topic Re: OpenVino IR representation generation fails in Intel® Distribution of OpenVINO™ Toolkit

OpenVino IR representation generation fails

GSrin4 — Sat, 19 Oct 2019 13:37:23 GMT

Trying to convert YOLOv3 model definde in Keras. I have frozen the session as frozen_graph.pb in text format (so that I can edit the input shapes). When I try the IR generation using mo_tf.py script part of openvino tools. I am getting the following errors.

[ ERROR ]  List of operations that cannot be converted to Inference Engine IR:
[ ERROR ]      Min (3)
[ ERROR ]          Min
[ ERROR ]          Min_1
[ ERROR ]          Min_2
[ ERROR ]      Where (6)
[ ERROR ]          boolean_mask/Where
[ ERROR ]          boolean_mask_1/Where
[ ERROR ]          boolean_mask_2/Where
[ ERROR ]          boolean_mask_3/Where
[ ERROR ]          boolean_mask_4/Where
[ ERROR ]          boolean_mask_5/Where
[ ERROR ]      Round (3)
[ ERROR ]          Round
[ ERROR ]          Round_1
[ ERROR ]          Round_2
[ ERROR ]      NonMaxSuppressionV2 (3)
[ ERROR ]          non_max_suppression/NonMaxSuppressionV2
[ ERROR ]          non_max_suppression_1/NonMaxSuppressionV2
[ ERROR ]          non_max_suppression_2/NonMaxSuppressionV2
[ ERROR ]  Part of the nodes was not converted to IR. Stopped.

Details of OpenVino version: 2019.3.0-375-g332562022

Any suggestions/references would be much appreciated.

Re: OpenVino IR representation generation fails

David_C_Intel — Tue, 22 Oct 2019 03:42:25 GMT

Hi GSrin4,

Thank you for contacting Intel Customer Support, my name is David, I am glad to assist you.

I see you are having trouble with the IR generation. Could you please provide the following information:

What was the model optimizer command used to convert to IR format?
Is it a pre-trained model or a custom model?
Could you provide the orignal model?

Regards,

David C.

Intel Customer Support Technician

A Contingent Worker at Intel

Re: OpenVino IR representation generation fails

GSrin4 — Tue, 22 Oct 2019 08:32:45 GMT

Hi David,

Thank you for the reply.

Please find my comments for the query.

model optimizer command: python mo_tf.py --input_model_is_text --input_model frozen_graph.pb. (Input shape has been updated in the .pb file)
It's custom model.
Original model in the sense? I used this github repo: https://github.com/qqwweee/keras-yolo3
If you need the .h5 weight file, I can attach that. (The output of training is .h5 file in this case)

Hope it helps,

Gunasekaran

Re: OpenVino IR representation generation fails

David_C_Intel — Wed, 23 Oct 2019 02:10:14 GMT

Hi GSrin4,

Thank you for the information given.

It seems there is a flag missing: --tensorflow_use_custom_operations_config "yolo_v3.json".

You can check the documentation on how to convert YOLO models to IR here. As you are using a custom model, you will need to modify the "yolo_v3.json" file to match your model.

I hope this will help you. If you are still experiencing some errors, you can send me your ".pb" file, so we can manage to replicate your issue and get to a solution.

Regards,

David C.

Intel Customer Support Technician

A Contingent Worker at Intel

Re: OpenVino IR representation generation fails

David_C_Intel — Mon, 28 Oct 2019 22:54:51 GMT

Hi GSrin4,

I am contacting you to give a follow up on this case. Could you please tell me if the issue is solved?

I will help you further to any other questions you may have.

Regards,

David C.

Intel Customer Support Technician

A Contingent Worker at Intel