this error,

[ ERROR ] Device with "CPU" name is not registered in the InferenceEngine

It seems to support MYRIAD only. And I want to do th comparison among CPU, GPU, or MYRIAD.

[Rizal_Intel]

Hi Lynn,

The CPU and GPU plugin support Intel hardware only.

ARM based processor devices such as Jetson is not compatible for the CPU plugin.

The NVIDIA graphics card is also not compatible with OpenVINO and should be used with the CUDA interface instead.

The only compatible devices are MYRIAD devices such as the Neural Compute Stick 2 or Intel® Vision Accelerator.

You can refer to the list of supported devices for more information.

Regards,

Rizal

[lynn1050]

Thank you for your information. It's very helpful.

Then I want to know how to build the python libraries so that I  can develop in python code and do some comparison. 

[jonnymo]

I have the same question. I followed the instructions but there does not appear to be any python folder created.

Also, the SqueezeNetv1.1 example is not present.

How to fix?

[Rizal_Intel · Accepted solution]

You can follow this building OpenVINO from source wiki guide.

Depending on your OS distribution, you may also need to compile Python from source.

Regards,

Rizal

[jonnymo]

Ah, yeah I had seen this previously and did try to specify the Python version but it was failing.  With the latest OpenVINO code downloaded, 2021.2, this seemed to work. I was able to run cmake and make without issue.

I added the following to the cmake options:

NOTE: Not specifying the Python version results in a Python2.7 build.

    -DENABLE_PYTHON=ON \
    -DPYTHON_EXECUTABLE=`which python3.8` \
    -DPYTHON_LIBRARY=/usr/lib/aarch64-linux-gnu/libpython3.8.so \
    -DPYTHON_INCLUDE_DIR=/usr/include/python3.8

This did require installing Cython before running cmake though.

sudo apt install python3-pip
pip3 install Cython

Thanks.

[jonnymo]

From the following link, I was able to test that the OpenVINO Python libs were created and working as expected:

https://www.intel.com/content/www/us/en/support/articles/000057448/software/development-software.html 

odroid@odroid:~/develop/openvino2$ export PYTHONPATH=$PYTHONPATH:/home/odroid/develop/openvino2/openvino/bin/aarch64/Release/lib/python_api/python3.8/
odroid@odroid:~/develop/openvino2$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/odroid/develop/openvino2/openvino/bin/aarch64/Release/lib/
odroid@odroid:~/develop/openvino2$ python3.8
Python 3.8.5 (default, Jul 28 2020, 12:59:40)
[GCC 9.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from openvino.inference_engine import IENetwork, IECore
>>>

There was no error on the import, so it appears to be good.

The issue I found with the setupvars.sh script is that it is expecting the OpenVINO Python libs to be located under {INTEL_OPENVINO_DIR}/python and {INTEL_OPENVINO_DIR}/python/python$python_version}. 

However, with the source build as I did, the Python libs are not placed here, but rather in my case at:

 $INTEL_OPENVINO_DIR/bin/aarch64/Release/lib/python_api/python$python_version

Also, when the script is run from say the OpenVINO directory, it sets the BASE_DIR to the SCRIPT_DIR, and other paths, to the script root directory.  Printing these out from the script, you can see what they are set to:

source /home/odroid/develop/openvino/scripts/setupvars/setupvars.sh
SCRIPT_DIR=/home/odroid/develop/openvino/scripts/setupvars
BASE_DIR=/home/odroid/develop/openvino/scripts
INSTALLDIR=/home/odroid/develop/openvino/scripts

Even though the script fails, it still sets the PYTHONPATH, which is based on the INSTALLDIR which is what the INTEL_OPENVINO_DIR is set to:

~$ echo $PYTHONPATH
/home/odroid/develop/openvino/scripts/deployment_tools/model_optimizer

The model_optimizer folder does not exist, and the deployment_tools are located at the following with no model_optimizer folder:

$ find  /home/odroid/develop/openvino -name "deployment_tools"
/home/odroid/develop/openvino/build/ngraph/CMakeFiles/Export/deployment_tools

I ended up changing the setupvars.sh script a bit to accommodate the source build and now succeeds but still needs work.

I changed the beginning to:

     17 BASE_DIR="${PWD}"
     18 SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" && pwd )" >/dev/null 2        >&1 && pwd )"
     19 #BASE_DIR="$( dirname "$SCRIPT_DIR" )"
     20
     21 INSTALLDIR="${BASE_DIR}"
     22
     23 echo "INSTALLDIR=${INSTALLDIR}"

and added an elif to accommodate the the actual path for the build:

    144     elif [[ -d $INTEL_OPENVINO_DIR/bin/aarch64/Release/lib/python_api/py        thon$python_version ]]; then
    145         echo "Python path found!"
    146         # add path to OpenCV API for Python 3.x
    147         #export PYTHONPATH="$INTEL_OPENVINO_DIR/python/python3:$PYTHONPA        TH"
    148         pydir=$INTEL_OPENVINO_DIR/bin/aarch64/Release/lib/python_api/pyt        hon$python_version
    149         if [[ -d $pydir ]]; then
    150             # add path to Inference Engine Python API
    151             export PYTHONPATH="${pydir}:${PYTHONPATH}"
    152         else
    153             echo "[setupvars.sh] WARNING: Can not find OpenVINO Python m        odule for python${python_version} by path ${pydir}"
    154             echo "[setupvars.sh] WARNING: OpenVINO Python environment do        es not set properly"
    155         fi
    156     else

The "-pyver" option can be used to set the Python version for the Python libs. In my case it was 3.8.

Thus running the script ends with:

$ source /home/odroid/develop/openvino2/openvino/scripts/setupvars/setupvars.sh -pyver 3.8
INSTALLDIR=/home/odroid/develop/openvino2/openvino
python_version = 3.8
Python path found!
[setupvars.sh] OpenVINO environment initialized

The PYTHONPATH is set as such:

$ echo $PYTHONPATH                    /home/odroid/develop/openvino2/openvino/bin/aarch64/Release/lib/python_api/python3.8:/home/odroid/develop/openvino2/openvino/deployment_tools/model_optimizer

The script needs more work, but it is close.

Cheers,

Jon

[lynn1050]

 It works well. Thanks a lot!

[Munesh_Intel]

Hi Lynn,

This thread will no longer be monitored since this issue has been resolved. If you need any additional information from Intel, please submit a new question.

Regards,

Munesh