1.3.0

• Support for giant graph handling using NeighborSampler and bipartite message passing operators
• Debugging support using the new debug API
• Fixed TUDataset download errors
• Added FeasStConv module
• Improved networkx conversion functionality
• Improved Data and DataLoader handling with customizable number_of_nodes (e.g. for holding two graphs in a single Data object)
• Added GeniePath example
• Added SAGPool module
• Added geodesic distance computation using gdist (optional)
• Improved PointNet and DGCNN classification and segmentation examples
• Added subgraph functionality
• Fixed GMMConv
• Added a bunch of new datasets
• Added fast implementations for random graph generation
• Improved loop API
• Minor bugfixes

Thanks to all contributors!

1.2.1

• More convenient self-loop API (including addition of edge weights)
• Small bugfixes, .e.g., DiffPool NaNs and empty edge indices treatment
• New datasets have been added:
  • GEDDataset
  • DynamicFAUST
  • TOSCA
  • SHREC2016

1.2.0

• New models and operators, e.g., RENet, Signed Graph Convolution, Deep Graph Infomax, PPFNet, ...
• Minor bugfixes
• New converts

1.1.2

• bugfixes for bipartite message passing API

1.1.1

PointConv bugfix for bipartite graphs.

1.1.0

This release includes:

• All Variants of Graph Autoencoders
• Gated Graph Conv
• DataParallel bugfixes
• New transforms (Line Graph Transformation, Local Degree Profile, Sample Points with Normals)
• PointNet++ example

1.0.3

• SGC and APPNP layer

1.0.2

• Added remove_faces parameter for face transforms

1.0.1

• Finally completed documentation
• Finally achieved 100% code coverage (every single line is tested)
• Fixed a few minor bugs
• Added the GlobalAttention layer from Li et al.

1.0.0

We made a bunch of improvements to PyTorch Geometric and added various new convolution and pooling operators, e.g., top_k pooling, PointCNN, Iterative Farthest Point Sampling, PointNet++, ...