ONNX
An open model-exchange format for representing machine-learning graphs, operators and tensor initializers across training and inference environments.
Format vs Runtime
ONNX represents model structure and parameters in a common format. Exporting a model to ONNX does not by itself decide which CPU/GPU kernels will run or how fast inference will be.
Execution behavior belongs to the Inference Engine.
Operator Sets
Graph operations target an operator-set version. A source-framework operation that is unsupported by the target opset or runtime can create conversion or execution problems.
Dynamic shapes, custom operators and numerical precision are additional portability boundaries.
Validation
A successful export does not prove semantic equivalence. Representative inputs should be compared between source and target runtimes within an explicitly chosen tolerance.
Related Concepts
Direct source: The primary paper or official specification for ONNX is linked here for verification.