Parallel Inference Demo using Python

[damonzzz]

Please provide complete information as applicable to your setup.

• Hardware Platform GPU
• DeepStream Version 6.1
• Issue Type questions

Hi, I’m trying to write a parallel inference demo using python, I found some reference repo on github:

• deepstream parallel inference app
• deepstream demux demo

the first one was written in C, so i build a pipeline like it does expect I didn’t use metamux which only provides a .so file in this repo. Instead, I use fakesink for both branch.

the result is confusing, the first branch is normal, its fps is alright and can also be displayed if change to nveglglessink. However, the second branch always has 0 fps.

**PERF:  {'stream0': 25.09} 


**PERF:  {'stream0': 0.0} 


**PERF:  {'stream0': 24.99} 


**PERF:  {'stream0': 0.0} 


**PERF:  {'stream0': 24.98} 


**PERF:  {'stream0': 0.0} 


**PERF:  {'stream0': 25.18} 

then I change the sink property asycn which was set to false before. If asycn was true, the whole pipeline just stuck at first frame.

So is the plugin metamux necessary if I don’t need to mux the outputs. Why only the first branch of tee has fps.

Thanks in advance.

[Fiona.Chen]

Do you mean you added nvstreamdemux to replace nvmultistreamtiler in the * deepstream parallel inference app sample to output seperated streams?

[damonzzz]

thanks for your quick reply.
I didn’t change the front part of origin pipeline in deepstream parallel inference app. I just rewrite them in Python. the pipeline look like this:

and as I continue to test, I remove the second streammux and streamdemux, and it works. I guess something wrong with my second branch’s streammux and streamdemux.

[Fiona.Chen]

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

I don’t think your pipeline is our deepstream parallel inference app

We don’t know how did you implement the pipeline. Please debug with our deepstream parallel inference app to understand the right settings.

MetaMux is not to merge the output, it only collects all output together in unified NvDsMetaData. You can filter the output according to your requirement.