Question about PGD graph global attack with CE/CW loss

[yuChen-XD]

Hi,
I have a doubt about the [test_pgd.py] (

DeepRobust/examples/graph/test_pgd.py

Lines 90 to 96 in 756453e

	# model.attack(features, adj, labels, idx_train, perturbations, epochs=args.epochs)
	# Here for the labels we need to replace it with predicted ones
	fake_labels = target_gcn.predict(features.to(device), adj.to(device))
	fake_labels = torch.argmax(fake_labels, 1).cpu()
	# Besides, we need to add the idx into the whole process
	idx_fake = np.concatenate([idx_train,idx_test])
	model.attack(features, adj, fake_labels, idx_fake, perturbations, epochs=args.epochs)

).
When I use the CE loss, I get basically the same result as you showed in the previous issues, as shown below.

=== testing GCN on clean graph ===
Test set results: loss= 0.8108 accuracy= 0.8180
=== setup attack model ===
100%|██████████| 100/100 [01:35<00:00, 1.05it/s]
=== testing GCN on Evasion attack ===
Test set results: loss= 1.0158 accuracy= 0.7250
=== testing GCN on Poisoning attack ===
Test set results: loss= 0.9889 accuracy= 0.7320

However, when I replace the loss type in line87 to 'CW', the result, shown below, is not as the same level as the origin paper. I think the CW loss should have a similar ASR as the CE loss.

=== testing GCN on clean graph ===
Test set results: loss= 0.8108 accuracy= 0.8180
=== setup attack model ===
100%|██████████| 100/100 [01:28<00:00, 1.13it/s]
=== testing GCN on Evasion attack ===
Test set results: loss= 0.9067 accuracy= 0.7750
=== testing GCN on Poisoning attack ===
Test set results: loss= 0.9512 accuracy= 0.7450

Do you get this difference when you test the PGD method?
Thanks for your help in advance.