Recently, in a massive multiple-input multiple-output (MIMO) system, deep learning (DL)-based beamforming method has been proposed for reducing the overhead associated with downlink training and uplink feedback. However, the DL-based approach is sensitive to the variation of the communication environment and requires a huge number of training data to ensure a certain level of performance. To reduce the number of required channel data for training a deep neural network (DNN), we introduce deep transfer learning (DTL), which exploits the information from the pre-trained DNN for training other DNNs to find the beamforming vector in the specific channel. Through DTL, DNN can be trained suitably for the communication environment at each BS with fewer channel data. Moreover, we propose 'step-by-step' DTL to flexibly apply DTL considering the uncertainties of the realistic system. Simulation results show that DTL has better performance than the conventional DL-approaches even with a small amount number of channel data. Therefore, the DTL-based approach can be a good framework to train DNN when high overhead occurs or designing the beamformer is complicated such as a massive MIMO system.