Deepstream Application freezes the system with YoloV3 int8

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Hardware Platform: GeForce GTX 1080 Ti/PCIe/SSE2
DeepStream Version: 5.0
JetPack Version: 4.3
TensorRT Version: Not sure

Hi, I’m facing a random system freeze after around 5-10 mins of running the deepstream app continuously with multiple rtsp streams. The problem only occurs while running YoloV3 pretrained network with int8 calibration. We’ve checking running the same configuration with fp32 configuration and it doesn’t freeze. We confirmed that both the system and GPU memory stayed at low values so that definitely isn’t the case. I faced a similar issue on Jetson Nano in which assigning a value out of image bounds to the RectParams in ObjMeta caused a system freeze. I’ll post that as an issue in another thread but I believe it’s relevant. I am thinking that might be the case here that somehow the yolov3 custom bbox parser is assigning out of bound values in int8 calibration case but doesn’t do so in fp32 case.
I am attaching the configuration files we’ve used. Note that the actual uri values are filtered in the config.
config.txt (2.7 KB)
pgie_config.txt (3.7 KB)

3

HI,
You also see the issue on NANO, right? if yes, can you share the config you used for NANO?

4

Hi, I have not tested this same scenario on nano since yolov3 runs very slow on nano to check it out anyway. The issue I faced on Nano was posted here: Deepstream freezes the system. I believe they are linked but not too sure about it.

5

ok, got it. and on dGPU, can you repro the issue with output encoding to file?

6

Hi @Amycao,

Sorry for a late reply. By output encoding to file, you mean changing the sink to filesink?

7

Yes.

8

Hi saifullah3396,

Any result can be shared? Is this still an issue to support?

Thanks

9

Hi,

I no longer have access to the specified GPU ( GeForce GTX 1080 Ti/PCIe/SSE2 ), so I cannot provide the results.

Thanks

11

I’m closing this topic due to can’t get the result to move this issue forward.
If need further support, please open a new one.
Thanks