High frequency oscillations (HFOs) in intracranial electroencephalography (iEEG) recordings are considered as promising clinical biomarker of epileptogenic regions of the brain. Previous automatic detection methods for HFOs require plenty of labeled training samples to build effective computational models, while expert labeling of HFOs is an expensive, subjective, time-consuming, and laborious task. Therefore, to relieve the problem of the severe lacking of HFOs labels and reduce labeling costs, we propose a label generation method for HFOs called Consistency Residual Memory (CRM). Firstly, to focus on temporal and spatial features, we design a network architecture composed of a Recurrent Neural Network (RNN) and a Convolutional Neural Network (CNN). Secondly, to take full advantage of a large amount of unlabeled data, we explore a semi-supervised algorithm consisting of supervised and unsupervised components. The supervised component minimizes the cross-entropy of the input and output labels based on labeled training samples. Meanwhile, the unsupervised component maximizes the consistency of the output, based on the original input and the disturbed input, utilizing the entire training set ( labeled and unlabeled). We validated the high accuracy and sensitivity of CRM method on a large-scale public benchmark dataset of HFOs and a private clinical dataset. Our method achieves state-of-the-art performance, with high accuracy (91.83%) and sensitivity (96.19%) on the public datasets, and achieves an accuracy (87.83%) and sensitivity (91.10%) on the private clinical datasets. It is demonstrated that CRM method can be applied in clinical applications, which relieves the difficulty of the clinician in the signal labeling and visual analysis of HFOs.