GAMMA Lab officially released version 0.5 of GammaGL today. You can download the latest GammaGL from the OpenI community, Github, or through pip.

Highlights

• Implemented interfaces for GraphStore and FeatureStore, and provided an example of training GraphSAGE using the Reddit dataset. #215
• The segment operator now supports data inputs of types float16, float64, and int. #197
• Implemented the bspmm operator, achieving a performance improvement of 2.3X compared to previous methods. #210
• Improved the invocation of the spmm and bspmm operators, allowing users to simply implement the message_aggregate method in convolution layers. #216
• Support for sampling under CUDA. #209
• Use of a unified extension interface for project compilation. #219

New Additions

• Add Hid-Net, a general diffusion equation framework with fidelity terms [AAAI 2023] #202
• Add FusedGAT, a graph neural network model optimized from computation, IO, and memory aspects [MLSys 2022] #195
• Add GLNN, a graph neural network model combining GNN and MLP through knowledge distillation [ICLR 2022] #205
• Add GNNLF-HF, a graph neural network model revealing essential links between different GNN propagation mechanisms [WWW 2021] #9f86d5e
• Add Sp2GCL, a graph contrastive learning framework integrating spatial and spectral views [NeurIPS 2023] #201
• Add DFAD-GNN, a data-free adversarial knowledge distillation graph neural network model [IJCAI 2022] #212
• Add HEAT, a trajectory prediction framework for heterogeneous traffic participants [T-ITS 2022] #206
• Add HeCo, a self-supervised heterogeneous graph neural network contrastive learning framework [KDD 2021] #204
• Add DHN, a distance encoding technique to enhance the expressive power of heterogeneous graph neural networks [TKDE 2023] #207
• Add DNA, an attention-based dynamic neighborhood aggregation method for graph representation learning [ICLR 2019] #214
• Add homogeneous graph datasets: Facebook (#201) and Yelp (#5c7ccfe)
• Add heterogeneous graph datasets: ACM4DHN and ACM4HeCo #204
• Add NGSIM dataset for traffic simulation #d741516
• Add SVDFeatureReduction, an SVD method for feature decomposition #6ddba3a
• Add get_train_val_test_split, a new interface for dataset partitioning #080b4d5
• Add dataset partitioning fields to the Amazon dataset #dc0511c

Bugs

• Fix repeated definition issue of forward dropout in MLP #2403822
• Fix calculation errors of the segment_mean operator in CUDA environments #fb00d7f
• Fix efficiency issues with the get_laplacian utility class when computing on large graphs #201

GAMMA Lab officially released version 0.4 of GammaGL today. You can download the latest gammaGL from the OpenI community, GitHub, or through pip.

New Additions

• Add CaGCN model based on confidence correction [NeurIPS 2021] (#157)
• Add Graph structure learning model CoGSL [WWW 2022] (#170)
• Add DR-GST model for data distribution shift [WWW 2022] (#161)
• Add spectral domain self-attention model SpecFormer [ICLR 2023] (#172)
• Add Adaptive multi-channel graph convolutional network model AM-GCN [KDD 2020] (#184)
• Add graph contrastive "provable training" model GNN-POT [NeurIPS 2023] (#171)
• Add graph contrastive learning model GGD based on group discrimination [NeurIPS 2022] (#181)
• Add spectral cotrastive learning model Grace-SpCo [NeurIPS 2022] (#189)
• Add knowledge distillation framework LTD [WSDM 2022] (#186)
• Add Graphormer based on Transformer [NeurIPS 2021] (#193)
• Add homogeneous graph datasets WikiCS and PolBlogs and BlogCatalog (#176)
• Add dataset for molecular machine learning MoleculeNet (#192)
• Optimize the issue of console output redundant content when training the model (#183)
• Add .clang-format to adjust C++ code format (#182)
• Added Github Action to help users verify the correctness when submitting code (#191)

Bugs

• Fixe the problem that HGB dataset cannot be downloaded (#167)
• Fix the abnormal edge feature problem of tu dataset (commit 6fabdfd)
• Fix data type issue in Reddit dataset (commit c3c06d8)
• Fixed the problem that the RGCN model does not support the latest version of Numpy (#183)
• Fix the calculation error of segment operator during backpropagation (#173)

Support Latest PyTorch and MindSpore

Currently, GammaGL supports PyTorch version 2.1.2 and MindSpore version 2.2.10.

We have made compatibility fixes for TensorLayerX. For PyTorch, GammaGL supports not only old versions before 1.10, but also the latest version 2.1.2. The relevant TensorLayerX repository link is https://github.com/dddg617/TensorLayerX/tree/nightly.

Database Support

The latest version of GammaGL adds support for graph databases. We developed the graph database interface gdbi, which implements a unified standard interface for four graph databases (neo4j, NebulaGraph, AtlasGraph, and g-store). You can get the example here (#194). The relevant repository link is https://github.com/xy-Ji/gdbi.

1. New models
2. Faster message passing APIs
3. Renew the computation structure with pybind11
4. Dataset cache

1. New models
2. New datasets
3. Faster API
4. GPU sample and heterogeneous graph sample