Pipeline goes to sleep

PepperoniPingu

Hi,

I'm using an oak-1 for periodic qr code decoding. On the oak I have a DetectionModel for detecting the qr codes. The result and the frame in full resolution get sent upstream to an dai.node.HostNode that crops the image to the detection bounding box and then decodes it with pyzbar.

But it's not supposed to run all the time since the application does other stuff to. Runs about every 20th second. So I run pipeline.start() at start and then every time I want to scan I poll queue.tryGet(). Since the whole pipeline is blocking, this means the pipeline is "frozen" when I don't poll. I do not want to shut down the pipeline every time since it's so slow to start it up again.

First time after the pipeline is started this works. But when the app has been doing something else for a while and comes back for a second scan it seems the pipeline is asleep. I get nothing when I poll. But then magically after about 4 minutes of polling it wakes up again and starts working correctly.

I've looked with top and my HostNode(5) thread shows very little activity (compared to the first run when it always works). Furthermore the XLinkInHost threads are completely dead second time around.

With DEPTHAI_LEVEL=trace on the first scan I see the data shuffling around and get the usual:

[18443010513845F500] [3.1] [2439.221] [DetectionParser(4)] [trace] DetectionParser took '12.115082' ms

Second time, I get no information about data shuffling around. I only get the following three lines repeating:

[18443010513845F500] [3.1] [1810.589] [system] [info] Temperatures - Average: 54.80C, CSS: 55.29C, MSS 54.64C, UPA: 54.21C, DSS: 55.07C
[1844301041F542F500] [1.1] [303.865] [system] [info] Cpu Usage - LeonOS 2.35%, LeonRT: 0.18%
[18443010513845F500] [3.1] [1753976327.793] [host] [trace] Log vector decoded, size: 3

Until it starts working again after the mentioned minutes of polling.

Unfortunately I don't know if it's a node in the camera or a host thread going to sleep and blocking the rest of the pipeline.

Setup:

depthai==3.0.0rc3
oak 1
raspberry pi 5 8gb

PepperoniPingu

It seems to work when I do queue.get() instead of queue.tryGet() but I don't want to do that since I have other stuff in the same thread that needs to run if there is no data in the queue.

PepperoniPingu

Finally got it working. Had to put a time.sleep(0) to yield the thread in between every queue.tryGet() that returned None . My guess is that this forces the depthai threads (HostNode, XLinkInHost etc) to run.

I still regard this as a bug in depthai though. Tested on depthai==3.0.0rc4 as well and same issue.

MatevzMorato

@PepperoniPingu thanks for posting the issue and how you solved it.

The issue likely stems in the Python interpreter hot-looping on the main thread and consequently the host nodes cannot run.

MatevzMorato

@PepperoniPingu would you perhaps be willing to share a reprodicble example?

PepperoniPingu

Sorry don't have the time to distill in to a MRE right now. The code is part of a bigger proprietary project so can't share. When I run the Depthai code outside of the project everything works fine.