In recent years, increasing attention has been paid to network representation learning, which aims to map nodes into low dimensional vectors while preserving topology and node attribute information, which are both backbone information of the network. Existing studies mainly focus on fusing structure and node attributes on single granularity for the attributed network. However, many complex networks present multi-granular characteristics. In this paper, we propose MultI-granular attributed network Representation Learning (MIRL), an algorithm that captures the relationship between different granular attributed networks. Firstly, topological structure and attributes are fused from fine to coarse under different granularities to mine the node potential relationship between different granular networks. The coarser-grained node is composed of a number of fine-grained nodes that are similar in structure and attributes. For the attributed network at the coarsest granularity which is much smaller than the original attributed network, one of the existing network representation learning methods can be used to learn the representation of the coarsest granularity. To obtain more accurate representation of the original network, we train a graph convolutional neural network (GCN) at the coarsest granulation. The parameters of GCN passing from coarse to fine are shared between two adjacent granularities, so as to trade off time consumption and embedding performance. We evaluate our algorithm on three real-world datasets and two benchmark applications. Our experimental results demonstrate that MIRL significantly increases effectiveness compared to state-of-art network representation methods.