It's a useful and commonly function to use a trained model to init the network. But there is a bug by using from_tar function. This interface is only suitable that all the parameter names are same in the initialized model and the network. If the users want to init partial network parameters through the initialized model, it will go wrong. The bug is that the Parameters only contains the parameters in the initialized model and will omit some parameters in the defined network by using from_tar.