v0.4.0

GraphNeuralNetworks v0.4.0

Diff since v0.3.15

Closed issues:

• hope can load dataset with GraphMLDatasets.jl (#149)
• GPU memory filling up (#150)
• Citing GraphNeuralNetworks.jl (#151)

Merged pull requests:

• Citing (#152) (@CarloLucibello)
• MLUtils and Flux v0.13 compatibility (#155) (@CarloLucibello)

v0.3.15

GraphNeuralNetworks v0.3.15

Diff since v0.3.14

Merged pull requests:

• better printing for GNNGraphs and check that features are arrays (#145) (@CarloLucibello)
• Added GMMConv (#147) (@melioristic)
• Update conv.md (#148) (@melioristic)

v0.3.14

GraphNeuralNetworks v0.3.14

Diff since v0.3.13

Closed issues:

• Explainer vs GeometricFlux (#81)
• Gradient of edge weights is nothing with fused e_mul_xj (#113)
• Roadmap to merge GeometricFlux.jl and GraphNeuralNetworks.jl (#132)
• Failed to compile PTX code (#133)
• GINConv not working on GPU when not all nodes are connected (#138)
• Batchnorm for Integers after GCNConv or GINConv on GPU (#140)
• Slow interaction with DataLoader (#141)

Merged pull requests:

• differentiable adjacency_matrix and degree (#123) (@CarloLucibello)
• remove cast rrule not needed anymore (#135) (@CarloLucibello)
• handle isolated nodes (#139) (@CarloLucibello)
• dataloader support for vector of graphs (#143) (@CarloLucibello)

v0.3.13

GraphNeuralNetworks v0.3.13

Diff since v0.3.12

Closed issues:

• Failure to combine SparseDiffTools.autoback_hesvec and GCNConv (#125)
• Support edge attributes in as many layers as possible (#128)

Merged pull requests:

• Improve support in conv layers for edge features (#130) (@CarloLucibello)
• GATv2 fix with edge features + doc improves (#131) (@CarloLucibello)

v0.3.12

GraphNeuralNetworks v0.3.12

Diff since v0.3.11

Closed issues:

• batching scales quadratically (#99)
• aggregate_neighbors() is 100x slower than equivalent sparse matrix operation (#124)
• GATv2Conv show method errors (#126)

Merged pull requests:

• faster batching (#122) (@CarloLucibello)
• fix GATv2Conv show (#129) (@CarloLucibello)

v0.3.11

GraphNeuralNetworks v0.3.11

Diff since v0.3.10

Merged pull requests:

• CompatHelper: bump compat for NNlibCUDA to 0.2, (keep existing compat) (#120) (@github-actions[bot])
• implement hash for GNNGraph (#121) (@CarloLucibello)

v0.3.10

GraphNeuralNetworks v0.3.10

Diff since v0.3.9

Closed issues:

• Implement add_reverse_edges (#103)
• propagate() is 20x slower than built-in sparse matmul (#106)
• add fusing propagate specialization for e_mul_xj (#108)
• implement set_edge_weights (#110)

Merged pull requests:

• fused e_mul_xj and weighted option for adjacency_matrix (#107) (@CarloLucibello)
• Typecast GNNGraph.num_nodes to Int (#109) (@abieler)
• Potential Spelling fix (#111) (@umbriquse)
• add w_mul_xj and set_edge_weight (#114) (@CarloLucibello)
• support for aggregating edge feature in remove_multi_edges (#115) (@CarloLucibello)
• add to_bidirected (#116) (@CarloLucibello)
• add Google Summer of Code [GSOC] doc page (#117) (@CarloLucibello)
• CompatHelper: bump compat for NNlib to 0.8, (keep existing compat) (#119) (@github-actions[bot])

v0.3.9

GraphNeuralNetworks v0.3.9

Diff since v0.3.8

Closed issues:

• Problem with InlineStrings.jl (#98)
• Include undirected graphs (#101)
• Merging multiple feature arrays (#102)
• conflict with CSV and GNNGraphs when running Flux.batch (#104)

Merged pull requests:

• workaround for sort ambiguity in cat_features (#105) (@CarloLucibello)

v0.3.8

GraphNeuralNetworks v0.3.8

Diff since v0.3.7

Closed issues:

• Custom Function GPU Compatibitlity Issue: Indexing (#91)
• Flux.batch Overloading for Generators (#92)
• outputsize for GNNChain (#96)

Merged pull requests:

• add sample_neighbors (#93) (@CarloLucibello)
• CompatHelper: bump compat for LearnBase to 0.6, (keep existing compat) (#95) (@github-actions[bot])
• Adds GATv2 layer (#97) (@abieler)

v0.3.7

GraphNeuralNetworks v0.3.7

Diff since v0.3.6

Closed issues:

• Hash function for GNNGraph (#87)
• no method matching getobs(::NamedTuple{(:x,), Tuple{Matrix{Float32}}} (#88)

Merged pull requests:

• GeometricFlux cora comparison (#82) (@CarloLucibello)
• implement MEGNetConv (#83) (@CarloLucibello)
• small doc update (#84) (@CarloLucibello)
• export get_edge_weight + use weights from adjacency matrix (#86) (@CarloLucibello)
• don't rely on LearnBase.getobs (#90) (@CarloLucibello)