Link Search Menu Expand Document

Getting Started Converting TensorFlow to ONNX

TensorFlow models (including keras and TFLite models) can be converted to ONNX using the tf2onnx tool.

Full code for this tutorial is available here.


First install tf2onnx in a python environment that already has TensorFlow installed.

pip install tf2onnx (stable)


pip install git+ (latest from GitHub)

Converting a Model

Keras models and tf functions

Keras models and tf functions and can be converted directly within python:

import tensorflow as tf
import tf2onnx
import onnx

model = tf.keras.Sequential()
model.add(tf.keras.layers.Dense(4, activation="relu"))

input_signature = [tf.TensorSpec([3, 3], tf.float32, name='x')]
# Use from_function for tf functions
onnx_model, _ = tf2onnx.convert.from_keras(model, input_signature, opset=13), "dst/path/model.onnx")

See the Python API Reference for full documentation.


Convert a TensorFlow saved model with the command:

python -m tf2onnx.convert --saved-model path/to/savedmodel --output dst/path/model.onnx --opset 13

path/to/savedmodel should be the path to the directory containing saved_model.pb

See the CLI Reference for full documentation.


tf2onnx has support for converting tflite models.

python -m tf2onnx.convert --tflite path/to/model.tflite --output dst/path/model.onnx --opset 13

NOTE: Opset number

Some TensorFlow ops will fail to convert if the ONNX opset used is too low. Use the largest opset compatible with your application. For full conversion instructions, please refer to the tf2onnx README.

Verifying a Converted Model

Install onnxruntime with:

pip install onnxruntime

Test your model in python using the template below:

import onnxruntime as ort
import numpy as np

# Change shapes and types to match model
input1 = np.zeros((1, 100, 100, 3), np.float32)

# Start from ORT 1.10, ORT requires explicitly setting the providers parameter if you want to use execution providers
# other than the default CPU provider (as opposed to the previous behavior of providers getting set/registered by default
# based on the build flags) when instantiating InferenceSession.
# Following code assumes NVIDIA GPU is available, you can specify other execution providers or don't include providers parameter
# to use default CPU provider.
sess = ort.InferenceSession("dst/path/model.onnx", providers=["CUDAExecutionProvider"])

# Set first argument of to None to use all model outputs in default order
# Input/output names are printed by the CLI and can be set with --rename-inputs and --rename-outputs
# If using the python API, names are determined from function arg names or TensorSpec names.
results_ort =["output1", "output2"], {"input1": input1})

import tensorflow as tf
model = tf.saved_model.load("path/to/savedmodel")
results_tf = model(input1)

for ort_res, tf_res in zip(results_ort, results_tf):
    np.testing.assert_allclose(ort_res, tf_res, rtol=1e-5, atol=1e-5)

print("Results match")

Conversion Failures

If your model fails to convert please read our README and Troubleshooting guide. If that fails feel free to open an issue on GitHub. Contributions to tf2onnx are welcome!

Next Steps