openvino inference

[Farhâd]

Trying to run inference with openvino model with yolov5 and get error:

net = ie.read_network(model=model)
File "ie_api.pyx", line 367, in openvino.inference_engine.ie_api.IECore.read_network
File "ie_api.pyx", line 410, in openvino.inference_engine.ie_api.IECore.read_network
RuntimeError: Cannot create Interpolate layer /model.11/Resize id:187 from unsupported opset: opset11

[Farhâd]

what would you do to convert ONNX to IR using Model Optimizer command, if the ONNX model has layers like RESIZE and INTERPOLATE in it? I think I am getting the above error in inference because of these unsupported layers in ONNX which i tried converting to IR.

[Farhâd]

[Farhâd]

I used this command for converting ONNX to IR ( yolov5 )

mo --input_model best_.onnx --model_name best_IR -s 255 --input_shape=[1,3,640,640] --input=images --reverse_input_channels --output /model.24/m.0/Conv,/model.24/m.1/Conv,/model.24/m.2/Conv

[Aznie_Intel]

Hi Farhad,

Thanks for reaching out.

RESIZE is a supported layer for ONNX but INTERPOLATE is not a supported layer in ONNX framework. You may refer to this ONNX Supported Operators.

For the unsupported layer, you may use OpenVINO Extensibility Mechanism to customize a new operation.

Hope this helps.

Regards,

Aznie

[Farhâd]

I did run the same ONNX to IR conversion command line a few months ago and it worked.

The error says: "RuntimeError: Cannot create Interpolate layer /model.11/Resize id:187 from unsupported opset: opset11". Is there a version problem? If so please give me a conversion command line with a version that works.

[Aznie_Intel]

Hi Farhad,

Which OpenVINO version have you been using before? Are you converting with the same OpenVINO version or you are using a different version now?

On another note, you may share your model for me to validate from my end.

Regards,

Aznie

[Farhâd]

I am using the latest MO as well as the latest openvino-dev .  i am running this job on colab which always install the latest versions. the new version of openvino must be incompatible with the older version of MO .

! pip install openvino-dev

! git clone --depth 1 https://github.com/openvinotoolkit/openvino.git

! mo --input_model best.onnx --model_name best -s 255 --input_shape=[1,3,640,640] --input=images --reverse_input_channels --output /model.24/m.0/Conv,/model.24/m.1/Conv,/model.24/m.2/Conv

[Aznie_Intel]

Hi Farhad,

Starting from the 2020.4 release, OpenVINO™ supports reading native ONNX models. You may use your ONNX model without converting if nothing's wrong with it/supported. Can you try to run your ONNX model with benchmark_app?

For us to further validate this, is it possible for you to share your model?

Regards,

Aznie

[Farhâd1]

My question is about openvino IR not ONNX.

[Aznie_Intel]

Hi Farhad,

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. 

Regards,

Aznie