deep_gcns_torch

Why we included this project

This is the PyTorch reference implementation of DeepGCNs, the line of research that borrowed depth tricks from CNNs, like residual and dense connections and dilated convolutions, and applied them to graph convolutional networks. Anyone working with 3D point clouds will find ready-to-run pipelines for semantic segmentation on S3DIS and PartNet and classification on ModelNet40, each folder carrying its own instructions, data, and pretrained models. The later DeeperGCN and GNN1000 variants are here too, so it works as a single place to compare very deep and memory-efficient GNN designs on OGB graph datasets. The core graph layers sit in a modular gcn_lib with dense and sparse backends, which makes it feasible to lift the code into your own architectures instead of only reproducing the papers.