Nework build pattern using `self.call` not compatible with Keras functional API

[vpratz]

When using the Keras functional API, the networks are automatically built. This includes passing a shape with the batch_size specified as None.

inputs = keras.layers.Input(shape=(10, 2))
outputs = bf.networks.DeepSet()(inputs)
model = keras.Model(inputs=inputs, outputs=outputs)

This errors in the second line with the following message (TensorFlow backend):

[bayesflow/networks/deep_set/deep_set.py#line=82), in DeepSet.build(self, input_shape)
     81 def build(self, input_shape):
     82     super().build(input_shape)
---> 83     self.call(keras.ops.zeros(input_shape))

ValueError: Attempt to convert a value (None) with an unsupported type (<class 'NoneType'>) to a Tensor.

Do we want to support using the networks in the functional API? If so, we should extend the self.call(keras.ops.zeros(input_shape)) by an appropriate check for None in the first dimension, and replace it with a dummy batch size.

[stefanradev93]

I personally never use the functional API but if at least one person thinks it could be relevant, we should probably do it using a simple helper function that checks if a tuple contains None entries and fills them out with dummy dimensions.

[vpratz]

Yes, I have a use case where this would be nice to have. I will take a look at it...