Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow, 3rd Edition

Appendix D. TensorFlow Graphs

In this appendix, we will explore the graphs generated by TF functions (see Chapter 12).

TF Functions and Concrete Functions

TF functions are polymorphic, meaning they support inputs of different types (and shapes). For example, consider the following tf_cube() function:

@tf.function
def tf_cube(x):
    return x ** 3

Every time you call a TF function with a new combination of input types or shapes, it generates a new concrete function, with its own graph specialized for this particular combination. Such a combination of argument types and shapes is called an input signature. If you call the TF function with an input signature it has already seen before, it will reuse the concrete function it generated earlier. For example, if you call tf_cube(tf.constant(3.0)), the TF function will reuse the same concrete function it used for tf_cube(tf.constant(2.0)) (for float32 scalar tensors). But it will generate a new concrete function if you call tf_cube(tf.constant([2.0])) or tf_cube(tf.constant([3.0])) (for float32 tensors of shape [1]), and yet another for tf_cube(tf.constant([[1.0, 2.0], [3.0, 4.0]])) (for float32 tensors of shape [2, 2]). You can get the concrete function for a particular combination of inputs by calling the TF function’s get_concrete_function() method. It can then be called like a regular function, but it will only support one input signature (in this example, float32 scalar tensors):

>>> concrete_function = tf_cube.get_concrete_function ...