Triton detector

For models served by NVIDIA Triton Inference Server. Uses gRPC to communicate with the Triton server.

models:
  - name: "triton-detector"
    type: "triton_detector"
    triton:
      model_name: "my_model"
      server_url: "localhost:8001"
    label_file: /path/to/labels.txt

note

Triton models do not use onnx_file or architecture. The model definition lives on the Triton server.

Fields

Field	Type	Required	Default	Description
triton	object	Yes	—	Triton server configuration (see below)
label_file	string	Yes	—	Path to label file
batch_size	integer	No	next power of 2 ≥ stream count	Inference batch size
operate_on	string	No	—	Name of parent model (for secondary detection)
operate_on_classes	list of int	No	all parent classes	Parent class IDs to process

Triton configuration

Field	Type	Required	Default	Description
model_name	string	Yes	—	Name of the model on the Triton server
server_url	string	Yes	—	Triton gRPC endpoint (e.g., "localhost:8001")
model_version	integer	No	latest	Specific model version to use
preprocess	object	No	see defaults	Preprocessing overrides

Triton preprocessing

Field	Type	Default	Description
scale_factor	float	0.00392156862 (1/255)	Input normalization scale
color_format	string	"RGB"	Input color format: "RGB", "BGR", or "GRAY"
offsets	list of float	none	Channel mean offsets (e.g., [0.485, 0.456, 0.406])
maintain_aspect_ratio	integer	0	1 to maintain aspect ratio during resize
symmetric_padding	integer	0	1 to use symmetric padding

Example: Triton detector with custom preprocessing

models:
  - name: "triton-detector"
    type: "triton_detector"
    triton:
      model_name: "yolov8"
      server_url: "localhost:8001"
      model_version: 1
      preprocess:
        scale_factor: 0.00392156862
        color_format: "RGB"
        maintain_aspect_ratio: 1
        symmetric_padding: 1
    label_file: /path/to/labels.txt
    batch_size: 4