Principle:Onnx Onnx Model Validation
| Knowledge Sources | |
|---|---|
| Domains | Validation, Model_Integrity |
| Last Updated | 2026-02-10 00:00 GMT |
Overview
A verification process that checks an ONNX model for structural correctness, type consistency, and compliance with the ONNX specification.
Description
Model validation ensures that an ONNX model conforms to the specification before it is saved or deployed to a runtime. The checker verifies multiple aspects of model correctness: the IR version is valid and properly set, opset imports are declared and compatible, all operator nodes reference valid operator types with correct input/output counts, tensor types match operator expectations, and there are no duplicate keys in metadata. Optionally, a full check can also run shape inference to verify type consistency throughout the graph.
This principle addresses the critical need for early error detection: catching malformed models at creation time prevents silent failures at runtime, which can be much harder to debug. The validation is implemented as a C++ engine for performance, wrapped with a Python API for convenience.
Usage
Use this principle after constructing or modifying an ONNX model, and before serialization or deployment. Validation should be applied after model creation, after version conversion, after model composition, and after any transformation that modifies the graph structure. For models larger than 2GB, path-based validation must be used.
Theoretical Basis
Model validation performs a series of structural and semantic checks:
- IR version check: Verify model.ir_version is set and within supported range
- Opset import check: Verify opset_import is consistent with IR version
- Node validation: For each node, verify:
- op_type exists in the declared opset
- Input/output count matches operator schema
- Attribute types match schema expectations
- Type consistency: (with full_check) Run shape inference and verify no conflicts
- Metadata check: No duplicate keys in metadata_props
Pseudo-code:
# Abstract validation algorithm
def validate(model):
check_ir_version(model.ir_version)
check_opset_imports(model.opset_import)
for node in model.graph.node:
schema = lookup_schema(node.op_type, opset_version)
check_inputs_outputs(node, schema)
check_attributes(node, schema)
if full_check:
inferred = shape_inference(model)
check_type_consistency(inferred)