Community
cancel
Showing results for 
Search instead for 
Did you mean: 
Highlighted
Beginner
31 Views

Unable to convert yolo_v3 Model

Jump to solution

I was following the guide for creating a yolo_v3.pb file and the converting that with the model optimizer to the intermediate format. Unfortunately I'm hitting this error and I'm not really sure what the problem is. ​​​​ Debug log:

 
0 Kudos

Accepted Solutions
Highlighted
Valued Contributor I
31 Views

Hello MESO,

Good to hear it is fixed now.

> Would you mind elaborating where you got those numbers from and what do they mean exactly in this scenario.

Based on 

https://pjreddie.com/darknet/yolo/

-- YOLOv3-416COCO trainvaltest-dev55.365.86 Bn35cfgweights

and since the sample yolo app in OpenVino supports 416x416 we converted size YOLO v3 416 first

You could also try 608x608 or 320x320 but may need to modify the sample app.

--data_type=FP32 tells the model optimizer to generate FP32 IR (bin and xml files). You could use FP32 for CPU or GPU (-d CPU or -d GPU)

For MYRIAD NCS or GPU you could also run with FP16 IR . Then you need to generate IR again using --data_type=FP16

cheers, 

nikos

View solution in original post

0 Kudos
6 Replies
Highlighted
Valued Contributor I
31 Views

Hello MESO,

There is good OpenVino documentation and many threads in this forum dealing with this issue; also good examples in github.

You may find good solution after reading for example:

https://software.intel.com/en-us/forums/computer-vision/topic/800813

https://software.intel.com/en-us/forums/computer-vision/topic/800049

https://software.intel.com/en-us/forums/computer-vision/topic/801626

Let us know if you still have issues.

Please post command to export to pb and also for the model optimizer.

nikos

 

0 Kudos
Highlighted
Beginner
31 Views

Hi Nikos,

thank you for those links. The solution from you from the first link did the trick.
Would you mind elaborating where you got those numbers from and what do they mean exactly in this scenario.
I'm referring to these two parameters that seem to fix the conversion:

--input_shape=[1,416,416,3] --data_type=FP32

How do I know that the model is in a 32bit floating point format and what does the input shape mean and how do I get to that number?

Thanks

0 Kudos
Highlighted
Valued Contributor I
32 Views

Hello MESO,

Good to hear it is fixed now.

> Would you mind elaborating where you got those numbers from and what do they mean exactly in this scenario.

Based on 

https://pjreddie.com/darknet/yolo/

-- YOLOv3-416COCO trainvaltest-dev55.365.86 Bn35cfgweights

and since the sample yolo app in OpenVino supports 416x416 we converted size YOLO v3 416 first

You could also try 608x608 or 320x320 but may need to modify the sample app.

--data_type=FP32 tells the model optimizer to generate FP32 IR (bin and xml files). You could use FP32 for CPU or GPU (-d CPU or -d GPU)

For MYRIAD NCS or GPU you could also run with FP16 IR . Then you need to generate IR again using --data_type=FP16

cheers, 

nikos

View solution in original post

0 Kudos
Highlighted
Beginner
31 Views

Ah I see but what is the meaning of the 1 and 3 then? According to the help for the model optimzer it should describe the dimensional size.
Intuitively I would think we are dealing with a 2D input not 4D? 3 I can see maybe relate to the color channels but then I'm not really sure what the 1 would mean.

0 Kudos
Highlighted
Innovator
31 Views
@MESO It is --input_shape=[batch size, height, width ,channel]
0 Kudos
Highlighted
Beginner
31 Views

Ah ok I see, did not see that info the the help the cli was printing. Thanks for all the help.

0 Kudos