dgi

Deep Graph Infomax (DGI)

• Paper link: https://arxiv.org/abs/1809.10341
• Author's code repo (in Pytorch): https://github.com/PetarV-/DGI

How to run

Run with following:

# use tensorflow backend
TL_BACKEND=paddle python dgi_trainer.py --dataset cora --lr 0.002 --patience 50
TL_BACKEND=paddle python dgi_trainer.py --dataset citeseer --lr 0.0005 --patience 20 --n_epoch 300
TL_BACKEND=paddle python dgi_trainer.py --dataset pubmed --lr 0.001 --hidden_dim 256 --patience 20 

# use paddle backend
TL_BACKEND=tensorflow python dgi_trainer.py --dataset cora --lr 0.003 --patience 50
TL_BACKEND=tensorflow python dgi_trainer.py --dataset citeseer --lr 0.001 --patience 20 --n_epoch 100
TL_BACKEND=tensorflow python dgi_trainer.py --dataset pubmed --hidden_dim 256 --lr 0.001

# use pytorch backend
TL_BACKEND=torch python dgi_trainer.py --dataset cora 
TL_BACKEND=torch python dgi_trainer.py --dataset citeseer
TL_BACKEND=torch python dgi_trainer.py --dataset pubmed --lr 0.001 --patience 20

Results

Dataset	Cora	Citeseer	Pubmed
Author's Code	82.3	71.8	76.8
DGL	81.6	69.4	76.1
GammaGL(tf)	81.51 ± 0.55	69.01 ± 0.91	78.37 ± 0.37
GammaGL(th)	--.-	--.-	79.58 ± 0.52
GammaGL(pd)	81.19 ± 0.64	69.06 ± 0.50	78.58 ± 0.65
GammaGL(ms)	--.-	--.-	--.-

• The model performance is the average of 5 tests