Contradiction of weight sampling procedure between the code and paper

[MlWoo]

            sess_list1 = self.apply_ + self.sd_asn
            sess_list2 = [self.train_step, self.cross_entropy, self.accuracy]
            result1 = self.sess.run(sess_list1, feed_dict = feed_dict)
            result2 = self.sess.run(sess_list2, feed_dict = feed_dict)

As above, the code will execute as the following sequence,
apply_ : sample a new weight with mean (the original weight) and the standard deviation
sd_asn: update the standard deviation with the gradients
train_step: update the the weight with the gradients
cross_entropy: compute the loss after a forward pass

However, the whole procedure which I got from the paper is:
sd_asn: update the standard deviation with the gradients
train_step: update the the weight with the gradients
apply_ : sample a new weight with mean (the original weight) and the standard deviation
cross_entropy: compute the loss after a forward pass

Do I misunderstand something?