When you trained Z12, you do not need to train Z1 and Z2 again. They are subsets of Z12, so the parameters are shared.