pyg-team/pytorch_geometric 2.0.2

on GitHub

A new minor version release, including further bugfixes, official PyTorch 1.10 support, as well as additional features and operators:

Features

• Added video tutorials and Colabs from the PyTorch Geometric Tutorial project (thanks to @AntonioLonga)
• Added the GraphMultisetTransformer operator (thanks to @JinheonBaek)
• Added the PointTransformerConv operator (thanks to @QuanticDisaster)
• Added the HEATConv operator (thanks to @Xiaoyu006)
• Added the PNA GNN model (thanks to @RBendias)
• Added the AddMetaPaths transform, which will add additional edge types to a HeteroData object based on a list of metapaths (thanks to @wsad1)
• Added the Data.to_heterogeneous method to allow for the conversion from Data to HeteroData objects
• Added the AttributedGraphDataset, containing a variety of attributes graphs
• Added the Airports datasets
• Added the structured_negative_sampling_feasible method, which checks if structured_negative_sampling is feasible (thanks to @WuliangHuang)
• GATConv can now make use of multi-dimensional edge features to compute attention scores (thanks to @dongkwan-kim)
• RandomNodeSplit and RandomLinkSplit now support HeteroData as input
• MessagePassing inference can now be sped up via the decomposed_layers argument (thanks to @ZhouAo-ZA)
• negative_sampling and batched_negative_sampling now support negative sampling in bipartite graphs
• HeteroConv now supports the inclusion of arbitrary node-level or edge-level information for the underlying MessagePassing operators
• GNNExplainer now supports multiple node-level masks and explaining regression problems (thanks to @gregorkrz)

Minor Changes

• Data.to_homogeneous will now add node_type information to the homogeneous Data object
• GINEConv now allows to transform edge features automatically in case their dimensionalities do not match (thanks to @CaypoH)
• OGB_MAG will now add node_year information to paper nodes
• Entities datasets do now allow the processing of HeteroData objects via the hetero=True option
• Batch objects can now be batched together to form super batches
• Added heterogeneous graph support for Center, Constant and LinearTransformation transformations
• HeteroConv now allows to return "stacked" embeddings
• The batch vector of a Batch object will now be initialized on the GPU in case other attributes are held in GPU memory

Bugfixes

• Fixed the num_neighbors argument of NeighborLoader in order to specify an edge-type specific number of neighbors
• Fixed the collate policy of lists of integers/strings to return nested lists
• Fixed the Delaunay transformation in case the face attribute is not present in the data
• Fixed the TGNMemory module to only read from the latest update (thanks to @cwh104504)
• Fixed the pickle.PicklingError when Batch objects are used in a torch.multiprocessing.manager.Queue() (thanks to @RasmusOrsoe)
• Fixed an issue with _parent state changing after pickling of Data objects (thanks to @zepx)
• Fixed the ToUndirected transformation in case the number of edges and nodes are equal (thanks to @lmkmkrcc)
• Fixed the from_networkx routine in case node-level and edge-level features share the same names
• Removed the num_nodes warning when creating PairData objects
• Fixed the initialization of the GeneralMultiLayer module in GraphGym (thanks to @fjulian)
• Fixed custom model registration in GraphGym
• Fixed a clash in the run_dir naming of GraphGym (thanks to @fjulian)
• Includes a fix to prevent a GraphGym crash in case ROC-score is undefined (thanks to @fjulian)
• Fixed the Batch.from_data_list routine on dataset slices (thanks to @dtortorella)
• Fixed the MetaPath2Vec model in case there exists isolated nodes
• Fixed torch_geometric.utils.coalesce with CUDA tensors