Implementation:Onnx Onnx ReferenceEvaluator Run

Knowledge Sources	ONNX ONNX Reference Implementation
Domains	Model_Evaluation, Inference
Last Updated	2026-02-10 00:00 GMT

Overview

Concrete tool for executing ONNX models using the pure Python reference evaluator provided by the ONNX reference module.

Description

The run method of ReferenceEvaluator executes the ONNX model by iterating through nodes in topological order. For each node, it collects inputs from previous results, invokes the Python operator implementation, and stores the outputs. The method supports requesting specific outputs by name, returning all intermediate results for debugging (via intermediate=True), and passing linked attributes for function evaluation.

Usage

Call run after initializing a ReferenceEvaluator instance. Pass a dictionary of input names to NumPy arrays as the feed dictionary. The method returns a list of NumPy arrays corresponding to the requested output names.

Code Reference

Source Location

• Repository: onnx
• File: onnx/reference/reference_evaluator.py
• Lines: 543-610

Signature

def run(
    self,
    output_names,
    feed_inputs: dict[str, Any],
    attributes: dict[str, Any] | None = None,
    intermediate: bool = False,
) -> dict[str, Any] | list[Any]:
    """Executes the ONNX model.

    Args:
        output_names: Requested outputs by name, or None for all outputs.
        feed_inputs: Dictionary of {input_name: numpy_array}.
        attributes: Attribute values for FunctionProto evaluation.
        intermediate: If True, return all results (including intermediates)
            as a dictionary; if False, return only final outputs as a list.

    Returns:
        list[Any] of output arrays when intermediate=False,
        dict[str, Any] of all results when intermediate=True.
    """

Import

from onnx.reference import ReferenceEvaluator

I/O Contract

Inputs

Name	Type	Required	Description
output_names	list[str] or None	Yes	Output names to return (None for all graph outputs)
feed_inputs	dict[str, Any]	Yes	Input name to NumPy array mapping
attributes	dict[str, Any] or None	No	Attribute values for FunctionProto
intermediate	bool	No	Return all intermediate results (default: False)

Outputs

Name	Type	Description
return (intermediate=False)	list[Any]	List of NumPy arrays, one per requested output
return (intermediate=True)	dict[str, Any]	Dictionary of all tensor names to their values

Usage Examples

Basic Inference

import numpy as np
from onnx.reference import ReferenceEvaluator

evaluator = ReferenceEvaluator("model.onnx")

# Prepare input data
X = np.random.randn(1, 3, 224, 224).astype(np.float32)

# Run inference
outputs = evaluator.run(None, {"input": X})
print(f"Output shape: {outputs[0].shape}")

Debugging with Intermediate Results

import numpy as np
from onnx.reference import ReferenceEvaluator

evaluator = ReferenceEvaluator("model.onnx", verbose=1)

X = np.array([[1.0, -2.0, 3.0]], dtype=np.float32)
results = evaluator.run(None, {"X": X}, intermediate=True)

# Inspect all intermediate tensor values
for name, value in results.items():
    if value is not None:
        print(f"{name}: {value}")

Request Specific Outputs

import numpy as np
from onnx.reference import ReferenceEvaluator

evaluator = ReferenceEvaluator(model)

X = np.random.randn(1, 784).astype(np.float32)

# Only request specific outputs
logits = evaluator.run(["logits"], {"input": X})
print(f"Logits: {logits[0]}")

Related Pages

Implements Principle

• Principle:Onnx_Onnx_Reference_Inference

Requires Environment

• Environment:Onnx_Onnx_Python_Runtime_Environment