An op outside of the function building code is being passed a "Graph" tensor

Hello, I'm trying to make a differentially private GAN, however when computing the gradients for my discriminator I get the error message

TypeError: An op outside of the function building code is being passed
a "Graph" tensor. It is possible to have Graph tensors
leak out of the function building context by including a
tf.init_scope in your function building code.
For example, the following function will fail:
  @tf.function
  def has_init_scope():
    my_constant = tf.constant(1.)
    with tf.init_scope():
      added = my_constant * 2
The graph tensor has name: strided_slice:0

What's odd is that I can run the training step for the discriminator (computing the gradients) twice, however on the third iteration I get this error.

I have tried my code with TensorFlow 2.x and 1.x but I seem to be getting the same error on either.

No repository file or test is identified. Start by reducing the reported discriminator gradient computation to a reproducible Python/TensorFlow example, focusing on why the third iteration differs from the first two. Done means a minimal reproduction and an actionable diagnosis or fix, with regression coverage if the issue can be reproduced.