Implementation:Onnx Onnx Shape Inference Infer Shapes

Knowledge Sources	ONNX ONNX Shape Inference API
Domains	Type_System, Static_Analysis
Last Updated	2026-02-10 00:00 GMT

Overview

Concrete tool for propagating shape and type information through ONNX models, wrapping a C++ shape inference engine.

Description

The infer_shapes function applies shape inference to a ModelProto, returning a new ModelProto with inferred type information populated in the graph's value_info field. It is a Python wrapper around the C++ shape inference engine (onnx_cpp2py_export.shape_inference). The function serializes the model to bytes, passes it to the C++ engine, and deserializes the result. For models exceeding 2GB, the companion infer_shapes_path function operates directly on file paths.

Usage

Import infer_shapes when you need to enrich a model with shape information for runtime optimization or validation. Use infer_shapes_path for models with external data that exceed 2GB. The function returns a new model (it does not modify in-place).

Code Reference

Source Location

Repository: onnx
File: onnx/shape_inference.py
Lines: 32-70 (infer_shapes), 73-108 (infer_shapes_path)

Signature

def infer_shapes(
    model: ModelProto | bytes,
    check_type: bool = False,
    strict_mode: bool = False,
    data_prop: bool = False,
) -> ModelProto:
    """Apply shape inference to the provided ModelProto.

    Args:
        model: ModelProto or serialized bytes.
        check_type: Check type equality for input/output.
        strict_mode: Throw errors instead of stopping on failure.
        data_prop: Enable data propagation for limited operators.

    Returns:
        New ModelProto with inferred shape information in value_info.
    """

def infer_shapes_path(
    model_path: str | os.PathLike,
    output_path: str | os.PathLike = "",
    check_type: bool = False,
    strict_mode: bool = False,
    data_prop: bool = False,
) -> None:
    """Apply shape inference for models >2GB (path-based).

    Args:
        model_path: Input model file path.
        output_path: Output path (default: overwrite input).
        check_type: Check type equality.
        strict_mode: Throw errors on failure.
        data_prop: Enable data propagation.
    """

Import

from onnx import shape_inference

I/O Contract

Inputs

Name	Type	Required	Description
model	ModelProto or bytes	Yes	Model to infer shapes for
check_type	bool	No	Check type equality for inputs/outputs (default: False)
strict_mode	bool	No	Throw errors instead of stopping (default: False)
data_prop	bool	No	Enable data propagation (default: False)

Outputs

Name	Type	Description
return	ModelProto	New model with value_info populated with inferred shapes/types

Usage Examples

Basic Shape Inference

import onnx
from onnx import shape_inference

model = onnx.load_model("model.onnx")

# Apply shape inference
inferred_model = shape_inference.infer_shapes(model)

# Check inferred value_info
for vi in inferred_model.graph.value_info:
    print(f"{vi.name}: {vi.type}")

Path-based for Large Models

from onnx import shape_inference

# For models >2GB with external data
shape_inference.infer_shapes_path(
    "large_model.onnx",
    output_path="large_model_inferred.onnx",
    strict_mode=True,
)

Related Pages

Implements Principle

Principle:Onnx_Onnx_Shape_Inference

Requires Environment

Uses Heuristic

Heuristic:Onnx_Onnx_Shape_Inference_Limitations

Page Connections

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

Principle

Implementation

Heuristic

Environment