Failure loading a Keras model after conversion from Sequential to Functional

[pyhopp]

Hi, this issue is observed on Tensorflow version 2.16 (it works as expected on version 2.15 and earlier).

Problem Description:
The following sequence causes a load_model error:

1. Create a Sequential model, and save it
2. Load the Sequential model
3. Convert the sequential model to a functional model
4. Save the functional model
5. Load the function model ---> error

Reproducible example
`

code

seq = keras.Sequential(
    [
        keras.Input(shape=(3,)),
        keras.layers.Dense(5),
        keras.layers.Softmax(),
    ],
)

seq.save("keras_model.keras")

# Load sequential model
seq = keras.models.load_model("keras_model.keras")

# Create functional model
func = keras.Model(inputs=seq.inputs, outputs=seq.outputs)

# Save functional model
func.save("functional_model.keras")

# Load functional model
func = keras.models.load_model("functional_model.keras")`

Error output

func = keras.models.load_model("functional_model.keras") ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.11/dist-packages/keras/src/saving/saving_api.py", line 176, in load_model return saving_lib.load_model( ^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.11/dist-packages/keras/src/saving/saving_lib.py", line 155, in load_model model = deserialize_keras_object( ^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.11/dist-packages/keras/src/saving/serialization_lib.py", line 711, in deserialize_keras_object instance = cls.from_config(inner_config) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.11/dist-packages/keras/src/models/model.py", line 492, in from_config return functional_from_config( ^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.11/dist-packages/keras/src/models/functional.py", line 522, in functional_from_config process_node(layer, node_data) File "/usr/local/lib/python3.11/dist-packages/keras/src/models/functional.py", line 466, in process_node args, kwargs = deserialize_node(node_data, created_layers) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.11/dist-packages/keras/src/models/functional.py", line 666, in deserialize_node args = tree.map_structure(convert_revived_tensor, args) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.11/dist-packages/tree/__init__.py", line 435, in map_structure [func(*args) for args in zip(*map(flatten, structures))]) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.11/dist-packages/tree/__init__.py", line 435, in <listcomp> [func(*args) for args in zip(*map(flatten, structures))]) ^^^^^^^^^^^ File "/usr/local/lib/python3.11/dist-packages/keras/src/models/functional.py", line 656, in convert_revived_tensor raise ValueError( ValueError: Layer node index out of bounds. inbound_layer = <Dense name=dense, built=True> inbound_layer._inbound_nodes = [<Node operation=<Dense name=dense, built=True>, id=140157363310928>] inbound_node_index = 1

Useful (?) observation

If when I create the functional model i pass all the outputs of the sequential model instead of just model.outputs then I get no errors (however I end up with a model that outputs the result of every layer).
outputs=[layer.output for layer in seq.outputs]

Please note that it doesn't matter whether the models are built/compiled/trained/ we still experience the error as described above.

[SuryanarayanaY]

Hi @pyhopp ,

Thanks for reporting. Replicated the reported issue with Keras and attached gist for reference.

[fchollet]

Thanks for the report! I fixed the issue at HEAD.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No

[pyhopp]

Thanks for the help!

Will the fix get rolled out in the next major (minor?) version update?

[fchollet]

In the next pip release. Within a few weeks.

[pyhopp]

Great. Thanks again!