Compressive sensing (CS) based methods have been widely used for sparse inverse synthetic aperture radar (ISAR) imaging. However, many CS-based methods are sensitive to the selection of model parameters, and the residual phase error of the echo also causes trouble for imaging and autofocusing. To address these problems, a novel deep learning approach, named as 2D-IADIANet, is proposed to achieve 2D sparse ISAR imaging with 2D phase error estimation in this paper. First, a 2D ISAR sparse echo model with 2D phase error into account is established, and a 2D-ADMM frame-work-based method, dubbed as 2D-IADIA, is presented to solve this compound reconstruction problem. Second, 2D-IADIA is further unfolded and mapped into a deep network form by integrating with 2D phase error compensation network. Moreover, all adjustable parameters can be learned adaptively by training the network through back propagation algorithm in complex domain directly. At last, experiment results verify that the well-learned 2D-IADIANet, which is only trained by a small amount of simulation samples, can also be generalized to measured data application. Especially, owing to the good performance of the network, the proposal has a superior reconstruction performance than 2D-IADIA under the low 2D sample rate and/or signal-to-noise ratio scenarios.