Introduction to ONNX

This documentation describes the ONNX concepts (Open Neural Network Exchange). It shows how it is used with examples in python and finally explains some of challenges faced when moving to ONNX in production.

• ONNX Concepts
  • Input, Output, Node, Initializer, Attributes
  • Serialization with protobuf
  • Metadata
  • List of available operators and domains
  • Supported Types
  • What is an opset version?
  • Subgraphs, tests and loops
  • Extensibility
  • Shape (and Type) Inference
  • Tools
• ONNX with Python
  • A simple example: a linear regression
  • Serialization
  • Initializer, default value
  • Attributes
  • Opset and metadata
  • Subgraph: test and loops
  • Parsing
  • Checker and Shape Inference
• Challenges
  • What is a converting library?
  • Opsets
  • Other API
  • Tricks learned from experience
• ONNX operators and function

previous

Tutorials

next

ONNX Concepts