build from source and install - please help

[Pedreira_Gonzalez__I]

Hello

I want to use last version openzoo_models on raspbian, and with the current release seems not possible.

That seems to be because is need to apply those patch... dont know if those are on current master branch of openvino
https://github.com/NervanaSystems/ngraph/pull/4329
https://github.com/NervanaSystems/ngraph/pull/4745

So im trying to build from source, i managed to build but now i see that the output estructure is not the same estructure as the openvino package you download.

I want to know how to install the package that i built or how can i built with the same estructure as the one that you download.

And if you accept a suggestion from an end user, please add some notes to build readme to help people to install or make a new release from rpi.... please!

Thank you

[Iffa_Intel]

Greetings,

May I get the clarification specifically on which version of your previous and current version(eg 2020.2)?

Sincerely,

Iffa

[Pedreira_Gonzalez__I]

Hello

Thanks for your reply

i tried packages 2020.1 and 2020.2 and i get an error trying to use 2020 models, so i have to use 2019 R3 models but i want to test new face models that are on 2020 releases

For example model downloaded from this
https://download.01.org/opencv/2020/openvinotoolkit/2020.2/open_model_zoo/models_bin/3/face-detection-adas-0001/FP16/

Produce this error
terminate called after throwing an instance of 'InferenceEngine::details::InferenceEngineException' what(): Check 'input_shape_product.get_length() == shape_size(m_output_shape)' failed at /teamcity/work/scoring_engine_build/releases_2020_2/ngraph/src/ngraph/op/reshape.cpp:100: While validating node 'v0::Reshape Reshape_971(PriorBox_744[0]:f32{2,3888336896}) -> (dynamic?)': Product of output shape dimensions does not match product of argument shape dimensions (output shape: Shape{1, 2, 3888336896}, argument shape: {2,3888336896}).

[Pedreira_Gonzalez__I]

Note

Im using opencv as current release native engine is compiled for python3.5 and buster has 3.7, so on next release... please update python version if you can

[Pedreira_Gonzalez__I]

I finally manage to build and install from source but still getting the same error

With opencv

Traceback (most recent call last):

File "./demo.py", line 23, in <module>
out = net.forward()
cv2.error: OpenCV(4.3.0-openvino) ../opencv/modules/dnn/src/ie_ngraph.cpp:600: error: (-2:Unspecified error) in function 'initPlugin'

> Failed to initialize Inference Engine backend (device = MYRIAD): Check 'input_shape_product.get_length() == shape_size(m_output_shape)' failed at /home/pi/openvino/ngraph/src/ngraph/op/reshape.cpp:100:
> While validating node 'v0::Reshape Reshape_18831(PriorBoxClustered_18234[0]:f32{2,2760708352}) -> (dynamic?)':
> Product of output shape dimensions does not match product of argument shape dimensions (output shape: Shape{1, 2, 2760708352}, argument shape: {2,2760708352}).

And same error on native engine

>>> import cv2

>>> from openvino.inference_engine import IECore, IENetwork

>>> ie = IECore()

>>> net = ie.read_network('face-detection-0100.xml', 'face-detection-0100.bin')

Traceback (most recent call last):

  File "<stdin>", line 1, in <module>

  File "ie_api.pyx", line 136, in openvino.inference_engine.ie_api.IECore.read_network

  File "ie_api.pyx", line 157, in openvino.inference_engine.ie_api.IECore.read_network

RuntimeError: Check 'input_shape_product.get_length() == shape_size(m_output_shape)' failed at /home/pi/openvino/ngraph/src/ngraph/op/reshape.cpp:100:

While validating node 'v0::Reshape Reshape_1991(PriorBoxClustered_1394[0]:f32{2,3443079936}) -> (dynamic?)':

Product of output shape dimensions does not match product of argument shape dimensions (output shape: Shape{1, 2, 3443079936}, argument shape: {2,3443079936}).

[Pedreira_Gonzalez__I]

Hello

Can someone please tell me if im loosing my time trying to make this work¿

I don't understand what happens and will be nice to have info from intel about that.

Best Regards

[Iffa_Intel]

Hi,

These are some steps that you can try:

1. Try to run the default sample from the latest version of openvino that you use and see whether you get the same error.

2. Try to duplicate the original sample into a new folder and copy your old algo there & run... same concept as using (cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_FLAGS="-march=armv7-a" /opt/intel/openvino/deployment_tools/inference_engine/samples) 

Note :  /opt/intel/openvino/deployment_tools/inference_engine/samples is the location of file that is being duplicated, put your own sample there.

3. Another option you can try is to get your old .bin and .xml into new Openvino package.

Here is the official and full tutorial on properly run Openvino samples in Raspbian:

https://docs.openvinotoolkit.org/latest/_docs_install_guides_installing_openvino_raspbian.html#run-sample

4. Your error seems to originate with sizing issues,here is some ideas of how you can try to troubleshoot size error:

https://docs.openvinotoolkit.org/2020.2/_docs_IE_DG_ShapeInference.html

Hope my answer helps!

Sincerely,

Iffa

[Pedreira_Gonzalez__I]

Thank you

Error is a BUG in openvino for raspbian, is solved now on openvino master branch.
https://github.com/openvinotoolkit/openvino/issues/411

[Iffa_Intel]

Glad that you found that,

Openvino version 2020.1 does have problems previously with Raspbian/Raspberry Pi where most of the time  users end up reverting back to 2019 R3.

I've tried both 2020.1 and 2019 R3 . Seems 2020.1 doesn't works for me, while 2019 R3 works.

The Openvino 2020.2 should fix the issue, however, I didn't have a chance to test it out yet.

Nonetheless, someone from this community tried and it works for him :  https://github.com/openvinotoolkit/openvino/issues/411

Sincerely,

Iffa

[Pedreira_Gonzalez__I]

The problem happens when you want to use 2020 models, you cant.

Im natxopedreira on github and yes it works, but you have to build openvino and opencv from source.

Hope next release will solve

[Iffa_Intel]

Noted & thanks for your feedback.

Sincerely,

Iffa