const std::vector<Gradient*>& in_grads, absl::Span<Gradient*> out_grads) {
  /* This function is approximately equivalent to this Python code:

  forwardprop_aids = tf.ones_like(output_tensors)
  with tf.GradientTape() as g:
    g.watch(forwardprop_aids)
    grad = backward_function(forwardprop_aids)
  forward_grads = g.gradient(grad, forwardprop_aids, output_gradients=in_grads)