The tff.learning.reconstruction.Model returned by this function uses
keras_model for its forward pass and autodifferentiation steps. During
reconstruction, variables in local_layers are initialized and trained.
Post-reconstruction, variables in global_layers are trained and aggregated
on the server. All variables must be partitioned between global and local
layers, without overlap.
Iterable of global layers to be aggregated across users. All
trainable and non-trainable model variables that can be aggregated on the
server should be included in these layers.
Iterable of local layers not shared with the server. All
trainable and non-trainable model variables that should not be aggregated
on the server should be included in these layers.
A structure of tf.TensorSpecs specifying the type of arguments
the model expects. Notice this must be a compound structure of two
elements, specifying both the data fed into the model to generate
predictions, as its first element, as well as the expected type of the
ground truth as its second.