AttributeError: 'Functional' object has no attribute '_get_save_spec'

[kerighan]

I get this error when trying to convert/quantize any .keras model using tflite (model was trained using TF):

converter = tf.lite.TFLiteConverter.from_keras_model(model)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.target_spec.supported_types = [tf.float16]
tflite_quant_model = converter.convert()

Traceback:

  input_signature = _model_input_signature(
  File "/[...]/tflite_keras_util.py", line 84, in model_input_signature
    input_specs = model._get_save_spec(  # pylint: disable=protected-access
AttributeError: 'Functional' object has no attribute '_get_save_spec'

[fchollet]

A .keras model is not meant to be consumed by anything other than keras.models.load_model(). To use the TF-Lite converter, first create a TF SavedModel via tf.saved_model.save(model).

[innat]

Supporting this might be useful. Usually onnx is needed to convert to torch model to tf-lite. So supporting this from Keras-Core might be effective. However, torch.onnx.export exist.

[batrlatom]

@innat @fchollet what is the working pipeline to go from the keras model through torch training to the tf-lite export? viz. https://github.com/keras-team/keras-core/issues/746 . unfortunatelly, even onnx export does not work as expected

[fchollet]

I'm told the TFLite team have got TFLite export working internally for Keras Core at Google, but the code has not yet been released. Hopefully you'll get an update soon.

[pwuertz]

Probably the same issue tf2onnx is running into when moving from Keras2 to Keras3.

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

[mehtamansi29]

Hi @kerighan -

By loading .keras model using keras.saving.load_model, then try to use TFLite to convert/quantize the model will work fine. Attached gist for your reference.