2.0.3

PyG 2.0.3 🎉

A new minor PyG version release, including a variety of new features and bugfixes:

Features

• GLNN: Graph-less Neural Networks [Example] (#3572)
• LINKX: Large Scale Learning on Non-Homophilous Graphs [Example] (#3654)
• Added an example for heterogeneous link classification (#3350) - thanks to @anniekmyatt
• HANConv: The Heterogenous Graph Attention operator [Example] (#3444, #3577, #3581) - thanks to @rishubhkhurana and @wsad1
• LGConv and LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation (#3685) - thanks to @LukasHaas and @KathyFeiyang
• PyTorch Lightning DataModule wrappers for PyG+PL multi-GPU training/inference without replicating datasets across processes :
  • torch_geometric.data.LightningDataset for multi-GPU training via PL on graph-level tasks [Example] (#3596, #3634)
  • torch_geometric.data.LightningNodeData for multi-GPU training via PL on node-level tasks [Example] (#3613, #3634)
• NeighborLoader: Added CUDA support leading to major runtime improvements [Example] (#3736)
• MessagePassing: Added the edge_updater/edge_update interface for updating edge features (#3450) - thanks to @Padarn
• GNNExplainer: Added an example that reproduces the official BA-Shapes experiment (#3386) - thanks to @RBendias
• torch_geometric.graphgym: Support for heterogeneous graphs and lazy initialization (#3460) - thanks to @JiaxuanYou
• MLP: Added a basic MLP implementation (#3553)
• PointTransformer: Classification and segmentation examples (#3344) - thanks to @QuanticDisaster and @wsad1
• ShaDowKHopSampler: Added an example (#3411) - thanks to @SubhajitDuttaChowdhury
• Data.subgraph(...) implementation (#3521)

Datasets

• HGBDataset benchmark suite (#3454)
• MalNetTiny dataset (#3472) - thanks to @rampasek
• OMDB: Organic Materials Database (#3506)
• BAShapes: The BA-Shapes dataset (#3386) - thanks to @RBendias
• PolBlogs and EmailEUCore datasets (#3534) - thanks to @AlexDuvalinho
• StochasticBlockModel and RandomPartition graph datasets (#3586) - thanks to @dongkwan-kim
• LINKXDataset: A subset of the non-homophilous benchmark datasets from LINKX
• FakeDataset and FakeHeteroDataset for testing purposes (#3741) - thanks to @levulinh

Minor Changes

• torch_geometric.nn.norm: Improved the runtimes of normalization layers - thanks to @johnpeterflynn
• DataLoader and NeighborLoader: Output tensors are now written to shared memory to avoid an extra copy in case num_workers > 0 (#3401 and #3734) - thanks to @johnpeterflynn
• GATv2Conv: Support for edge features (#3421) - thanks to @Kenneth-Schroeder
• Batch.from_data_list: Runtime improvements
• TransformerConv: Runtime and memory consumption improvements (#3392) - thanks to @wsad1
• mean_iou: Added IoU computation via omitting NaNs (#3464) - thanks to @GericoVi
• DataLoader: follow_batch and exclude_keys are now optional arguments
• Improvements to the package metadata (#3445) - thanks to @cthoyt
• Updated the quick start widget to support PyTorch 1.10 (#3474) - thanks to @kathyfan
• NeighborLoader and HGTLoader: Removed the persistent_workers=True default
• voxel_grid: The batch argument is now optional (#3533) - thanks to @QuanticDisaster
• TransformerConv: JIT support (#3538) - thanks to @RobMcH
• Lazy modules can now correctly be saved and loaded via state_dict() and load_state_dict() (#3651) - thanks to @shubham-gupta-iitr
• from_networkx: Support for nx.MultiDiGraph (#3646) - thanks to @max-zipfl-fzi
• GATv2Conv: Support for lazy initialization (#3678) - thanks to @richcmwang
• torch_geometric.graphgym: register_* functions can now be used as decorators (#3684)
• AddSelfLoops: Now supports the full argument set of torch_geometric.utils.add_self_loops (#3702) - thanks to @dongkwan-kim
• Documentation: Added shape information to a variety of GNN operators, e.g., GATConv or ChebConv (#3697) - thanks to @saiden89
• GATv2Conv and HEATConv: Removed unnecessary size argument in forward (#3744) - thanks to @saiden89

Bugfixes

• GNNExplainer: Fixed a bug in the GCN example normalization coefficients were wrongly calculated (#3508) - thanks to @RBendias
• HGTConv: Fixed a bug in the residual connection formulation - thanks to @zzhnobug
• torch_geometric.grapghym: Fixed a bug in the creation of MLP (#3431) - thanks to @JiaxuanYou
• torch_geometric.graphgym: Fixed a bug in the dimensionality of GeneralMultiLayer (#3456) - thanks to @JiaxuanYou
• RandomLinkSplit: Fixed a bug in negative edge sampling for undirected graphs (#3440) - thanks to @panisson
• add_self_loops: Fixed a bug in adding self-loops with scalar-valued weights
• SchNet: Fixed a bug in which a bias vector was not correctly initialized as zero - thanks to @nec4
• Batch.from_data_list: Replaced the torch.repeat_interleave call due to errors in forked processes (#3566) - thanks to @Enolerobotti
• NeighborLoader: Fixed a bug in conjunction with PyTorch Lightning (#3602) - thanks to @pbielak
• NeighborLoader and ToSparseTensor: Fixed a bug in case num_nodes == num_edges (#3683) - thanks to @WuliangHuang
• ToUndirected: Fixed a bug in case num_nodes == 2 (#3627) - thanks to @aur3l14no
• FiLMConv: Fixed a bug in the backward pass due to the usage of in-place operations - thanks to @Jokeren
• GDC: Fixed a bug in case K > num_nodes - thanks to @Misterion777
• LabelPropagation: Fixed a bug in the order of transformations (#3639) - thanks to @Riyer01
• negative_sampling: Fixed execution for GPU input tensors - thanks to @Sticksword and @lmy86263
• HeteroData: Fixed a bug in which node types were interpreted as edge types in case they were described by two characters (#3692)
• FastRGCNConv: Fixed a bug in which weights were indexed on destination node index rather than source node index (#3690) - thanks to @Jokeren
• WikipediaNetwork: Fixed a bug in downloading due to a change in URLs - thanks to @csbobby and @Kousaka-Honoka