Parallel Multiple Models App

Introduction

The parallel inferencing application constructs the parallel inferencing branches pipeline as the following graph, so that the multiple models can run in parallel in one piepline.

Main Features

• Support multiple models inference with nvinfer(TensorRT) or nvinferserver(Triton) in parallel
• Support sources selection for different models with nvstreammux and nvstreamdemux
• Support new nvstreammux. you can set "export USE_NEW_NVSTREAMMUX=yes" to use new streammux.
• Support to mux output meta from different sources and different models with gst-nvdsmetamux plugin newly introduced in DeepStream 6.1.1 or above

Prerequisites

• If you are using a deepstream docker above DeepStream 6.1.1 version, please execute /opt/nvidia/deepstream/deepstream/user_additional_install.sh to install tools such as x264enc which is used in this sample.
• DeepStream 6.1.1 or above, especially
  • nvmsgbroker if you want to enable nvmsgbroker sink, e.g. Kafka
  • nvinferserver if running model with Triton
• Cloud server, e.g. Kafka server (version >= kafka_2.12-3.2.0), if you want to enable broker sink

Steps To Run

The sample should be downloaded and built with root permission.

1. Download

   apt install git-lfs
   git lfs install --skip-repo
   git clone https://github.com/NVIDIA-AI-IOT/deepstream_parallel_inference_app.git

   If git-lfs download fails for bodypose2d and YoloV4 models, get them from Google Drive link

2. Generate Inference Engines

   Below instructions are only needed on Jetson (Jetpack 5.0.2 or above)

   apt-get install -y libjson-glib-dev libgstrtspserver-1.0-dev
   /opt/nvidia/deepstream/deepstream/samples/triton_backend_setup.sh
   ## Only DeepStream 6.1.1 GA need to copy the metamux plugin library. Skip this copy command if DeepStream version is above 6.1.1 GA
   cp tritonclient/sample/gst-plugins/gst-nvdsmetamux/libnvdsgst_metamux.so /opt/nvidia/deepstream/deepstream/lib/gst-plugins/libnvdsgst_metamux.so
   ## set power model and boost CPU/GPU/EMC clocks
   nvpmodel -m 0 && jetson_clocks

   Below instructions are needed for both Jetson and dGPU (DeepStream Triton docker - 6.1.1-triton or above)

   cd tritonserver/
   ./build_engine.sh

3. Build and Run

   cd tritonclient/sample/
   source build.sh
   ./apps/deepstream-parallel-infer/deepstream-parallel-infer -c configs/apps/bodypose_yolo_lpr/source4_1080p_dec_parallel_infer.yml

Directory

Application Configuration Semantics

The parallel inferencing app uses the YAML configuration file to config GIEs, sources, and other features of the pipeline. The basic group semantics is the same as deepstream-app.

Please refer to deepstream-app Configuration Groups part for the semantics of corresponding groups.

There are additional new groups introduced by the parallel inferencing app which enable the app to select sources for different inferencing branches and to select output metadata for different inferencing GIEs:

Branch Group

The branch group specifies the sources to be infered by the specific inferencing branch. The selected sources are identified by the source IDs list. The inferencing branch is identified by the first PGIE unique-id in this branch. For example:

branch0:
  key1: value1
  key2: value2

The branch group properties are:

Metamux Group

The metamux group specifies the configuration file of gst-dsmetamux plugin. For example:

meta-mux:
  key1: value1
  key2: value2

The metamux group properties are:

The gst-dsmetamux configuration details are introduced in gst-dsmetamux plugin README.

Sample Models

The sample application uses the following models as samples.

Sample Configuration Files

The application will create new inferencing branch for the designated primary GIE. The secondary GIEs should identify the primary GIE on which they work by setting "operate-on-gie-id" in nvinfer or nvinfereserver configuration file.

To make every inferencing branch unique and identifiable, the "unique-id" for every GIE should be different and unique. The gst-dsmetamux module will rely on the "unique-id" to identify the metadata comes from which model.

There are five sample configurations in current project for reference.

• The sample configuration for the open source YoloV4, bodypose2d and TAO car license plate identification models with nvinferserver.

  • Configuration folder

  tritonclient/sample/configs/apps/bodypose_yolo_lpr

  "source4_1080p_dec_parallel_infer.yml" is the application configuration file. The other configuration files are for different modules in the pipeline, the application configuration file uses these files to configure different modules.

  • Pipeline Graph:

  

  • App Command:

  ./apps/deepstream-parallel-infer/deepstream-parallel-infer -c configs/apps/bodypose_yolo_lpr/source4_1080p_dec_parallel_infer.yml

• The sample configuration for the TAO vehicle classifications, carlicense plate identification and peopleNet models with nvinferserver.

  • Configuration folder

  tritonclient/sample/configs/apps/vehicle_lpr_analytic

  "source4_1080p_dec_parallel_infer.yml" is the application configuration file. The other configuration files are for different modules in the pipeline, the application configuration file uses these files to configure different modules.

  • Pipeline Graph:

  

  • App Command:

  ./apps/deepstream-parallel-infer/deepstream-parallel-infer -c configs/apps/vehicle_lpr_analytic/source4_1080p_dec_parallel_infer.yml

• The sample configuration for the TAO vehicle classifications, carlicense plate identification and peopleNet models with nvinferserver and nvinfer.

  • Configuration folder

  tritonclient/sample/configs/apps/vehicle0_lpr_analytic

  "source4_1080p_dec_parallel_infer.yml" is the application configuration file. The other configuration files are for different modules in the pipeline, the application configuration file uses these files to configure different modules. The vehicle branch uses nvinfer, the car plate and the peoplenet branches use nvinferserver.

  • Pipeline Graph:

  

  • App Command:

  ./apps/deepstream-parallel-infer/deepstream-parallel-infer -c configs/apps/vehicle0_lpr_analytic/source4_1080p_dec_parallel_infer.yml

• The sample configuration for the open source YoloV4, bodypose2d with nvinferserver and nvinfer.

  • Configuration folder

  tritonclient/sample/configs/apps/bodypose_yolo/

  "source4_1080p_dec_parallel_infer.yml" is the application configuration file. The other configuration files are for different modules in the pipeline, the application configuration file uses these files to configure different modules. The bodypose branch uses nvinfer, the yolov4 branch use nvinferserver. The output streams is tiled.

  • Pipeline Graph:

  

  • App Command:

  ./apps/deepstream-parallel-infer/deepstream-parallel-infer -c configs/apps/bodypose_yolo/source4_1080p_dec_parallel_infer.yml

• The sample configuration for the open source YoloV4, bodypose2d with nvinferserver and nvinfer.

  • Configuration folder

  tritonclient/sample/configs/apps/bodypose_yolo_win1/

  "source4_1080p_dec_parallel_infer.yml" is the application configuration file. The other configuration files are for different modules in the pipeline, the application configuration file uses these files to configure different modules. The bodypose branch uses nvinfer, the yolov4 branch use nvinferserver. The output streams is source 2.

  • Pipeline Graph:

  

  • App Command:

  ./apps/deepstream-parallel-infer/deepstream-parallel-infer -c configs/apps/bodypose_yolo_win1/source4_1080p_dec_parallel_infer.yml