How to beginner · 3 min read

How to verify ONNX model export

Q: How to verify ONNX model export

To verify an ONNX model export, load the model using onnx.load() and check its structure with onnx.checker.check_model(). Then run inference using onnxruntime.InferenceSession to confirm the model produces expected outputs.

Quick answer

To verify an ONNX model export, load the model using onnx.load() and check its structure with onnx.checker.check_model(). Then run inference using onnxruntime.InferenceSession to confirm the model produces expected outputs.

PREREQUISITES

Python 3.8+
pip install onnx onnxruntime numpy

Setup

Install the required packages onnx, onnxruntime, and numpy using pip to load, check, and run inference on ONNX models.

bash

pip install onnx onnxruntime numpy

Step by step verification

This example loads an ONNX model file, verifies its structure, and runs a sample inference to validate the export.

python

import onnx
import onnxruntime as ort
import numpy as np
import os

# Path to your ONNX model file
model_path = "model.onnx"

# Load the ONNX model
model = onnx.load(model_path)

# Check the model's structure and validity
onnx.checker.check_model(model)
print("ONNX model structure is valid.")

# Create an inference session
session = ort.InferenceSession(model_path)

# Prepare dummy input data matching the model's input shape
input_name = session.get_inputs()[0].name
input_shape = session.get_inputs()[0].shape
input_type = session.get_inputs()[0].type

# Replace dynamic dimensions (None or 'None') with 1 for dummy input
input_shape = [dim if isinstance(dim, int) else 1 for dim in input_shape]

# Create a dummy input array of the correct shape and type
dummy_input = np.random.rand(*input_shape).astype(np.float32)

# Run inference
outputs = session.run(None, {input_name: dummy_input})

print(f"Inference output shapes: {[output.shape for output in outputs]}")

output

ONNX model structure is valid.
Inference output shapes: [(1, 1000)]

Common variations

Use onnxruntime.InferenceSession with different execution providers like CUDA for GPU acceleration.
Validate multiple inputs by preparing a dictionary with all input names and dummy data.
Use onnxruntime.SessionOptions to customize session behavior.

python

import onnxruntime as ort

# Use CUDA execution provider if available
session = ort.InferenceSession(model_path, providers=["CUDAExecutionProvider", "CPUExecutionProvider"])

# Prepare inputs for multiple inputs example
inputs = {}
for input_meta in session.get_inputs():
    shape = [dim if isinstance(dim, int) else 1 for dim in input_meta.shape]
    inputs[input_meta.name] = np.random.rand(*shape).astype(np.float32)

outputs = session.run(None, inputs)
print(f"Output shapes with multiple inputs: {[o.shape for o in outputs]}")

output

Output shapes with multiple inputs: [(1, 1000)]

Troubleshooting

If onnx.checker.check_model() raises errors, the model export is invalid or corrupted.
If inference fails, verify input shapes and types match the model's expected inputs.
Use onnx.helper.printable_graph(model.graph) to inspect the model graph for debugging.

python

import onnx

# Print the model graph for inspection
print(onnx.helper.printable_graph(model.graph))

output

graph ( ... ) { ... }  # Detailed graph structure printed

✅

Key Takeaways

Always use onnx.checker.check_model() to validate ONNX model structure after export.
Run inference with onnxruntime.InferenceSession and dummy inputs to verify runtime correctness.
Inspect model inputs and outputs carefully to prepare matching dummy data for testing.
Use execution providers like CUDA in onnxruntime for accelerated inference if available.
Print the ONNX graph to debug structural issues when verification or inference fails.

Verified 2026-04 · onnx, onnxruntime

Verify ↗