Electrocardiogram (ECGs) is an important tool to diagnose cardiovascular disease, and is widely used in the medical field. Many machine learning approaches have been propose to automatically diagnose irregular ECGs. Among these approaches, classification model using deep neural networks (DNNs) have shown significant result. In contrast to the previous approaches which required manual feature engineering, DNNs extracts features automatically from training data. In the classification of irregular ECGs with DNNs, one dimensional convolutional neural networks with residual connections are often used which was inspired by ResNet used for image classification. On the other hand, Efficient Net have shown better accuracy than ResNet in image classification with large dataset and object detection. Thus in this study, we develop Efficient Net based one dimensional sequence classification model for irregular ECG classification. In addition, we apply data augmentation and utilize unlabeled data to improve classification accuracy. For data augmentation, we apply RandAugment, which is known to improve generalization performance in image classification with large dataset. RandAugment in image classification randomly combines the augmentations used in traditional image classification, such as contrast adjustment, image padding, and rotation. In this study, we apply a random combination of amplitude changes and padding as applicable augmentations to ECG data. We also apply self-training with noisy student, which generates pseudo-labels for unlabeled data and train them together with labeled data. We improve the generalization performance of ECG classification models by increasing data variation by combining the use of unlabeled external data and data augmentation approaches, which are known to improve the accuracy in image classification tasks. Currently our model achieved score of 0.661 on hidden test set.