A new major release, introducing TorchScript support, memory-efficient aggregations, bipartite GNN modules, static graphs and much more!

Major Features

• TorchScript support, see here for the accompanying tutorial (thanks to @lgray and @liaopeiyuan)
• Memory-efficient aggregations via torch_sparse.SparseTensor, see here for the accompanying tutorial
• Most GNN modules can now operate on bipartite graphs (and some of them can also operate on different feature dimensionalities for source and target nodes), useful for neighbor sampling or heterogeneous graphs:

conv = SAGEConv(in_channels=(32, 64), out_channels=64)
out = conv((x_src, x_dst), edge_index)

• Static graph support:

conv = GCNConv(in_channels=32, out_channels=64)

x = torch.randn(batch_size, num_nodes, in_channels)
out = conv(x, edge_index)
print(out.size())
>>> torch.Size([batch_size, num_nodes, out_channels])

Additional Features

• PNAConv (thanks to @lukecavabarrett and @gcorso)
• Pre-Trained DimeNet on QM9
• SEAL link prediction example (thanks to @muhanzhang)
• ClusterGCNConv
• Cluster-GCN PPI example (thanks to @CFF-Dream)
• WeightedEdgeSampler for GraphSAINT (thanks to @KiddoZhu)
• Better num_workers support for GraphSAINT
• The automatic addition of self-loops can now be disabled via the add_self_loops argument, e.g., for GCNConv

Breaking Changes

• Memory-efficient RGCNConv: The old RGCNConv implementation has been moved to FastRGCNConv

Complementary Frameworks

• DeepSNAP: A PyTorch library that bridges between graph libraries such as NetworkX and PyTorch Geometric
• PyTorch Geometric Temporal: A temporal GNN library built upon PyTorch Geometric

Datasets

• GNNBenchmarkDataset suite from the Benchmarking Graph Neural Networks paper
• WordNet18

Bugfixes

• Fixed a bug in the VGAE KL-loss computation (thanks to @GuillaumeSalha)