Model Optimizer of OpenVino Release R5 2018 corrupts models

[Ezzat__Ahmed]

Environment

• OpenVino R5
• Neural Compute Stick 2
• Windows 10 (Ubuntu also tested and the same issue exists)
• SSD_Inception_V2
• YOLOv3

I have the model SSD Inception V2 working fine with the same setup with the MODEL_OPTIMIZER of Release R4. When Compile the models with the new release (after upgrading the openvino setup to the R5), the model comipiled with no errors but the inference results were corrupted.

I opened an Issue about the same problem for YOLOv3, but never got a helpful answer. I do not know what is the change in the new release that breaks my models?

I attached the 2 xml files for the SSD_Inception_V2 (working was compiled using R4 model_opt and not_working was compiled using r5 model_opt). 

For complete debug setup, see the linked Issue for detailed SW package for YOLOv3.

I have a clean environment. I removed the old version of the tool and installed the R5 SDK, so I do not think it is a problem in the environment. The old model for SSD_Inception_V2 works with the new inference engine. 

That is not limited to these models only, that is for all SSD Models I am trying. The new model optimizer causes wrong inference results. The symptom of the error is exactly the same:

There was a permute layer before the reshape layer in the working models. The new model optimizer removes the permute layer before the reshape. See the 2 attached files for the case (1st appearance of this strange behavior at Line 1883 of <ssd_inception_v2_vino_WORKING.xml> and <ssd_inception_v2_vino_NOT_WORKING.xml>

Regards,

Ahmed

[Anna_B_Intel]

Hi Ahmed,

Make sure you converted the model using these instructions: https://docs.openvinotoolkit.org/R5/_docs_MO_DG_prepare_model_convert_model_tf_specific_Convert_Object_Detection_API_Models.html

Instructions for YOLO: https://docs.openvinotoolkit.org/R5/_docs_MO_DG_prepare_model_convert_model_tf_specific_Convert_YOLO_From_Tensorflow.html

Best wishes,

Anna

[Ezzat__Ahmed]

Anna B. (Intel) wrote:

Hi Ahmed,

Make sure you converted the model using these instructions: https://docs.openvinotoolkit.org/R5/_docs_MO_DG_prepare_model_convert_model_tf_specific_Convert_Object_Detection_API_Models.html

Instructions for YOLO: https://docs.openvinotoolkit.org/R5/_docs_MO_DG_prepare_model_convert_mo...

Best wishes,

Anna

Thanks Anna for response.

I know how to convert a model. I already had a working model with the OPENVINO.R4 for the SSD Inception V2. 

My issue is in the model optimizer of OPENVINO.R5. I used the same command to convert the model, but the resulting model gives corrupted detection results which was not the case with the model converted using the older release. Is there a change in the command line parameters for the R5 release which I am not aware of?

Regards,

Ahmed

[Shubha_R_Intel]

Dear Ahmed:

Is your issue similar to this forum issue ? Can you kindly confirm ?

https://software.intel.com/en-us/forums/computer-vision/topic/805570

Thanks for using OpenVino !

Shubha