search

DSE-MSU / DeepRobust

A pytorch adversarial library for attack and defense methods on images and graphs
MIT License
994 stars 192 forks source link

Cannot load cora-ml with 15% and 20% perturbation rate #52

Closed liu-jc closed 3 years ago

liu-jc commented 3 years ago

Hi,

I found we cannot load cora-ml datasets with 15% and 20% perturbation rate when I am trying to run test_sgc.py. The problem is that we cannot find cora_ml_meta_adj_0.15.npz and cora_ml_meta_adj_0.2.np in the github repo Pro-GNN. As you are from the same research group, I just open an issue here. Hope these two perturbed graphs can be uploaded there. Then the problem will be resolved and it will make the example more complete. Thanks in advance.

ChandlerBang commented 3 years ago

Hi,

Thanks for the comment. I redid the attacks (0.5, 0.1, 0.15, 0.2, 0.25 perturbation rates) and uploaded the data to Pro-GNN. Note that the graphs under 0.05/0.1/0.25 perturbation rates are also different from previous ones.

liu-jc commented 3 years ago

Great! Thank you Wei Jin.

Regards, Juncheng On 27 Mar 2021, 6:46 AM +0800, Wei Jin @.***>, wrote:

Hi, Thanks for the comment. I redid the attacks (0.5, 0.1, 0.15, 0.2, 0.25 perturbation rates) and uploaded the data to Pro-GNN. Note that the graphs under 0.05/0.1/0.25 perturbation rates are also different from previous ones. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

liu-jc commented 3 years ago

Hi,

Thanks for the comment. I redid the attacks (0.5, 0.1, 0.15, 0.2, 0.25 perturbation rates) and uploaded the data to Pro-GNN. Note that the graphs under 0.05/0.1/0.25 perturbation rates are also different from previous ones. Hi Wei Jin,

I use SGC and GCN on the dataset with 0.25 perturbation rate. The test accuracies are both larger than 60%. On the dataset with 0.2 and 0.15 perturbation rates, the accuracies are usually less than 60%. I think it's quite strange. Could you check it? Maybe there're some mistakes?

ChandlerBang commented 3 years ago

Please remove the original files of the perturbed graphs. For example, if you stored them in "/tmp", you should remove /tmp/cora_ml_meta_adj_0.25.npz or rm /tmp/cora_ml_meta_*.

liu-jc commented 3 years ago

Please remove the original files of the perturbed graphs. For example, if you stored them in "/tmp", you should remove /tmp/cora_ml_meta_adj_0.25.npz or rm /tmp/cora_ml_meta_*.

Thanks. Now all look good!