Custom neural network execution.

martin181818

Hi all ,

I am working on a project based on object detection and segmentation with 2 different neural networks.
The part of object detection is always working when I run the script but the segmentation part using the NeuralNetwork node is executed only on demand ( user request via TCP or with a timer ) so it will be executed only once or twice an hour. The segmentation model is large and will decrease the FPS value to below 1 FPS if I link it directly to the pipeline and make it always active. Is there any solution to make it execute only on specific image when the user asks for it and then stops while the other pipeline keeps working ? what are your suggestions ?

Thank you very much

erik

Hi @martin181818 ,
I'd suggest using Script node:
https://docs.luxonis.com/projects/api/en/latest/samples/Script/script_change_pipeline_flow/#script-change-pipeline-flow
It's a good example, as it turns NN inferencing on/off based on user input (via XLinkIn). I hope this helps!

martin181818

Perfect. That's exactly what I needed. Thank you very much for the support

Link92

@erik My project is based on standalone mode, In the example you did provide there are the following three lines of code:

buf = dai.Buffer()
buf.setData(DETECTION)
switch.send(buf)

Any idea on how to make it work in standalone? The problem is related to the call to dai.Buffer() which cannot be imported within the script node.

Thanks in advance for your help.

erik

Hi @Link92 ,
You can just string (or maybe JSON) instead of dai.Buffer. Example here:
https://docs.luxonis.com/projects/api/en/latest/samples/Script/script_json_communication/#script-json-communication

Unrelated (as you should use JSON), but you can create dai messages inside script node:
https://docs.luxonis.com/projects/api/en/latest/components/nodes/script/#using-depthai-messages

Link92

erik Unfortunately dai.Buffer() is also used in this example (lines 40:42):

buffer = dai.Buffer()
buffer.setData(list(data))
device.getInputQueue("in").send(buffer)

erik

@Link92 If you are using standalone, then you'd just use networking protocol like http/tcp.. and you can send JSON that way (instead of inside dai.Buffer via Xlink)

Link92

I am not sure I am being clear enough, I am referring quite to the same type of application @martin181818 was referring to. With the only difference I am using standalone. So, the idea would be that of switching from one model to another. In order to do so I cannot think of any other different way that using dai.Buffer.

jakaskerl

Hi @Link92
Try to convert data to bytearray.

Thanks,
Jaka

Link92

@jakaskerl, @erik Thanks for your patience and support, unfortunately I cannot still see how these could be a solution to my problem. In order to stay simple, I'll refer directly to this example. As you can see, the latest three line of code are:

buf = dai.Buffer()
buf.setData(runNn)
inQ.send(buf)

Would you be so kind to show me how this example can be reproduced in standalone?

jakaskerl

Hi @Link92
If you wish to externally control the pipeline flow when in standalone mode, you can utilize a socket to send data through. Basically create a server inside a script node.

Thoughts?