Implementation:Kornia Kornia ONNXSequential

Knowledge Sources	Kornia
Domains	ONNX, Deployment, Pipeline_Design
Last Updated	2026-02-09 15:00 GMT

Overview

Concrete tool for chaining ONNX models into a single inference pipeline provided by Kornia.

Description

ONNXSequential accepts ONNX models (as ModelProto objects, file paths, or HuggingFace Hub identifiers), combines their graphs with io_maps routing, and creates a single ONNX Runtime InferenceSession. It supports:

Execution provider selection — CUDA, CPU, and others.
Session option configuration — thread count, optimization level, etc.
Automatic IR/opset version conversion — harmonize models with different ONNX versions.

The combined graph can be exported as a single ONNX file.

Usage

Initialize with models, io_maps, and providers. Call the instance with numpy arrays for inference.

Code Reference

Repository	https://github.com/kornia/kornia
File	`kornia/onnx/sequential.py`
Lines	L29–115
Signature	str, providers: Optional[list[str]] = None, session_options: Optional[ort.SessionOptions] = None, io_maps: Optional[list[tuple[str, str]]] = None, cache_dir: Optional[str] = None, auto_ir_version_conversion: bool = False, target_ir_version: Optional[int] = None, target_opset_version: Optional[int] = None)
Import	`from kornia.onnx import ONNXSequential`

I/O Contract

Inputs

Parameter	Type	Required	Description
`*args`	str	Yes	ONNX models or identifiers (paths, URLs, HF Hub)
`providers`	`list[str]`	No	Execution providers (e.g., `["CUDAExecutionProvider"]`)
`io_maps`	`list[tuple[str, str]]`	No	I/O routing between consecutive models
`auto_ir_version_conversion`	`bool`	No	Auto-convert IR versions across models
`session_options`	`ort.SessionOptions`	No	ONNX Runtime session configuration
`cache_dir`	`str`	No	Directory for caching downloaded models
`target_ir_version`	`int`	No	Target IR version for conversion
`target_opset_version`	`int`	No	Target opset version for conversion

Outputs

Configured ONNXSequential instance with an active inference session.

Usage Examples

Creating pipeline from HuggingFace Hub models

from kornia.onnx import ONNXSequential

pipeline = ONNXSequential(
    "hf://operators/resize",
    "hf://operators/normalize",
    "hf://operators/classifier",
    io_maps=[("resized", "input"), ("normalized", "images")],
)

Pipeline with CUDA provider

from kornia.onnx import ONNXSequential

pipeline = ONNXSequential(
    "hf://operators/resize",
    "hf://operators/classifier",
    providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
    io_maps=[("resized", "input_image")],
)

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment