When defining a tensorflow.function it is generally a bad practice to make this function recursive. TensorFlow does not support
recursive tensorflow.function and will in the majority of cases throw an exception. However it is possible as well that the execution of
such function succeeds, but with multiple tracings which has strong performance implications. When executing tensorflow.function, the
code is split into two distinct stages. The first stage call tracing creates a new tensorflow.Graph, runs the Python code
normally, but defers the execution of TensorFlow operations (i.e. adding two Tensors). These operations are added to the graph without being ran. The
second stage which is much faster than the first, runs everything that was deferred previously. Depending on the input of the
tensorflow.function the first stage may not be needed, see: Rules of
tracing. Skipping this first stage is what provides the user with TensorFlow’s high performance.
Having a recursive tensorflow.function prevents the user from benefiting of TensorFlow’s capabilities.
