Inquiry about deepstream parallel inference sample app

[clefable]

Deepstream Version 7.0

Hi I was dabbling with the deepstream_parallel_inference_app while referring to this link: deepstream_parallel_inference_app/README.md at master · NVIDIA-AI-IOT/deepstream_parallel_inference_app · GitHub

More specifically, I was trying out the bodypose_yolo_win1 example: deepstream_parallel_inference_app/tritonclient/sample/configs/apps/bodypose_yolo_win1/source4_1080p_dec_parallel_infer.yml at master · NVIDIA-AI-IOT/deepstream_parallel_inference_app · GitHub

I wanted to show all 4 sources instead, so I commented out show source: 2 in line 48 under [tiled-display].
I also wanted inference to be performed on all 4 sources, so under [branch0] and [branch1], I changed src-ids: to src-ids: 0;1;2;3 for both (lines 209 and 246.

However, this is what I got for my output:

It seems that the object detection model is not inferring on sources 2 and 3, while the bodypose is not inferring on sources 0 and 3.

Regarding this, I have a few inquiries:

1. Am I changing the correct configs? I’m under the impression that src-ids under the [branch] group is what I’m supposed to change.

2. Is this considered 1 or 4 sources? deepstream_parallel_inference_app/tritonclient/sample/configs/apps/bodypose_yolo_win1/sources_4.csv at master · NVIDIA-AI-IOT/deepstream_parallel_inference_app · GitHub

3. I noticed that this sample app does not send print information regarding FPS in the terminal. I referred to the link below to print out the statistics, but it seems like the FPS is capped at 30 (source FPS), although I can’t be completely sure. I was wondering whether there is a more straightforward way to produce the performance statistics.

My goal for now is to output a single source with the annotations of both models. I know I can simply do show-source: 1 in the [tiled-display] branch or something for that, but wouldn’t that mean I’m still running inference on 4 sources?

[junshengy]

@clefable

Am I changing the correct configs? I’m under the impression that src-ids under the [branch] group is what I’m supposed to change.

Yes, this modification is correct.

@clefable

Is this considered 1 or 4 sources? deepstream_parallel_inference_app/tritonclient/sample/configs/apps/bodypose_yolo_win1/sources_4.csv at master · NVIDIA-AI-IOT/deepstream_parallel_inference_app · GitHub

4, num-sources represents the number of times the source appears

@clefable

I noticed that this sample app does not send print information regarding FPS in the terminal. I referred to the link below to print out the statistics, but it seems like the FPS is capped at 30 (source FPS), although I can’t be completely sure. I was wondering whether there is a more straightforward way to produce the performance statistics.

deepstream-parallel-infer currently requires you to get the FPS by yourself. If you want to test the highest performance, you can use fakesink

[junshengy]

@clefable

It seems that the object detection model is not inferring on sources 2 and 3, while the bodypose is not inferring on sources 0 and 3.

Regarding this issue, you can refer to this configuration file

tritonclient/sample/configs/metamux/config_metamux0.txt

metamux is responsible for merging meta data from different branches together

[yingliu]

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