load_model_from_hdf5 doesn't load axis param for Concatenate layer

[gkoundry]

When loading a keras model from an hdf5 file the axis parameter is dropped from the model config.
The code at https://github.com/keras-team/keras/blob/6c3dd68c1b2e7783d15244279959e89fe88ee346/ pops the param from the config and only puts it back if it's a list which AFAIK it never is.

Script to reproduce the issue:

from tensorflow.keras.layers import Input, Dense, Concatenate
from tensorflow.keras.models import Model
from keras.src.legacy.saving import legacy_h5_format

input1 = Input(shape=(1,4), name='input1')
input2 = Input(shape=(1,4), name='input2')
input3 = Input(shape=(2,4), name='input3')
concat1 = Concatenate(axis=1)([input1, input2])
concat2 = Concatenate(axis=-1)([concat1, input3])
output = Dense(1, activation='sigmoid')(concat2)

model = Model(inputs=[input1, input2], outputs=output)

model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

legacy_h5_format.save_model_to_hdf5(model, "cat_test.h5")
legacy_h5_format.load_model_from_hdf5("cat_test.h5")

Running this script results in load_model_from_hdf5 failing with an error ValueError: A Concatenate layer requires inputs with matching shapes except for the concatenation axis. Received: input_shape=[(None, 1, 8), (None, 2, 4)] concatenate_1

[mehtamansi29]

Hi @gkoundry -

Thanks for reporting the issue. Here if you ran your reproducible code with latest keras(3.8.0) and instead of saving model .h5 try to use .keras extension like this. Code works fine without error.

keras.saving.save_model(model, "cat_test.keras")
model1 = keras.saving.load_model("cat_test.keras")

Attached gist for the reference here.

[gkoundry]

I know that there is a better way to do this but I had some older HDF5 models that I was trying to load which is why I ran into this. It does seem to be a bug and was difficult to track down so I figured I would just report it anyways just to save someone else a headache if they ran into it. However I understand if you decide it's not worth it to maintain old legacy code.