Extension NavigatorFILE EXTENSION LIBRARY
日本語
Home›AI & Machine Learning›.onnx
THE FILE FIELD GUIDE
.onnx

ONNX AI Model

AI & Machine Learning

ONNX exchanges machine-learning models. Inference depends on input names/shapes/types, operator versions and runtime support. Weights may be external, so the model file alone may be incomplete.

What it is used for
Exchange of machine-learning graphs and weights
What the file contains
Binary data
All sections, questions and sources

Content reviewed： · Editor：Kai

Opening steps

Choose your environment or purpose to see its steps.

Identify the purpose and distribution

Inspect graph, external weights and inputs.

  1. Read the publisher instructions and license; record the producer, version and purpose.
  2. Check the distribution for external tensor data accompanying the ONNX file.
  3. Retain the original and assemble configuration and companions in a working copy.
Inspect with the matching environment

Successful loading and correct inference are separate checks.

  1. Inspect IR version, opsets and inputs/outputs with compatible ONNX tooling.
  2. Preserve relative external-data paths and check that validation can read companions.
  3. Inspect session input names, types, shapes and available execution providers.
Test with a known input

Match the input contract and compare with a reference.

  1. Match color order, normalization, axis layout and batch dimension to the source contract.
  2. Compare named outputs and values against the source model with the same input.
  3. After checking CPU execution, inspect outputs and logs with the intended GPU provider.

Menu labels and support vary by app, OS and version. Preserve the original before making changes.

Troubleshooting by symptom

  • Input mismatch: inspect names, types, layout/batch dimensions and preprocessing.
  • Unsupported operation: check opset, custom operators and runtime compatibility.
  • Missing weights: inspect external-data names and paths relative to the model.
  • Shape mismatch: inspect axis layout and batching rather than merely changing dimensions.
  • GPU selection is slow: inspect actual provider use and CPU execution of unsupported operations.
  • Moving the model breaks loading: copy external weights with their relative structure.

Training model → ONNX

  1. Use the framework’s supported exporter with suitable sample inputs and target opset.
  2. Compare outputs on matching inputs and inspect precision, dynamic shapes and external data.
  3. Compare output shapes, values and timing with the same input before/after export; retain definitions and export settings.

Renaming is not conversion; export support is not universal.

Building the correct input

Treat names, types, dimensions and axis order as one contract; images also require matching value ranges and color order.

Numerical output alone is insufficient; align label order and postprocessing and compare known cases.

Source for these steps：ONNX Project · ONNX Concepts / ONNX Project · External Data / ONNX Runtime Project · Python API overview

Single-file models and external data

ONNX may keep large tensors outside the graph file; transferring only the graph can omit weights.

Include a manifest, runtime information and reference output; check opsets and custom operators at the destination.

Source for these steps：ONNX Project · ONNX Concepts / ONNX Project · External Data / ONNX Runtime Project · Python API overview

Related-format differences

Related-format differences
FormatMain roleWhat to check
ONNXGraph/inference interchangeInspect external data and opsets
PTPyTorch artifactIdentify architecture and save method
ENGINEOptimized TensorRT artifactDepends on hardware and build settings

Handling and security

Review origin, license and memory needs; avoid unfamiliar custom-operator libraries and use updated runtimes.

Questions and answers

Is the ONNX alone sufficient?

External weights, preprocessing and a compatible runtime may be required.

What is opset?

The operator-specification version; check runtime support.

Can SAFETENSORS be renamed ONNX?

No; the structures, including model graphs, differ.

Does dynamic input accept any size?

Dynamic dimensions still have operator/export constraints; test the intended sizes.

Does validation prove equal accuracy?

Structural validity is separate from accuracy; compare predictions with identical preprocessing.

Format facts: name, steward and MIME type
Formal name
Open Neural Network Exchange
Steward / specification
ONNX project
MIME type
Depends on the producer or server; an extension alone does not establish a MIME type.

Primary and official sources

  • Project documentationONNX Concepts

    Explains model graphs, operators and inputs/outputs.

    ONNX Project · Link checked：
  • Project documentationExternal Data

    Documents externally stored tensor weights.

    ONNX Project · Link checked：
  • Official documentationPython API overview

    Documents session inputs and execution providers.

    ONNX Runtime Project · Link checked：