In this paper, a phase retrieval method based on deep learning is proposed and applied to the phase-modulated holographic storage system. The phase-modulated holographic storage system has become a research hotspot because of its higher encoding rate and higher signal-to-noise ratio (SNR). Since the phase data cannot be detected directly by the detector, the intensity image is used to retrieve the phase. The traditional interferometric phase retrieval method is not suitable for the storage system because its optical system is complex and is easily affected by environmental disturbances. The non-interferometric phase modulation storage system uses iterative methods to solve the phase data, and the number of iterations will affect the data transmission rate in the holographic data system. In this paper, a simulated non-interferometric phase retrieval system based on deep learning is established, which uses a convolutional neural network to directly establish the relationship between phase and intensity images captured by CCD. The neural network is trained by learning the dataset of intensity images and phase data images. After training, the phase can be obtained by a single calculation, which greatly improves the data transmission speed. In the process of deep learning training, we introduced embedded data to improve the precision of phase reconstruction and reduce the bit error rate. According to our investigation, this is the first application of deep learning in phase retrieval of optical holographic storage.