Fourier ptychography (FP) can reconstruct the amplitude and phase distribution of objects with a wide field of view and high. With the continuous development of deep learning, neural network has become one of the important methods to deal with the nonlinear inverse problems in computational imaging. Aiming at the characteristics of FP system such as strong data specificity and small amount of data, this paper proposes an algorithm combining computational imaging prior knowledge and deep learning, to design a neural network framework based on physical model, and verifies it on simulation samples. Furthermore, a far-field transmission system is constructed to verify the FP reconstruction of image sequences of macroscopic objects. Experimental results show that the system can reconstruct the complex amplitude distributions of high-resolution samples using limited simulation and real data sets, with high robustness to optical aberration and background noise.