Driver fatigue is a major cause of traffic accidents, and electroencephalography (EEG) based driver fatigue classification is widely regarded as a future direction. In practical applications, different tasks may require different classification approaches. However, traditional fatigue classification approaches are task-specific, ignoring adaptive classification for different tasks. Achieving adaptive classification is challenging due to diversity of tasks, individual differences, etc. This paper proposes an adaptive fatigue classification framework to extend the applicability of EEG-based techniques. The proposed framework adjusts classification granularities based on requirements of different tasks. In order to capture information suitable for different granularities, the proposed framework includes an attention-based hybrid neural network (AHNN). AHNN combines 1D convolutions with attention mechanisms to enhance relationships between features and channels while suppressing the influence of insensitive information. To improve the generality of the proposed framework among individuals, an individual feature selection approach is added to generate required feature subsets for different individuals to weaken individual differences. We conduct subject-dependent and subject-independent experiments using the public SEED-VIG dataset to verify the feasibility of our proposed framework. In subject-dependent experiments, classification accuracy is 92.97% and 97.86% for fine-grained and coarse-grained classification, respectively, while it is 87.05% and 95.64% in subject-independent experiments. Experimental results show that the framework could be applied to different tasks, and outperforms some existing studies and classical models in specific tasks. AHNN could suit different classification granularities, and weaken individual differences, improving the generality of the framework. The proposed framework could provide a reference for future adaptive classification technique researches.