I have a squeeze net network trained with Matlab which is used for image classification
I converted it to an onnx file and then converted it to an IR file, hoping to use it on NCS2
The conversion process is very smooth, with no errors
But when executed on Pi4, an Error popped up, as shown in the attached picture.
I would like to ask if Matlab is still not accepted by ncs2 even if it is converted to onnx file?
Thanks for reaching out to us.
Are you using an older version of Intel® Distribution of OpenVINO™ Toolkit? For your information, the error you encountered: “Cannot create ShapeOf layer …” is due to the unsupported network.
You may refer to a similar thread here for more information.
It is recommended to use the latest version of the Intel® Distribution of OpenVINO™ Toolkit which some of the known issues have been fixed on the latest release (2021.4.2).
On another note, could you please share the following information with us to further assist you?
· Version of Intel® Distribution of OpenVINO™ Toolkit installed on your Raspberry Pi 4
· SqueezeNet topology that trained with MATLAB, command that you used to convert into ONNX and Intermediate Representation
· Sample script (classification_sample.py)
First of all thanks for your reply
My Intel® OpenVINO™ toolkit distribution on Raspberry Pi 4 is 2020.1
The attached word file is the SqueezeNet topology trained with MATLAB and the commands used to convert to ONNX and intermediate representation
I am new to Openvino so my classification_sample.py is from
The classification_sample.py in /opt/intel/openvino/deployment_tools/inference_engine/samples/python/classification_sample/ has not been changed
Thanks for sharing all the requested information.
Could you please share your ONNX model with us so that we can replicate your issue?
Also, please share the Model Optimizer conversion command that you used to convert your ONNX model into IR.
Thanks for your reply.
Since I changed to the latest version of the Intel® Distribution of OpenVINO™ Toolkit latest release (2021.4.2
The problem is solved. Now there is a new problem with the classification result. I have 10 kinds of pictures to be classified, but the classid obtained by inputting different classes of pictures are all the same.
Glad to know the latest version of the Intel® Distribution of OpenVINO™ Toolkit solved your previous problem.
Here is the command that I used:
python3 classification_sample_async.py -i <path_to_image_1> <path_to_image_2> -m <path_to_model>/squeezenet1.1.xml --labels <path_to_label>/squeezenet1.1.labels
Could you please share your Intermediate Representation files and labels mapping file with us to further assist you?
Thank you for your question.
If you need any additional information from Intel, please submit a new question as this thread is no longer being monitored.