Automatic detection and classification of cardiac abnormalities facilitates the prevention and diagnosis of cardiac diseases. Aiming at classify cardiac abnormalities into 27 classes with either 12-lead, 6-lead,3-lead or 2-lead multi-label ECG recordings, we develop a deep convolutional neural network with residual block and attention mechanism. For the convenience of training, all recordings are cropped or padded to 60 seconds with resampled rates of 300 Hz. The algorithm was trained and validated on the PhysioNet Challenge 2020 dataset, and the F1 scores on 12-lead, 6-lead, 3-lead and 2-lead recordings are 0.5,0.53,0.47 and 0.45 respectively. During unofficial phase, our model achieves 0.41,0.45,0.37 and 0.38 respectively on the official test data( Team name: Proton ) .