How to beginner · 3 min read

How to export TensorFlow model to ONNX

Q: How to export TensorFlow model to ONNX

Use the tf2onnx Python package to convert a TensorFlow model to ONNX format. Load your TensorFlow model, then run tf2onnx.convert.from_keras() or tf2onnx.convert.from_saved_model() to export it as an ONNX file.

Quick answer

Use the tf2onnx Python package to convert a TensorFlow model to ONNX format. Load your TensorFlow model, then run tf2onnx.convert.from_keras() or tf2onnx.convert.from_saved_model() to export it as an ONNX file.

PREREQUISITES

Python 3.8+
pip install tensorflow
pip install tf2onnx

Setup

Install the required packages tensorflow and tf2onnx using pip.

bash

pip install tensorflow tf2onnx

Step by step

This example shows how to export a simple TensorFlow Keras model to ONNX format using tf2onnx.

python

import tensorflow as tf
import tf2onnx

# Create a simple Keras model
model = tf.keras.Sequential([
    tf.keras.layers.InputLayer(input_shape=(28, 28)),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(10, activation='softmax')
])

# Save the model in SavedModel format
saved_model_path = "./saved_model"
model.save(saved_model_path)

# Convert the Keras model to ONNX
import tf2onnx.convert

spec = (tf.TensorSpec((None, 28, 28), tf.float32, name="input"),)

model_proto, external_tensor_storage = tf2onnx.convert.from_keras(
    model,
    input_signature=spec,
    opset=17
)

# Save the ONNX model to a file
onnx_model_path = "model.onnx"
with open(onnx_model_path, "wb") as f:
    f.write(model_proto.SerializeToString())

print(f"ONNX model saved to {onnx_model_path}")

output

ONNX model saved to model.onnx

Common variations

You can also convert a TensorFlow SavedModel directory directly to ONNX using tf2onnx.convert.from_saved_model(). For example:

python

import tf2onnx

saved_model_path = "./saved_model"

model_proto, _ = tf2onnx.convert.from_saved_model(
    saved_model_path,
    opset=17
)

onnx_model_path = "model_from_saved_model.onnx"
with open(onnx_model_path, "wb") as f:
    f.write(model_proto.SerializeToString())

print(f"ONNX model saved to {onnx_model_path}")

output

ONNX model saved to model_from_saved_model.onnx

Troubleshooting

If you get errors about unsupported ops, try upgrading tf2onnx to the latest version.
Ensure your TensorFlow model uses ops supported by ONNX opset 17 or adjust the opset parameter accordingly.
For complex models, exporting from the SavedModel directory is more reliable than from Keras directly.

✅

Key Takeaways

Use tf2onnx to convert TensorFlow models to ONNX format easily.
Export from Keras models or SavedModel directories depending on your workflow.
Always specify the ONNX opset version (e.g., 17) for compatibility.
Upgrade tf2onnx if you encounter unsupported operator errors.
Saving the ONNX model requires writing the serialized protobuf to a file.

Verified 2026-04 · tf2onnx, tensorflow

Verify ↗