how quickly convert a model (*.py) to a graph and upload to NCS2

[TNguy219]

Hi

I have this model

http://www.chezmoi.dk/upload/YOLOV3-Tutorial-master.zip  (8MB)

https://i.imgur.com/ekvadYd.png

yolov3.weights, 248MB download from https://pjreddie.com/media/files/yolov3.weights

It ran successful on Google Colab. I want now implement this on my Windows10 with an "Intel NCS2"

How can i convert my trained model to a graph file on my windows 10 and upload to NCS2?

I this a quick/smart way to do it, is this possible? please help.

NB: I cant run on the model on my laptop due to an old nvidia card.

d:\_events\big data\object detection with tensorflow\YOLOV3-Tutorial-master>python tutorialDetect.py --images circus.jpg --det output

Loading network.....

Network successfully loaded

Traceback (most recent call last):

 File "tutorialDetect.py", line 97, in <module>

   model.cuda()

 File "D:\Program Files\Python37\lib\site-packages\torch\nn\modules\module.py", line 265, in cuda

   return self._apply(lambda t: t.cuda(device))

 File "D:\Program Files\Python37\lib\site-packages\torch\nn\modules\module.py", line 193, in _apply

   module._apply(fn)

[David_C_Intel · Accepted solution]

Hi TNguy219,

Thanks for reaching out.

First you need to install the latest OpenVINO™ toolkit release (refer to this link for Windows* 10 installation). Then, use the model optimizer, you can check this Yolo* models guide for converting your YoloV3 model to IR format using this tool. After obtaining your IR files, you can run inference on the Intel® Neural Compute Stick 2.

You can use the Object Detection YOLO* V3 Python* Demo to test your model with the Intel® Nerual Compute Stick 2.

If you have more questions let us know.

Best regards,

David C.

Intel Customer Support Technician

A Contingent Worker at Intel

[TNguy219]

Hello David C,

Thanks for the reply and the guide, i manage to make it. I wrote a guide on my language, which you can see here

https://drive.google.com/file/d/1pUsvsmX1-Y4dC0zfW6x59aHTqaf7kZ7F/view

look at page 30, the chapter "Covert và optimize pre-train model để load vào NCS2 Usb Stick"

You dont need to understand the language, look at the screendumps and the links.

I need to change 3 files before it works. Thanks to some posts i found from the Intel forum, all are list in the doc.

yolo_v3.json

d:\Program Files (x86) 2\IntelSWTools\openvino\deployment_tools\model_optimizer\mo\front\tf\extractor.py

d:\Program Files (x86) 2\IntelSWTools\openvino\deployment_tools\model_optimizer\mo\middle\passes\fusing\decomposition.py

Still too slow, I need to switch to yolov3-tiny, it is a lot better. But i think my old laptop need to decode the video stream so it takes extra time, better with a webcam, so no decode, only raw data from the cam.

Hoping this help for other who will try NCS2 in the future.

ps: if you plan to make a NCS3, please do it so it is backward compatible (openvino) and it will be great if it is a "single computer board", just like "Odroid XU4" (jetson nano).

THANKS

Tuan

[TNguy219]

just to mention the script mentioned from David C Object Detection YOLO* V3 Python* Demo , it is for video stream detection, not for an image detection. Wonder if there is a script for NCS2 which can take a image (or a folder with images) , put the square/label around the objects and then write to a file.

[David_C_Intel]

Hello TNguy219,

Thank you for your reply.

Great job by doing that guide on your own! Now more people can read and understand more about how OpenVINO™ toolkit works.

The Object Detection YOLO* V3 Python* Demo you checked is able to detect objects on images as well. You also can modify the ".py" code for it to read a directory with images inside.

In case you have additional questions, be free to ask.

Best regards,

David C.

Intel Customer Support Technician

A Contingent Worker at Intel