Summary: Recently, many learning-based methods have explored logic learning task in the assumption that the training set and testing set are from the consistent distribution, achieving good performance. But, in most cases, this assumption does not hold. In this paper, we explore this topic on the open-set logic reasoning task where the digit length and the sequence length of the training set and testing set are from inconsistent distributions. To address this issue, inspired by multigranulation studies in granular computing, we propose a granulation logic reasoning machine, namely GLRM. In this method, this open-set task is granulated into a series of sub-tasks from two dimensions: the digit length and the sequence length, and then these sub-tasks are conquered one by one. Finally, the results of the sub-tasks are organized into the final result. The effectiveness of GLRM is demonstrated by experiments on the open-set Fashion-Logic data set and the open-set Fashion-Logic task proposed in this paper. This study provides a novel view for solving open-set logic reasoning tasks and promotes the research of data-driven logic learning.