v0.7.7

GeometricFlux v0.7.7

Diff since v0.7.6

• drop support of julia v1.4 and v1.5
• support CUDA v3.3
• support Flux v0.12
• fix stable doc
• add benchmark script
• migrate scatter to NNlib
• make gradient of GatedGraphConv available
• Implement GINConv layer. (#186)
• check consistency for vertex or edge number between graph and features
• add manual for pooling layers and bypass_graph
• deprecate FeatureSelector
• not export GraphNetwork and MessagePassing APIs
• new implementation for message-passing scheme

Closed issues:

• Increase test cases (#7)
• Method error when training GCNConv with FeaturedGraph input (#76)
• More "mutating arrays is not supported" (#103)
• missing adjoints for pooling functions with specified graph size (#147)
• drop ScatterNNlib in favor of NNlib.scatter (#191)
• drop support in layer for arbitrary graph types and use only FeaturedGraphs (#201)
• FeaturedGraph cannot be moved to gpu (#202)
• deprecate FeatureSelector (#207)
• broken documentation links (#210)

Merged pull requests:

• CompatHelper: bump compat for "CUDA" to "3.0" (#175) (@github-actions[bot])
• Use mapreduce in message-passing (#179) (@yuehhua)
• Refactor pool function into scatter function (#180) (@yuehhua)
• Update doc (#181) (@yuehhua)
• Fix building doc (#182) (@yuehhua)
• Some cleanup (#183) (@yuehhua)
• Update docs (#184) (@yuehhua)
• Implement GINConv layer. (#186) (@emsal0)
• Replace ScatterNNlib with NNlib (#187) (@yuehhua)
• Refactor (#189) (@yuehhua)
• remove learnable parameters when bias=false (#196) (@CarloLucibello)
• misc cleanup and add buildkite (#198) (@CarloLucibello)
• homogenize forward pass for all layers (#203) (@CarloLucibello)
• misc improvements (#206) (@CarloLucibello)
• Improve documentation (#208) (@yuehhua)
• Fix doc (#211) (@yuehhua)
• add benchmark script (#214) (@CarloLucibello)
• Fix GINConv docstring. (#219) (@emsal0)