In this letter, an efficient model-driven deep learning (DL) based massive multiple-input multiple-output (MIMO) detector is proposed by improving the approximate expectation propagation (EPA) algorithm, named EPANet. Specifically, EPANet is constructed by unfolding the iterative EPA detector and adding learnable parameters to enhance the performance and convergence robustness through the DL approach. Only one training procedure is required in advance for EPANet to be reused for multiple detection tasks with different antenna configurations. Numerical results indicate that DL can bring significant performance improvement to EPA with various antenna settings. Besides, the proposed EPANet can outperform state-of-the-art (SOA) DL-based detectors with hardware-friendly complexity, especially under highly-correlated channels.