This paper investigates a transfer learning application for predicting software faults. Detecting faulty modules in software projects is challenging due to two main issues 1) the low quality of existing handcrafted features leads to the bad performance of traditional learning models and 2) the shortage of annotated data hinders applying deep neural networks. Recently, transfer learning is a good solution to train deep neural networks with insufficient data. Our experiments for tasks of within-project and cross-project software fault prediction have shown the transferable possibility among project data. As a result, the performance of the base model is significantly improved and achieves competitive results with the state of the art method.