I am using tf.keras to construct an autoencoder that uses auxiliary information (such as the class label) in addition to the input image to perform better reconstructions. That\