TAO YoloV4 .etlt model with Triton server

[rsicak]

Hi, is there any guide how to implement Yolo v4 TAO model into Triton inference server? I have trained Yolo v4 custom data model via TAO toolkit and looking for an guide how to implement this model with Triton inference server. Thank you.

[imSrbh]

@rsicak Any update on How you made it working??

[imSrbh]

@sujitbiswas @morganh-nv

I have yolov4 .etlt model and generated trt.engine from nvinfer (deepstream-app) and I have generated libnvds_infercustomparser_tao.so

I wish to use same in triton inference server. model-repository:

Then Written a deepstream-app for the same with nvinferserver plugin.

For one model with the same configuration, I am able to do inference. But Unable to get any meta info & obj_count for other model with distinct classes.

Thanks...

Here is my config_infer.txt

infer_config {
  unique_id: 1
  gpu_ids: [0]
  max_batch_size: 16
  
  backend {
    inputs: [ {
      name: "Input"
    }]
    outputs: [
      {name: "BatchedNMS"},
      {name: "BatchedNMS_1"},
      {name: "BatchedNMS_2"},
      {name: "BatchedNMS_3"}
    ]
    triton {
      model_name: "Helmet"
      version: 1
      grpc {
        url: "172.17.0.2:8001"		
      }
    }
  }

  preprocess {
    network_format: MEDIA_FORMAT_NONE
    tensor_order: TENSOR_ORDER_NONE
    tensor_name: "Input"
    maintain_aspect_ratio: 0
    frame_scaling_hw: FRAME_SCALING_HW_DEFAULT
    frame_scaling_filter: 1
    normalize {
    scale_factor: 1.0
    channel_offsets: [0, 0, 0]
    }
  }

  postprocess {
    labelfile_path: "../../model_repository/Helmet_model/labels.txt"
    detection {
      num_detected_classes: 2
      custom_parse_bbox_func:"NvDsInferParseCustomBatchedNMSTLT"
      per_class_params {
          key: 0
          value { pre_threshold: 0.4 }
        }
      nms {
        confidence_threshold:0.2
        topk:20
        iou_threshold:0.5
      }
                             
    }
  }

  custom_lib {
    path:"../../.../customLib/libnvds_infercustomparser_tao.so"
  }
}

[Wesley-E]

Hi @imSrbh I believe that you have to write a client for your model to get the desired outputs that you are looking for and I may be able to assist you with this if necessary. Let me know if you still need help.

[monjha]

Hi @imSrbh

could you share your analytics config file where you are defining object count for a particular class id? you probably want to update the class-id if you are able to see counts in one model and no counts in another.

[morganh-nv]

You can firstly run the default steps in this github. Then, to run your own .etlt model in triton server, just need to replace your own .etlt model with the original one, and also set correct input shapes in config file(https://github.com/NVIDIA-AI-IOT/tao-toolkit-triton-apps/blob/main/model_repository/yolov3_tao/config.pbtxt#L9). The triton server will generate model.plan with your own .etlt model. For example, for yolov3, https://github.com/NVIDIA-AI-IOT/tao-toolkit-triton-apps/blob/main/scripts/download_and_convert.sh#L51

[h9945394143]

@morganh-nv @monjha could you help us with post-processor parser

[Wesley-E]

@morganh-nv @monjha could you help us with post-processor parser

I could help you with that

[Wesley-E]

@morganh-nv @monjha could you help us with post-processor parser

Do you have a branch that you're currently working on that you could share?

[morganh-nv]

The post-processing is available at https://github.com/NVIDIA-AI-IOT/tao-toolkit-triton-apps/blob/main/tao_triton/python/postprocessing/yolov3_postprocessor.py