Published spatiotemporal GNNs, packaged as PyTorch Geometric layers

What it does PyTorch Geometric Temporal layers temporal awareness on top of PyTorch Geometric. It bundles implementations of dynamic graph neural networks—recurrent convolutions like GConvGRU and DCRNN, plus attention-based aggregations such as STGCN and ASTGCN—into standard PyTorch modules. The library also provides temporal snapshot iterators, train-test splitters, and benchmark datasets drawn from traffic forecasting, epidemiology, energy production, and web traffic management.

The interesting bit Rather than forcing you to reimplement papers, the library treats each architecture as a reusable torch.nn.Module you can stack like any other PyTorch Geometric layer. A newer index-batching trick cuts memory usage on large spatiotemporal datasets without hurting accuracy, and the whole stack integrates with PyTorch Lightning and Dask-DDP if you want to scale out.

Key highlights

• Implements over a dozen published architectures—DCRNN, EvolveGCN, AGCRN, STGCN, ASTGCN, and others—as composable layers.
• Ships with benchmark datasets for traffic, epidemiological forecasting, web traffic, and energy production.
• Uses index-batching to keep memory efficient when scaling to larger temporal graphs.
• Works with PyTorch Lightning and Dask-DDP for distributed and multi-GPU training.
• Includes temporal snapshot iterators and splitters so you don’t hand-roll time-window logic.

Verdict Grab this if your PyTorch Geometric pipeline needs to handle graph sequences for forecasting or dynamic link prediction. If your graphs never change between snapshots, you don’t need the extra baggage.