Recently, coded caching scheme was proposed as the ability of alleviating the load of networks. Especially, the placement delivery array (PDA) used for characterizing the coded caching scheme has attracted vast attention. In this letter, a deep neural architecture is first proposed to learn the construction of PDAs for reducing the computational complexity. The problem of variable size of PDAs is solved using mechanism of neural attention and reinforcement learning. Different from previous works using combined optimization algorithms to get PDAs, our proposed deep neural architecture uses sequence-to-sequence model to learn construct PDAs. Numerical results are given to demonstrate that the proposed method can effectively implement coded caching meanwhile reducing the computational complexity.