Retval[0] does not have value : tf.cond(condition, net1, net2)

I was trying to use tf.cond() to create 2 different graphs based on a condition. On, both the graphs we want to have weight regularization loss and thus we use tf.losses.get_regularization_loss(). Here is a pseudo code of our project

 def net_1(x,y):
  statement 1 (has trainable params)
  statement 2 (has trainable params)
  return

def net_2(x,y):
 statement 1 (has trainable params)
 statement 2 (has trainable params)
 statement 3 (has trainable params)
 return

step = tf.get_or_create_global_step()


tf.cond(tf.greater(step, 100), net_1, net_2)
loss = 0.0
loss += tf.losses.get_regularization_loss()

If we keep the tf.losses.get_regularization_loss() we get the error:

Retval[0] does not have value

Otherwise, there is no error.

Is there any special care to be taken with tf.cond() if we want to impose tf.losses.get_regularization_loss().