- Paper link: http://papers.nips.cc/paper/6703-inductive-representation-learning-on-large-graphs.pdf
- Author's code repo: https://github.com/williamleif/graphsage-simple. Note that the original code is simple reference implementation of GraphSAGE.
Run with following (available dataset: "cora", "citeseer", "pubmed")
python train_full.py --dataset cora # full graphRun with following (available dataset: "cora", "citeseer", "pubmed")
# use fensorflow background
TL_BACKEND=tensorflow python train_full.py --dataset cora --lr 0.01 --hidden_dim 128 --drop_rate 0.7 --n_epoch 500
TL_BACKEND=tensorflow python reddit_sage.py --lr 0.0005 --hidden_dim 256 --drop_rate 0.8TL_BACKEND=paddle python train_full.py --dataset cora --n_epoch 500 --lr 0.005 --hidden_dim 512 --drop_rate 0.7 --n_epoch 500
CUDA_VISIBLE_DEVICES=5 TL_BACKEND=paddle python reddit_sage.py --lr 0.001 --hidden_dim 128 --drop_rate 0.8# use pytorch
TL_BACKEND=torch python train_full.py --dataset cora --n_epoch 500 --lr 0.005 --hidden_dim 512 --drop_rate 0.8
TL_BACKEND=torch python reddit_sage.py --lr 0.005 --hidden_dim 128 --drop_rate 0.8| Dataset | Cora | |
|---|---|---|
| DGL | 83.3 | 94.95 |
| Paper | 83.3 | 95.0 |
| GammaGL(tf) | 82.44 ± 0.88 | 95.0 |
| GammaGL(th) | 81.13 ± 1.08 | 94.9 |
| GammaGL(pd) | 82.04 ± 0.33 | 91.2 |
| GammaGL(ms) | --.- | --.- |
- We fail to reproduce the reported accuracy on 'Cora', even with the DGL's code.
- The model performance is the average of 5 tests