We propose a novel algorithm for unsupervised graph representation learning\nwith attributed graphs. It combines three advantages addressing some current\nlimitations of the literature: i) The model is inductive: it can embed new\ngraphs without re-training in the presence of new data; ii) The method takes\ninto account both micro-structures and macro-structures by looking at the\nattributed graphs at different scales; iii) The model is end-to-end\ndifferentiable: it is a building block that can be plugged into deep learning\npipelines and allows for back-propagation. We show that combining a coarsening\nmethod having strong theoretical guarantees with mutual information\nmaximization suffices to produce high quality embeddings. We evaluate them on\nclassification tasks with common benchmarks of the literature. We show that our\nalgorithm is competitive with state of the art among unsupervised graph\nrepresentation learning methods.\n