I can't reproduce the published results of unsupervised_Cora_Citeseer

[c747903646]

I have respectively run the codes using commands python -u execute.py --dataset citeseer --aug_type node --drop_percent 0.20 --seed 39 --save_name cite_best_dgi.pkl --gpu 0 and python -u execute.py --dataset cora --aug_type node --drop_percent 0.20 --seed 39 --save_name cite_best_dgi.pkl --gpu 0 on the following two environments:

1. soft env: CUDA 11.1, Python 3.6.7, PyTorch 1.8.0, hard env: RTX 3080 laptop
2. soft env: CUDA 10.0, Python 3.7.6, Pytorch 1.3.1, hard env: RTX 2080ti

and respectively got the results as the following:

1.

Citeseer

Average accuracy:[0.7279]
Mean:[72.7880]
Std :[0.1394]

Cora

Average accuracy:[0.8124]
Mean:[81.2360]
Std :[0.1274] 

2.

Citeseer

Average accuracy:[0.7256]
Mean:[72.5620]
Std :[0.1615]

Cora

Average accuracy:[0.8137]
Mean:[81.3660]
Std :[0.1649]

But according to the paper's Appendix G, the results should be 73.14±0.15 on Citeseer and 81.76±0.17 on Cora. How can I reproduce the results?

[yyou1996]

Hi @c747903646, thanks for asking.

@yongduosui any comments?

[yongduosui]

Different augmentation types, hyper-parameters, random seeds may lead to different results. Please try to adjust them and run again, thanks.