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 ECG signals to develop a lightweight end-to-end neural network. A soft voting scheme is applied to determining the prediction in a long record from multiple segments. The model has a ResNet-18 backbone. It integrates standard and dilated convolutions to extract multi-scale information. Anti-aliased blocks are used for shift invariance. Age and sex are included. To encourage the inter-class competition in the multi-label classification task, lovász softmax and weighted cross entropy loss are randomly selected in the training process, which facilitates model convergence. In order to derive a robust model, data augmentation approaches like Gaussian noise, random erasing and shifting are implemented. Our offline validation is carried out using databases from four sources. We score 0.328 using the challenge metric. False negatives are main errors.