Antimicrobial peptides (AMPs) are natural polypeptides with antibacterial activity and are an important part of the innate immune system. In order to solve the growing problem of conventional antibiotic resistance, AMPs have been applied in different fields as highly potential alternatives. It is of great significance that deep learning-based methods can quickly screen out candidate samples of AMPs from massive protein sequences to help discover new AMPs. In this paper, we designed a flexible and interpretative deep learning model(iAMP-CRA) using Convolutional Recurrent Neural Network with Self-Attention. Different sequence embedding encodings take into account both primary structural information and evolutionary information. Different feature extraction modules can learn feature representations from different aspects. Multiple feature descriptors of protein sequences were collected, and efficient feature descriptors were evaluated and screened out using different machine learning models as supplementary information. The introduction of attention mechanisms fuses complementary information and feature representations into a mixed feature representation that is used for the classification task of AMPs. Our model can learn both efficient sequence encodings and adaptively incorporate heterogeneous features. It shows excellent learning ability on the benchmark dataset. Compared with the latest methods, the accuracy on the independent testing set is 0.919, better than or comparable to the-state-of-art methods.