python-kerasHow can I convert a TensorFlow Keras model to ONNX using keras2onnx?

To convert a TensorFlow Keras model to ONNX using keras2onnx, you need to install the keras2onnx package and import the necessary modules.

pip install keras2onnx

from keras.models import load_model
import keras2onnx
import onnx

Then, you can load the Keras model and convert it to ONNX.

model = load_model('model.h5')
onnx_model = keras2onnx.convert_keras(model, model.name)
onnx.save_model(onnx_model, 'model.onnx')

The above example code will save the converted model as model.onnx.

You can also specify the target ONNX opset version, input and output names, and the input shape.

onnx_model = keras2onnx.convert_keras(model, model.name, target_opset=7,
    input_names=['input_1'], output_names=['output_1'],
    custom_conversion_functions={'MyLayer': my_conversion_function},
    inputs=[(input_shape, 'input_1')])

• target_opset: ONNX opset version (defaults to 9)
• input_names: list of input names (defaults to ['input_1'])
• output_names: list of output names (defaults to ['output_1'])
• custom_conversion_functions: dictionary of custom conversion functions for layers not supported by keras2onnx
• inputs: list of tuples containing input shape and name

For more information, see the keras2onnx documentation.

Edit this code on GitHub