This update includes index-batching. Index-batching is a technique that reduces the memory cost of training ST-GNNs with spatiotemporal data with no impact on accuracy, enabling greater scalability and training on the full PeMS dataset without graph partitioning for the first time. Leveraging the reduced memory footprint, this technique also enables GPU-index-batching - a technique that performs preprocessing entirely in GPU memory and utilizes a single CPU-to-GPU mem-copy in place of batch-level CPU-to-GPU transfers throughout training.

We also redid the documentation to use auto-api rather than auto-doc.

What's Changed

• Index-batching functionality by @OckermanSethGVSU in #294
• Index-batching documentation update by @OckermanSethGVSU in #295
• Documentation overhaul by @OckermanSethGVSU in #296

New Contributors

• @OckermanSethGVSU made their first contribution in #294

Full Changelog: 0.55.0...0.56.0

1. Fix SSL import in #262
2. Fix to_dense_adj function in tsgcn via #271

What's Changed

• Remove setup.py test (deprecated) by @jamesmyatt in #182
• Make tqdm an optional dependency by @jamesmyatt in #188
• Remove torch-scatter from mandatory dependencies by @jamesmyatt in #191
• Update installation instructions for PyTorch 1.11.0 and PyG latest by @jamesmyatt in #192
• Remove scipy from mandatory dependencies (not used) by @jamesmyatt in #190
• change on _set_hidden_state() by @h3dema in #187

New Contributors

• @jamesmyatt made their first contribution in #182
• @h3dema made their first contribution in #187

Full Changelog: v0.53.0...v0.54.0

What's Changed

• remove unused parameters and variables for MPNNLSTM by @SherylHYX in #154
• Support for slicing signals by @gfngoncalves in #160
• Change dicts in hetero_gc_lstm to nn.ParameterDicts by @xunil17 in #180
• Optimize code to prevent repeated calls to convolution operator by @xunil17 in #181

New Contributors

• @gfngoncalves made their first contribution in #160
• @xunil17 made their first contribution in #180

Full Changelog: v0.52.0...v0.53.0

What's Changed

• Hetero support by @doGregor in #142
• Fixed Montevideo Bus dataset by @dtortorella in #147
• Matrix multiplication bug fixed by @doGregor in #148
• Fixed lambda_max error when normalization != sym by @gravins in #150

New Contributors

• @gravins made their first contribution in #150

Full Changelog: v0.51.0...v0.52.0

What's Changed

• Temporal signal split fix by @tforgaard in #129
• Refactor getitem by @josephenguehard in #133
• First commit for heterogeneous graph support by @doGregor in #125
• Fix an error in documentation for TGCN cell. by @BraveDistribution in #136

New Contributors

• @tforgaard made their first contribution in #129
• @josephenguehard made their first contribution in #133
• @doGregor made their first contribution in #125
• @BraveDistribution made their first contribution in #136

Full Changelog: v0.50.0...v0.51.0

• Removed coupling to PyG 1.7.0
• Removed coupling to PyTorch 1.7.
• DevOps updated
• Documentation and Installation guide updated

What's Changed

• Feature extension: Additional named attributes in Signal/Batch objects by @Flunzmas 🤖
• Fixed EvolveGCN weight squeezing by @dtortorella 🌃
• Updated the A3TGCN_example by @poteman 🎇
• Make attention trainable in A3TGCN and make it support batches by @elmahyai 🌄

• Evolve GCN-O Fix of Weight Updates
• Evolve GCN-H Fix of Weight Updates

• GMAN move out - hyperparameters are not fixed.