Loading

Another advantage of tf.data is that all its operations are registered in the TensorFlow operational graph, and the extracted and processed samples are returned as Tensor instances. Therefore, we do not have much to do regarding the final step of ETL, that is, the loading. As with any other TensorFlow operation or tensor, the library will take care of loading them into the target devices—unless we want to choose them ourselves (for instance, wrapping the creation of datasets with tf.device()). When we start iterating over a tf.data dataset, generated samples can be directly passed to the models.