Feasible real-time ECG classification algorithms contribute to an early and correct diagnosis of cardiac abnormalities. In this paper, we (Team Triology) leverage 80 Hz rather than 500 Hz ECG signals to realize a smaller input size using a lightweight end-to-end neural network, because a 3.2-second-long (256/80) segment, containing around 3 beats, provides sufficient long-range information. A soft voting scheme is applied to determining the prediction in a long record. Our backbone is ResNet-18 with anti-aliased blocks so that the model possesses shift invariance and the shortcut connection preserves high-frequency information. Moreover, to incorporate the imbalance among classes, focal loss and weighted cross entropy are randomly selected in the training process. In this way, the correlation between labels accelerates model convergence, due that most of records have only one label in the multi-label task. In order to derive a robust algorithm, random erasing, random shifting and Gaussian noise are implemented as data augmentation approaches. Currently, based on 6,877 (male: 3,699; female: 3,178) 12-ECG recordings lasting from 6 seconds to 60 seconds, our model achieves good performance with an F-score of 0.755 and a Jaccard measure of 0.542 on the official unseen test set. Future work includes integrating multi-scale information and some traditional features into the current backbone. We will also perform offline 5-fold cross validation.