models

The models list defines the AI inference pipeline. It can contain detectors, pose estimators, segmentation models, classifiers, Triton models, and an optional tracker. Entries are processed in the order they appear, which controls the inference chain.

models:
  - name: "my-detector"
    type: "detector"
    architecture: "yolo11"
    onnx_file: /path/to/model.onnx
    label_file: /path/to/labels.txt

Common fields

These fields apply to all model types except tracker.

Field	Type	Required	Default	Description
name	string	No	auto-generated (e.g., detector_0)	Unique name for this model. Used by operate_on for chaining.
type	string	Yes	—	Model type — see model types
onnx_file	string	Yes (non-Triton)	—	Path to ONNX model file
label_file	string	Yes	—	Path to label file (one class name per line)
batch_size	integer	No	next power of 2 ≥ stream count	Inference batch size (always rounded up to a power of 2)
precision	string	No	"fp16"	Inference precision: "fp32" or "fp16"

tip

num_classes is determined automatically by counting non-empty lines in label_file. The DeepStream nvinfer config file is auto-generated — you only need to provide the ONNX model and labels.

Model types

type value	Description
detector	Object detection with bounding boxes
pose	Human keypoint estimation (17 COCO joints)
segmentation	Instance segmentation with mask overlay
classifier	Secondary classification on detected objects
triton_detector	Object detection via Triton Inference Server
triton_classifier	Classification via Triton Inference Server
tracker	Object tracking

See Model types for detailed usage of each.

Model chaining rules

• At least one primary model (without operate_on) is required in every pipeline.
• operate_on must reference a model defined before it in the list.
• Classifiers must always have operate_on.
• operate_on_classes indices must be valid class IDs of the parent model.

Supported chains:

Chain	Description
detector → classifier	Detect objects, then classify each detection
detector → detector	Detect objects, then run a secondary detector on each crop
triton_detector → triton_classifier	Same pattern using Triton models