dmlc/dgl 0.8.1

v0.8.1

on GitHub

This is a minor release that includes the following model updates, optimizations, new features and bug fixes.

Model update

• nn.GroupRevRes from Training Graph Neural Networks with 1000 layers [#3842]
• transforms.LaplacianPositionalEncoding from Graph Neural Networks with Learnable Structural and Positional Representations [#3869]
• transforms.RWPositionalEncoding from Graph Neural Networks with Learnable Structural and Positional Representations [#3869]
• dataloading.SAINTSampler from GraphSAINT [#3879]
• nn.EGNNConv from E(n) Equivariant Graph Neural Networks [#3901]
• nn.PNAConv from the baselines of E(n) Equivariant Graph Neural Networks [#3901]

Example update

• Position-aware GNN [#3823 @RecLusIve-F]
• EGES (Enhanced Graph Embedding with Side info) [#3756 @Wang-Yu-Qing]

Feature update (new functionalities, interface changes, etc.)

• Radius graph - construct a graph by connecting points within a given distance. [#3829 @ksadowski13]
  • It uses torch.cdist so the space complexity is O(N^2).
• Added a get_attention parameter in GlobalAttentionPooling. [#3837 @decoherencer]

Quality of life update

• Example to train with multi-GPU with PyTorch Lightning. [#3863]
• Multi-GPU inference with UVA. [#3827 @nv-dlasalle]
• Enable UVA sampling with CPU indices to save GPU memory. [#3892]
• Set stacklevel=2 for DGL-raised warnings. [#3816]
• Pure GPU example of GraphSAGE, with both node classification and link prediction. [#3796 @nv-dlasalle, #3856 @Kh4L]
• Tensoradapter DLPack 0.6 compatibility / PyTorch 1.11 support. [#3803]

System optimization

• Enable UVA for PinSAGE and RandomWalk. [#3857 @yaox12]
• METIS partition with communication volume minimization, reduces the communication volume by 13.4% compared with edge-cut minimization on ogbn-products. [#3821 @chwan1016]
• Change parameter of curand_init for reducing GPU latency [#3794 @paoxiaode]

Bug fixes

• Fix Python 3.10 import error [#3862]
• Fix repeated 0’s in DataLoader index iteration when shuffle=False [#3892]
• DataLoader device cannot be None [#3822 @yinpeiqi]
• Fix device error in negative sampling with UVA [#3904 @nv-dlasalle]
• Illegal instruction in ClusterGCNSampler (#3910)
• Include pin memory status in pickling and deep copy [#3914]
• Misc doc fixes (@lvcrek @AzureLeon1 @decoherencer @yaox12 @ketyi )