Huge overhead of beam training poses a significant challenge in millimeter-wave (mmWave) wireless communications. To address this issue, a wide beam based training method is proposed to calibrate the narrow beam direction according to the channel power leakage in this paper. To handle the complex nonlinear properties of the channel power leakage, deep learning is utilized to predict the optimal narrow beam directly. Specifically, three deep learning assisted calibrated beam training schemes are proposed. First, Convolution Neural Network (CNN) is adopted to implement the prediction based on the instantaneous received signals of wide beam training. Besides, we propose to perform the additional narrow beam training based on the predicted probabilities for further beam direction calibrations. In the second scheme, in order to enhance the robustness to noise, Long-Short Term Memory (LSTM) Network is adopted to track the movement of users and calibrate the beam direction according to the received signals of prior beam training. To further reduce the overhead of wide beam training, an adaptive beam training strategy is proposed, where partial wide beams are selected to be trained based on the prior received signals. Furthermore, two criteria, namely optimal neighboring criterion (ONC) and maximum probability criterion (MPC), are designed for the selection. To handle mobile scenarios, auxiliary LSTM is introduced to calibrate the directions of the selected wide beams more precisely. Simulation results demonstrate that our proposed schemes achieve significantly higher beamforming gain with smaller beam training overhead compared with the conventional counterparts.