Nest error on using keras.callbacks.ModelCheckpoint at multi-input(maybe output also) model

[WondooSeo]

Hi, I found some issue on multi-input model problem when I add keras.callbacks.ModelCheckpoint as callback in fit. I use tensorflow 2.6.0 and python 3.9. I don't use tf-nightly because of laptop environment(1660ti).

When I add

keras.callbacks.ModelCheckpoint(
        filepath='best_model.keras',
        monitor='val_loss',
        save_best_only=True
    )

in my callback method in model fit, it raise error below;

ValueError: The target structure is of type <class 'keras.engine.keras_tensor.KerasTensor'>
KerasTensor(type_spec=TensorSpec(shape=(None, 22), dtype=tf.float32, name='input_1'), name='input_1'...
However the input structure is a sequence (<class 'list'>) of length 0.
[]
nest cannot guarantee that it is safe to map one to the other.

But when I get rid of this option, it works!

I'm really sorry that I can't open my code since its secure. But I was really struggled with handling this error.

Roughly to say my model, there are two inputs and my pre-trained functional model, and concatenate two outputs of pre-trained functional model. Those two functional model is non-trainable(== freezed) things. And there is output with some layers.

Is there any solution about this problem?