def test_given_adj(p_adj: sp.csr_matrix, p_feat: np.ndarray,
p_label: np.ndarray, p_train_idx, p_val_idx, p_test_idx):
"""
Parameters:
p_adj : csr_matrix
"""
use_gpu = t.cuda.is_available()
adj = p_adj.A
features = p_feat
labels = p_label
# surrogate model
surrogate = GCN(nfeat=features.shape[1],
nclass=labels.max().item() + 1,
nhid=16,
dropout=0,
with_relu=True,
with_bias=True,
device='cuda:0').cuda()
# train model
surrogate.fit(features, adj, labels, p_train_idx, p_val_idx, patience=30)
# test_model
surrogate.test(p_test_idx)
return surrogate
model = GCN(nfeat=features.shape[1],
nhid=args.hidden,
nclass=labels.max().item() + 1,
dropout=args.dropout,
device=device)
if args.only_gcn:
perturbed_adj, features, labels = preprocess(perturbed_adj,
features,
labels,
preprocess_adj=False,
sparse=True,
device=device)
model.fit(features,
perturbed_adj,
labels,
idx_train,
idx_val,
verbose=True,
train_iters=args.epochs)
model.test(idx_test)
else:
perturbed_adj, features, labels = preprocess(perturbed_adj,
features,
labels,
preprocess_adj=False,
device=device)
prognn = ProGNN(model, args, device)
prognn.fit(features, perturbed_adj, labels, idx_train, idx_val)
prognn.test(features, labels, idx_test)
# no defense
model = GCNTorch(nfeat=features.shape[1],
nhid=16,
nclass=labels.max() + 1,
device=device)
model = model.to(device)
model.fit(features,
nettack.adj_preprocessed,
labels,
idx_train,
idx_val,
train_iters=200,
verbose=False,
normalize=False)
model.eval()
acc = model.test(idx_test)
probs_after_defense = model.predict().detach().cpu().numpy()[u]
preds['no'] = np.argmax(probs_after_defense)
###########################
# vae settings
model = 'arga_vae' #'arga_ae' or 'arga_vae'
iterations = 500 if dataname == "polblogs" else 200
train = True # train or load pretrain
###########################
percentile = (1 - adj.sum() / adj.shape[0]**2 * 20) * 100
reconstructor = Link_rec_Runner(nettack.adj, features, model, iterations)
reconstructor.erun(train=train)
print("finishing reconstruction")
torch.cuda.manual_seed(args.seed)
vgae = GCN(nfeat=features.shape[1],
nhid=16,
nclass=labels.max() + 1,
device=device)
vgae = vgae.to(device)
vgae.fit(features,
adj_rec_norm,
labels,
idx_train,
idx_val,
train_iters=200,
verbose=False,
normalize=False)
vgae.eval()
output = vgae.test(idx_val)
res.append(output.item())
print("retraining")
bst = np.argmax(res) + 1
percentile = (1 - perturbed_adj.sum() / adj.shape[0]**2 * bst) * 100
adj_rec = reconstructor.sparse(percentile)
adj_rec_norm = utils.preprocess_graph(adj_rec)
torch.manual_seed(args.seed)
if args.cuda:
torch.cuda.manual_seed(args.seed)
vgae = GCN(nfeat=features.shape[1],
nhid=16,
nclass=labels.max() + 1,
device=device)
reconstructor = Link_rec_Runner(adj, features, model, iterations)
reconstructor.erun(train=train)
print("finishing reconstruction")
result = []
for p in percentiles:
torch.manual_seed(args.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(args.seed)
print(p)
adj_rec = reconstructor.sparse(p)
adj_rec_norm = preprocess_graph(adj_rec)
model = GCN(nfeat=features.shape[1],
nhid=16,
nclass=labels.max() + 1,
device=device)
model = model.to(device)
model.fit(features,
adj_rec_norm,
labels,
idx_train,
idx_val,
train_iters=200,
verbose=False,
normalize=False)
model.eval()
acc = model.test(idx_test)
result.append(acc.item())
print(result)
