This rule raises an issue when a tensorflow.Variable created inside of a tensorflow.function is not a singleton.

Why is this an issue?

tensorflow.functions only supports singleton tensorflow.Variables. This means the variable will be created on the first call of the tensorflow.function and will be reused across the subsequent calls. Creating a tensorflow.Variable that is not a singleton will raise a ValueError.

import tensorflow as tf

@tf.function
def f(x):
  v = tf.Variable(1.0)
  return v

In the example above each time the function f will be called a new tensorflow.Variable will be created.

How to fix it

To fix this issue use the singleton pattern to avoid the creation of a tensorflow.Variable multiple times.

Code examples

Noncompliant code example

import tensorflow as tf

class Count(tf.Module):
  def __init__(self):
    self.count = None

  @tf.function
  def __call__(self):
    self.count = tf.Variable(0) # Noncompliant: the variable will be re-created for each call
    return self.count.assign_add(1)

c = Count()
c() # ValueError is raised
c()

Compliant solution

import tensorflow as tf

class Count(tf.Module):
  def __init__(self):
    self.count = None

  @tf.function
  def __call__(self):
    if self.count is None:
      self.count = tf.Variable(0) # Compliant: the variable is created only on during the first call
    return self.count.assign_add(1)

c = Count()
c()
c()

Resources

Documentation