Can gradient tape in tensorflow differentiate nested functions of the same network?

Question

That is to say if I have a differentiable model g, and a differentiable function f (which could also include models).

with tf.GradientTape(persistent=True) as tape:
  for_ in range(n):
     r = g(r)
  loss = f(r)
grad = tape.gradient(loss, g.trainable_variables)

would tape.gradient apply backpropagation through time for n steps on g?

Answer 1

You can use matplotlib to create two graphs and see how they compare. The red line is for the target and the blue line is for the prediction result.

tf.GradientTape(
persistent=False, watch_accessed_variables=True

enter image description here