Implementation:Onnx Onnx Compose Expand Out Dim

Knowledge Sources	ONNX ONNX Compose API
Domains	Model_Composition, Shape_Manipulation
Last Updated	2026-02-10 00:00 GMT

Overview

Concrete tool for inserting extra dimensions into ONNX model outputs provided by the ONNX compose module.

Description

The expand_out_dim function inserts an Unsqueeze node for each output in the model, adding a dimension of extent 1 at the specified index. It modifies the graph by renaming original output edges, appending Constant and Unsqueeze nodes, and updating the output ValueInfoProto with the expanded shape. The function can operate in-place or return a copy.

Usage

Import this function when a model's outputs need an additional dimension for compatibility with a downstream model during composition. Commonly used to add a batch dimension (dim_idx=0) before merging models.

Code Reference

Source Location

• Repository: onnx
• File: onnx/compose.py
• Lines: 718-751

Signature

def expand_out_dim(
    model: ModelProto,
    dim_idx: int,
    inplace: bool | None = False,
) -> ModelProto:
    """Inserts an extra dimension with extent 1 to each output.

    Args:
        model: Model whose outputs to expand.
        dim_idx: Index of the dimension to insert.
            Negative values count from the back.
        inplace: If True, mutate directly; otherwise copy first.

    Returns:
        ModelProto with expanded output dimensions.
    """

Import

from onnx import compose

I/O Contract

Inputs

Name	Type	Required	Description
model	ModelProto	Yes	Model whose outputs to expand
dim_idx	int	Yes	Dimension index to insert (negative for counting from back)
inplace	bool	No	Mutate directly or copy (default: False)

Outputs

Name	Type	Description
return	ModelProto	Model with Unsqueeze nodes appended, outputs have an extra dimension

Usage Examples

Add Batch Dimension

import onnx
from onnx import compose

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

# Add batch dimension at index 0
expanded = compose.expand_out_dim(model, dim_idx=0)

# Original output [10] becomes [1, 10]
for out in expanded.graph.output:
    dims = [d.dim_value for d in out.type.tensor_type.shape.dim]
    print(f"{out.name}: {dims}")

Related Pages

Implements Principle

• Principle:Onnx_Onnx_Output_Dimension_Expansion

Requires Environment

• Environment:Onnx_Onnx_Python_Runtime_Environment