The current data iterators are slow. We could use the tf.data iterators. Does that work with dynamic input/output dimensions?