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.
5541 Discussions

YoloV2 reported error in Neural Compute Stick2

Deng__Mike
Beginner
149 Views

hi , I run Yolo2 by Openvion, FP32 is ok In CPU mode, but it report error in MYRIAD with FP16, anybody know why?

E: [xLink] [         0] dispatcherEventReceive:308    dispatcherEventReceive() Read failed -4 | event 0x7f9979ffabb0 USB_READ_REL_RESP


E: [xLink] [         0] eventReader:254    eventReader stopped

E: [xLink] [         0] dispatcherWaitEventComplete:694    waiting is timeout, sending reset remote event

E: [ncAPI] [         0] ncFifoReadElem:2853    Packet reading is failed.

E: [ncAPI] [         0] ncFifoDestroy:2672    Failed to write to fifo before deleting it!

0 Kudos
4 Replies
Cary_P_Intel1
Employee
149 Views

Hello, Mike,

Yolo v2 has layer which MYRIAD VPU doesn't support called "RegionYolo" so you can't run YOLO example on NCS device.

Please refer to the table of supported layer for each device from link below - navigate to "Advanced Topics/supported Devices/Supported Configurations"

https://software.intel.com/en-us/articles/OpenVINO-InferEngine

 

Deng__Mike
Beginner
149 Views

Hi Cary,

   thanks for your response.

  as we know, Openvino can run custom layer in CPU, for Example in FPGA, we can add " -d  HETERO:FPGA,CPU" ,

do we have way to run RegionYolo in CPU?

thanks.

 

Shaukat__Maria
Beginner
149 Views

Hey Deng, 

I am also trying to run YOLO v2 and YOLO v2 tiny on NCS2 using OpenVINO. 

I have converted yolo.pb model to IR files with FP16 precision. 

However when I run it with OpenVINO via -d MYRIAD, the result of inference gives me all 0s. I think the problem here is that MYRIAD does not support Region Layers. Have you figured out a way to distribute the layers of YOLO v2 to run Region layers on CPU while other layers on MYRIAD? 

Monica__Zhao
Beginner
149 Views

Hello, Shaukat, Maria. Have you solved your problem? I converted tiny-yolo-v2 model into an IR file with FP16 accuracy. The result inferred to me was not a three-dimensional matrix, but a long eigenvector. There were many values of 0 and 1 in it. After analysis, the effect was very poor. But it was sure that the model before conversion was very good, and the effect of converting into graph on the generation of neural rods was also very good, but converted into bin and XML text. The effect on the second generation nerve rod is very poor. I don't know why?

Reply