In the previous section, we saw how the tf.data API simplifies and standardizes the input pipeline definition. Also, we saw that the tf.data API is completely integrated into the TensorFlow Keras implementation and the eager or graph-accelerated version of a custom training loop.
Just as for the input data pipelines, there are a lot of repetitive parts in the whole machine learning programming. In particular, after defining the first version of the machine learning model, the practitioner is interested in:
- Training
- Evaluating
- Predicting
After many iterations of these points, exporting the trained model for serving is the natural consequence.
Of course, defining a training loop, the evaluation process, and the predicting process are very similar for each machine learning process. For example, for a predictive model, we are interested in training the model for a...
