Intel® Distribution of OpenVINO™ Toolkit
Community support and discussions about the Intel® Distribution of OpenVINO™ toolkit, OpenCV, and all things computer vision-related on Intel® platforms.
5924 Discussions

Run yolox model with myriad vpu 2 & the inference results wrong

yangqb
Beginner
469 Views

VU9)RUXWG94YOO}@48X6WZY.pngGX{$17@FQALD8_~RF5UBIG7.png

 

Hello, 

I have trained a yolox for object detection and successfully converted yolox model -> onnx -> IR <Openvino format>. The inference results from cpu were right, while the inference results from myraid were wrong, and I found numbers in myriad results were all integers. 

Kindly help me resolving the issue. Thanks

0 Kudos
6 Replies
ChrisB_Intel
Moderator
427 Views

Thanks for the report, I am currently investigating and will update you shortly.

 

Update:

It looks like you are running OpenVINO on a local machine, versus in the DevCloud.  I am moving this over to the OpenVINO community and that team will be able to assist you.  

 

Thanks

Chris

 

 

Peh_Intel
Moderator
393 Views

Hi yangqb,


Try to re-generate IR model using the Model Optimizer by specifying the scale value and obtain the inference results from MYRIAD again. The scale value should be up to 255. Adding --scale will have the data values and weights analyzed by the MYRIAD plugin correctly and there will be no overflow.


For example:

python3 mo.py --input_model yolox.onnx --scale 255

 

 

Regards,

Peh


yangqb
Beginner
352 Views

Hi Peh,

Thank you fore your reply, I have tried your suggestion but only find it doesn't effect, the outputs still be integers.

 

Regards,

Yang

 

Peh_Intel
Moderator
316 Views

Hi yangqb,

 

I obtained the same runtime overflow issue as you when using OpenVINO 2021.3. However, this issue is solved when using OpenVINO 2021.4 or OpenVINO 2021.4.1.


I’ve validated the inference results from Intel® Neural Compute Stick 2 are correct when using the latest OpenVINO version. Do not re-generate IR model using the Model Optimizer by specifying the scale value which can lead to wrong results.


Hence, I would suggest you download the latest OpenVINO version, OpenVINO 2021.4.1.



Regards,

Peh


yangqb
Beginner
289 Views

Hi Peh,

I have got the right inference results from the Intel® Neural Compute Stick 2 after upgrading the OpenVINO to version 2021.4.1, and the issue troubled me for several months is solved, and I am so excited! Thank you very much for your help!!!

 

Best regards,

Yang

Peh_Intel
Moderator
286 Views

Hi Yang,

 

Thank you for confirming the issue is resolved. I'm glad too that I was able to help.  

 

This thread will no longer be monitored since this issue has been resolved. If you need any additional information from Intel, please submit a new question.

 

 

Regards,

Peh

 

Reply