Can't read openvino pre-trained model in opencv.

[chinmaydas96]

I was running the following line in raspberr pi:

import cv2

cv2.dnn.readNet('person-detection-retail-0002/person-detection-retail-0002.xml',
'person-detection-retail-0002/person-detection-retail-0002.bin')

the error is following :

**
terminate called after throwing an instance of 'InferenceEngine::details::InferenceEngineException'
what(): Check 'axis < static_cast<size_t>(input_rank)' failed at /teamcity/work/scoring_engine_build/releases_2020_1/ngraph/src/ngraph/op/gather.cpp:140:
While validating node 'Gather[Gather_928](patternLabel_924: float{10,20,30}, patternLabel_925: int64_t{5}, patternLabel_927: int64_t{1}) -> (??)':
The axis must => 0 and <= input_rank (axis: 4294967295).

Aborted
**

I have already build the opencv with openvino.
import cv2 as cv

In [2]: print(cv.file)
/home/pi/openvino/python/python3/cv2.abi3.so`

[maxlytkin]

Hello @chinmaydas96

There's a known case with the latest models due to changes in the Model Optimizer and ngraph in latest OpenVINO toolkit 2020.1 build. The "axis" error is caused by the new IR v10 format.

As a workaround, you can use one of the following:

1. Either use the models from the previous release (2019 R3) of open model zoo - https://download.01.org/opencv/2019/open_model_zoo/R3/20190905_163000_models_bin/
2. Or generate the IRv7 format models using the Model Optimizer from the OpenVINO toolkit 2020.1 build with the following parameter: --generate_deprecated_IR_V7

Please let us know if this works for you or not.

Thanks.
Best regards, Max.

[IRDonch]

This looks like an issue with OpenCV, so I'm moving it to the OpenCV repo.

[alalek]

This is DLDT issue actually (OpenCV uses IR reader as is).

[dkurt]

That's right. This is nGraph issue fixed by NervanaSystems/ngraph#4329

[dkurt]

@chinmaydas96, Can you please try the latest 2020.2 release?

[fischermario]

I got the same error message as @chinmaydas96 while trying to run the face detection test as advised by the OpenVINO Raspbian guide (https://docs.openvinotoolkit.org/latest/_docs_install_guides_installing_openvino_raspbian.html#run-inference-opencv).

With the new 2020.2 release I ran the test again and got this error message instead:

Traceback (most recent call last):
  File "openvino_fd_myriad.py", line 12, in <module>
    _, confidences, boxes = net.detect(frame, confThreshold=0.5)
cv2.error: OpenCV(4.3.0) /home/nvidia/Downloads/opencv-4.3.0/opencv/modules/dnn/src/ie_ngraph.cpp:600: error: (-2:Unspecified error) Failed to initialize Inference Engine backend (device = MYRIAD): Unexpected CNNNetwork format: it was converted to deprecated format prior plugin's call in function 'initPlugin'

This points to the same problem as before while the error description is much better this time.

Btw. I have used this model for testing with 2020.2:
https://download.01.org/opencv/2020/openvinotoolkit/2020.2/open_model_zoo/models_bin/1/face-detection-adas-0001/FP16/face-detection-adas-0001.xml
https://download.01.org/opencv/2020/openvinotoolkit/2020.2/open_model_zoo/models_bin/1/face-detection-adas-0001/FP16/face-detection-adas-0001.bin

(Out of curiosity: What is the difference between the models in the three directories "1", "2", "3" inside https://download.01.org/opencv/2020/openvinotoolkit/2020.2/open_model_zoo/models_bin/?)

@dkurt: If this behavior was really fixed by the mentioned pull request (which was merged into master on February 13th) then reading the new model format should work with OpenVINO 2020.2 because it relies on a ngraph version (edc65ca0111f86a7e63a98f62cb17d153cc2535c) from March 31st.

[alalek]

@fischermario Mentioned problem with "deprecated format" in VPU plugin should be workarounded in OpenCV here: opencv/opencv#17134

[fischermario]

@alalek: Thank you. I will patch OpenCV 4.3.0 with the changes from the pull request and report back.

[fischermario]

@alalek: The patch solves the issue entirely 😊 Thanks again! Will there be a patch release for the OpenCV 4.3.x series?

[ilya-lavrenov]

@fischermario is the issue resolved? can be closed?

[s-t-e-p-a-n-o-v]

@fischermario is the issue resolved? can be closed?

Nop, mensioned patch for opencv opencv/opencv#17134 is already included in opencv-4.3-openvino-2020.3 release
https://github.com/opencv/opencv/tree/4.3.0-openvino-2020.3.0/modules/dnn/src
dnn.cpp and test_ie_models.cpp

Just tested on openvino 2020.3 and opencv 4.3.0-openvino-2020.3 on ARM64 target:
Models from here https://download.01.org/opencv/2019/open_model_zoo/R3/20190905_163000_models_bin/ works, but fresh models from here https://download.01.org/opencv/2020/openvinotoolkit/2020.3/open_model_zoo/models_bin/1/ doesn't work.

Same error:
cv2.error: OpenCV(4.3.0-openvino-2020.3.0) /home/ubuntu/opencv-4.3.0-openvino-2020.3.0/modules/dnn/src/op_inf_engine.cpp:881: error: (-2:Unspecified error) Failed to initialize Inference Engine backend (device = MYRIAD): Unexpected CNNNetwork format: it was converted to deprecated format prior plugin's call in function 'initPlugin'

[dkurt]

@s-t-e-p-a-n-o-v, Can you please try to test the latest master branch of OpenCV?

[s-t-e-p-a-n-o-v]

I tried without success with 4.4.0-pre

There was build error:
[ 76%] Linking CXX executable ../../bin/opencv_perf_dnn
/usr/bin/ld: /home/ubuntu/openvino/bin/aarch64/Release/lib//libinference_engine_s.a(ie_core.cpp.o): undefined reference to symbol 'dlsym@@GLIBC_2.17'
/usr/lib/gcc/aarch64-linux-gnu/7/../../../aarch64-linux-gnu/libdl.so: error adding symbols: DSO missing from command line
collect2: error: ld returned 1 exit status
modules/dnn/CMakeFiles/opencv_perf_dnn.dir/build.make:175: recipe for target 'bin/opencv_perf_dnn' failed
make[2]: *** [bin/opencv_perf_dnn] Error 1
CMakeFiles/Makefile2:2328: recipe for target 'modules/dnn/CMakeFiles/opencv_perf_dnn.dir/all' failed
make[1]: *** [modules/dnn/CMakeFiles/opencv_perf_dnn.dir/all] Error 2
make[1]: *** Waiting for unfinished jobs....

I created separate issue for this - #1221