0.7.1 Release Notes

0.7.1 is a minor release with multiple fixes and a few new models/features/optimizations included as follows.

Note: We noticed that 0.7.1 for Linux is unavailable on our anaconda repository. We are currently working on this issue. For now, please use pip installation instead.

New models

• GCN-based spam review detection (#3145, @kayzliu)
• CARE-GNN (#3187, @kayzliu)
• GeniePath (#3199, @kayzliu)
• EEG-GCNN (#3186, @JOHNW02)
• EvolveGCN (#3190, @maqy1995)

New Features

• Allows providing username in tools/launch.py (#3202, @erickim555)
• Refactor and allows customized Python binary names in tools/launch.py (#3205, @erickim555)
• Add support for distributed preprocessing for heterogeneous graphs (#3137, @ankit-garg)
• Correctly pass all DGL client server environment variables for user-defined multi-command (#3245, @erickim555)
• You can configure the DGL configuration directory with environment variable DGLDEFAULTDIR (#3277, @konstantino)

Optimizations

• Improve usage of pinned memory in sparse optimizer (#3207, @nv-dlasalle)
• Optimized counting of nonzero entries of DistTensor (#3203, @freeliuzc)
• Remove activation cache if not required (#3258)
• Edge excluding in EdgeDataLoader on GPU (#3226, @nv-dlasalle)

Fixes

• Update numbers for HiLANDER model (#3175)
• New training and test scripts for HiLANDER (#3180)
• Fix potential starving in socket receiver (#3176, @JingchengYu94)
• Fix typo in Tensorflow backend (#3182, @lululxvi)
• Add WeightBasis documentation (#3189)
• Default ntypes/etypes consistency between dgl.DGLGraph and dgl.graph (#3198)
• Set sharing strategy for SEAL example (#3167, @KounianhuaDu)
• Remove DGL_LOADALL in doc builds (#3150, @lululxvi)
• Fix distributed training hang with multiple samplers (#3169)
• Fix random_walk documentation inconsistency (#3188)
• Fix curand_init() calls in rowwise sampling leading to not-so-random results (#3196, @nv-dlasalle)
• Fix force_reload parameter of FraudDataset (#3210, @Orion-wyc)
• Fix check for num_workers for using ScalarDataBatcher (#3219, @nv-dlasalle)
• Tensoradapter linking issues (#3225, #3246, @nv-dlasalle)
• Diffpool loss did not consider the loss of first diffpooling layer (#3233, @yinpeiqi)
• Fix CUDA 11.1 SPMM crashing with duplicate edges (#3265)
• Fix DotGatConv attention bug when computing edge_softmax (#3272, @Flawless1202)
• RelGraphConv reshape argument is incorrect (#3256, @minchenGrab)
• Documentation typos and fixes (#3214, #3221, #3244, #3231, #3261, #3264, #3275, #3285, @amorehead, @blokhinnv, @kalinin-sanja)