Based on Diffpool paper - notes on torch implementation. This is based on work within torch-geometric; and we extend it for node-level classification (on clustered nodes)
See: Hierarchical Graph Representation Learning with Differentiable Pooling https://arxiv.org/abs/1806.08804