In the PhysioNet/Computing in Cardiology Challenge 2021, our team, DrCubic, develops a novel approach to classify cardiac abnormalities using reduced-lead ECG recordings. In our approach, we incorporate peak detection as a self-supervised auxiliary task. We build the model based on SE-ResNet, and ensemble models of different input lengths and sampling rates. Inspired by last year's challenge results, we investigate various settings and techniques, and select the best ones, considering the intra-source performance and inter-source generalization simultaneously. Our classifiers receive scores of 0.666, 0.643, 0.642, 0.651, and 0.639 (ranked 3rd, 3rd, 4th, 4th, and 3th out of the teams) for the 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead versions of the hidden validation set with the Challenge evaluation metric.