In recent years, deep learning has emerged in the audio field with many excellent models and beats non-depth methods in the quality of generated audio. This paper implements a symbol-based end-to-end music generation model. This model generates piano music corresponding to the pitch of the musical score using a two-dimensional “Piano-roll” liked structure as input. The experiments show the generated music obtains good performance and achieves a result similar to the original song in pitch, melody, and timbre. Compared with other generation methods, the input of our model is simple, easy to obtain, and can generate music through an end-to-end method.