Tensorflow 2.16 / Keras 3 support

[pwuertz]

The switch from Keras 2 to Keras 3 in Tensorflow 2.16 apparently breaks tf2onnx:

import tf2onnx
...
onnx_model, _ = tf2onnx.convert.from_keras(model)
# AttributeError: ... object has no attribute '_get_save_spec'

This is probably the same issue people are seeing with tf.lite.TFLiteConverter since Keras 3:
keras-team/keras#18430

Is there an alternative route like tf2onnx.convert.from_function we could use as a workaround?

[MidnessX]

I've discovered that explicitly passing an input signature makes the function work again:

import tf2onnx
import keras

inputs = [
     keras.Input((128, 2), batch_size=64, dtype="float32", name="input_1"),
     keras.Input((64, 4), batch_size=64, dtype="float32", name="input_2")
]
# [...] Rest of Keras code
model = keras.Model(inputs=inputs, outputs=outputs)

model_proto, _ = tf2onnx.convert.from_keras(
    model,
    input_signature=(
        tf.TensorSpec(
            model.inputs[0].shape,
            dtype=model.inputs[0].dtype,
            name=model.inputs[0].name,
        ),
        tf.TensorSpec(
            model.inputs[1].shape,
            dtype=model.inputs[1].dtype,
            name=model.inputs[1].name,
        ),
    ),
    output_path="model.onnx",
)

However, I do not know how robust this workaround is.
I suggest exporting the model to the TensorFlow saved_model format first using model.export("path/to/saved_model_folder") and then using:

python -m tf2onnx.convert --saved-model "path/to/saved_model_folder" --output "path/to/model.onnx"

to convert it to the ONNX format.

[dmagee]

Is there any news of getting t2fonnx with keras models working on recent versions of Keras/Tensorflow? I've tried the workarounds suggested above and tf2onnx.convert.from_keras() gives me an error "Cannot convert a symbolic tf.Tensor (input_1:0) to a numpy array." and model.export() gives error "AttributeError: module 'keras._tf_keras.keras.backend' has no attribute 'set_learning_phase'."

python 3.11 / tensorflow 2.16.1 / windows 10