GINN

Developing completed (Prototype)

• GAT
• ConvKB-Attention
• Merge
• Testing
• Code Review

Requirement

• Require pytorch>=1.6.0 for CPU (torch.sparse.softmax)
• Require pytorch>=1.7.0 for GPU (CUDA torch.sparse.softmax)

History

Version 3.1 2020-11-30

• fix bugs in val and test dataset and dataloader
• add kinship dataset
• add k-hop option for dataset
• val and test running on cpu

Version 3.0 2020-11-20

• enable batch training
• enable batch to val and test also
• add customized collate function to dataloader for different length batch input

Version 2.1 2020-11-16

• modify filter to filter all other triples
• remove for loop in load data and clean up
• add None to attention type
• change ranking data from scipy to torch
• add reverse function

Version 2.0 2020-10-14

• add GAT for models.py
• add ConvE and separate DistMult for score_function.py
• add early stopping
• ConvAttenion code polished
• add GPU support

Version 1.1 2020-09-14

• add label smoothing
• add filter for model results
• change name to GINN
• ranking part code polished

Version 1.0 2020-08-31

• change loss to BCEloss as ConvE (1-N scoring)
• remove negative sampling
• add new ops_new dataset for extensive test triple

Version 0.3 2020-08-17

• add WN18RR dataset
• change measurement from ACC to MRR
• clean layers.py and parameterized

Version 0.2 2020-08-12

• create parser.py
• add negative sampling
• create utils.py

Version 0.1 2020-08-05

• initial commit
• add README